Skip to content

armgit / ECC

Public Imported from GitHub
0

The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond.

videodb skills update: add reference files for videodb skills

Rohit Gargcff03085682e
diff --git a/skills/videodb-skills/SKILL.md b/skills/videodb-skills/SKILL.md
index e4b3d8275..4f6b11ff6 100644
--- a/skills/videodb-skills/SKILL.md
+++ b/skills/videodb-skills/SKILL.md
@@ -1,109 +1,368 @@
---
name: videodb-skills
-description: The only video skill your agent needs — upload any video, connect real-time streams, search inside by what was said or shown, build complex editing workflows with overlays, generate AI media, add subtitles, and get instant streaming links.
+description: See, Understand, Act on video and audio. See- ingest from local files, URLs, RTSP/live feeds, or live record desktop; return realtime context and playable stream links. Understand- extract frames, build visual/semantic/temporal indexes, and search moments with timestamps and auto-clips. Act- transcode and normalize (codec, fps, resolution, aspect ratio), perform timeline edits (subtitles, text/image overlays, branding, audio overlays, dubbing, translation), generate media assets (image, audio, video), and create real time alerts for events from live streams or desktop capture.
origin: ECC
+allowed-tools: Read Grep Glob Bash(python:*)
+argument-hint: "[task description]"
---
-# VideoDB Skills
+# VideoDB Skill
-The only video skill your agent needs. Upload any video, connect real-time streams, search inside by what was said or shown, build complex editing workflows with overlays, generate AI media, add subtitles, and get instant streaming links — all via the VideoDB Python SDK.
+**Perception + memory + actions for video, live streams, and desktop sessions.**
-## When to Activate
+Use this skill when you need to:
-- Uploading or ingesting videos from YouTube URLs, web URLs, or local files
-- Searching spoken words or visual scenes across video content
-- Generating transcripts or auto-styling subtitles
-- Editing clips — trim, combine, multi-timeline composition
-- Adding overlays — text, images, audio, music
-- Generating AI media — images, video, music, sound effects, voiceovers
-- Transcoding — resolution, codec, bitrate, FPS changes
-- Reframing video for social platforms (vertical, square, etc.)
-- Real-time screen or audio capture with AI transcription
-- Getting playable HLS streaming links for any output
+## 1) Desktop Perception
+- Start/stop a **desktop session** capturing **screen, mic, and system audio**
+- Stream **live context** and store **episodic session memory**
+- Run **real-time alerts/triggers** on what's spoken and what's happening on screen
+- Produce **session summaries**, a searchable timeline, and **playable evidence links**
-## Setup
+## 2) Video ingest + stream
+- Ingest a **file or URL** and return a **playable web stream link**
+- Transcode/normalize: **codec, bitrate, fps, resolution, aspect ratio**
+
+## 3) Index + search (timestamps + evidence)
+- Build **visual**, **spoken**, and **keyword** indexes
+- Search and return exact moments with **timestamps** and **playable evidence**
+- Auto-create **clips** from search results
+
+## 4) Timeline editing + generation
+- Subtitles: **generate**, **translate**, **burn-in**
+- Overlays: **text/image/branding**, motion captions
+- Audio: **background music**, **voiceover**, **dubbing**
+- Programmatic composition and exports via **timeline operations**
+
+## 5) Live streams (RTSP) + monitoring
+- Connect **RTSP/live feeds**
+- Run **real-time visual and spoken understanding** and emit **events/alerts** for monitoring workflows
+
+---
+
+## Common inputs
+- Local **file path**, public **URL**, or **RTSP URL**
+- Desktop capture request: **start / stop / summarize session**
+- Desired operations: get context for understanding, transcode spec, index spec, search query, clip ranges, timeline edits, alert rules
+
+## Common outputs
+- **Stream URL**
+- Search results with **timestamps** and **evidence links**
+- Generated assets: subtitles, audio, images, clips
+- **Event/alert payloads** for live streams
+- Desktop **session summaries** and memory entries
+
+---
+
+## Canonical prompts (examples)
+- "Start desktop capture and alert when a password field appears."
+- "Record my session and produce an actionable summary when it ends."
+- "Ingest this file and return a playable stream link."
+- "Index this folder and find every scene with people, return timestamps."
+- "Generate subtitles, burn them in, and add light background music."
+- "Connect this RTSP URL and alert when a person enters the zone."
+
+
+## Running Python code
+
+Before running any VideoDB code, change to the project directory and load environment variables:
+
+```python
+from dotenv import load_dotenv
+load_dotenv(".env")
+
+import videodb
+conn = videodb.connect()
+```
+
+This reads `VIDEO_DB_API_KEY` from:
+1. Environment (if already exported)
+2. Project's `.env` file in current directory
+
+If the key is missing, `videodb.connect()` raises `AuthenticationError` automatically.
+
+Do NOT write a script file when a short inline command works.
+
+When writing inline Python (`python -c "..."`), always use properly formatted code — use semicolons to separate statements and keep it readable. For anything longer than ~3 statements, use a heredoc instead:
```bash
-# Install the skill
-npx skills add video-db/skills
+python << 'EOF'
+from dotenv import load_dotenv
+load_dotenv(".env")
-# Or setup manually
+import videodb
+conn = videodb.connect()
+coll = conn.get_collection()
+print(f"Videos: {len(coll.get_videos())}")
+EOF
+```
+
+## Setup
+
+When the user asks to "setup videodb" or similar:
+
+### 1. Install SDK
+
+```bash
pip install "videodb[capture]" python-dotenv
-export VIDEO_DB_API_KEY=sk-xxx
```
-Run `/videodb setup` inside your agent for guided setup ($20 free credits, no credit card).
+If `videodb[capture]` fails on Linux, install without the capture extra:
+
+```bash
+pip install videodb python-dotenv
+```
+
+### 2. Configure API key
+
+The user must set `VIDEO_DB_API_KEY` using **either** method:
+
+- **Export in terminal** (before starting Claude): `export VIDEO_DB_API_KEY=your-key`
+- **Project `.env` file**: Save `VIDEO_DB_API_KEY=your-key` in the project's `.env` file
-## Core Patterns
+Get a free API key at https://console.videodb.io (50 free uploads, no credit card).
-### Upload and Process
+**Do NOT** read, write, or handle the API key yourself. Always let the user set it.
+
+## Quick Reference
+
+### Upload media
```python
-import videodb
+# URL
+video = coll.upload(url="https://example.com/video.mp4")
-conn = videodb.connect()
-video = conn.upload(url="https://www.youtube.com/watch?v=VIDEO_ID")
+# YouTube
+video = coll.upload(url="https://www.youtube.com/watch?v=VIDEO_ID")
+
+# Local file
+video = coll.upload(file_path="/path/to/video.mp4")
+```
+
+### Transcript + subtitle
+
+```python
+# force=True skips the error if the video is already indexed
+video.index_spoken_words(force=True)
+text = video.get_transcript_text()
+stream_url = video.add_subtitle()
+```
+
+### Search inside videos
+
+```python
+from videodb.exceptions import InvalidRequestError
+
+video.index_spoken_words(force=True)
-transcript = video.get_transcript()
-for entry in transcript:
- print(f"[{entry['start']:.1f}s] {entry['text']}")
+# search() raises InvalidRequestError when no results are found.
+# Always wrap in try/except and treat "No results found" as empty.
+try:
+ results = video.search("product demo")
+ shots = results.get_shots()
+ stream_url = results.compile()
+except InvalidRequestError as e:
+ if "No results found" in str(e):
+ shots = []
+ else:
+ raise
```
-### Search Across Videos
+### Scene search
```python
-# Index for semantic search
-video.index_spoken_words()
+import re
+from videodb import SearchType, IndexType, SceneExtractionType
+from videodb.exceptions import InvalidRequestError
+
+# index_scenes() has no force parameter — it raises an error if a scene
+# index already exists. Extract the existing index ID from the error.
+try:
+ scene_index_id = video.index_scenes(
+ extraction_type=SceneExtractionType.shot_based,
+ prompt="Describe the visual content in this scene.",
+ )
+except Exception as e:
+ match = re.search(r"id\s+([a-f0-9]+)", str(e))
+ if match:
+ scene_index_id = match.group(1)
+ else:
+ raise
-# Search by what was said
-results = video.search("product demo")
-for r in results:
- print(f"{r.start:.1f}s - {r.end:.1f}s: {r.text}")
+# Use score_threshold to filter low-relevance noise (recommended: 0.3+)
+try:
+ results = video.search(
+ query="person writing on a whiteboard",
+ search_type=SearchType.semantic,
+ index_type=IndexType.scene,
+ scene_index_id=scene_index_id,
+ score_threshold=0.3,
+ )
+ shots = results.get_shots()
+ stream_url = results.compile()
+except InvalidRequestError as e:
+ if "No results found" in str(e):
+ shots = []
+ else:
+ raise
```
-### Timeline Editing
+### Timeline editing
+
+**Important:** Always validate timestamps before building a timeline:
+- `start` must be >= 0 (negative values are silently accepted but produce broken output)
+- `start` must be < `end`
+- `end` must be <= `video.length`
```python
-from videodb import Timeline, VideoAsset, AudioAsset
+from videodb.timeline import Timeline
+from videodb.asset import VideoAsset, TextAsset, TextStyle
timeline = Timeline(conn)
-asset = VideoAsset(asset_id=video.id, start=10, end=30)
-timeline.add_inline(asset)
+timeline.add_inline(VideoAsset(asset_id=video.id, start=10, end=30))
+timeline.add_overlay(0, TextAsset(text="The End", duration=3, style=TextStyle(fontsize=36)))
+stream_url = timeline.generate_stream()
+```
+
+### Transcode video (resolution / quality change)
+
+```python
+from videodb import TranscodeMode, VideoConfig, AudioConfig
+
+# Change resolution, quality, or aspect ratio server-side
+job_id = conn.transcode(
+ source="https://example.com/video.mp4",
+ callback_url="https://example.com/webhook",
+ mode=TranscodeMode.economy,
+ video_config=VideoConfig(resolution=720, quality=23, aspect_ratio="16:9"),
+ audio_config=AudioConfig(mute=False),
+)
+```
+
+### Reframe aspect ratio (for social platforms)
+
+**Warning:** `reframe()` is a slow server-side operation. For long videos it can take
+several minutes and may time out. Best practices:
+- Always limit to a short segment using `start`/`end` when possible
+- For full-length videos, use `callback_url` for async processing
+- Trim the video on a `Timeline` first, then reframe the shorter result
+
+```python
+from videodb import ReframeMode
+
+# Always prefer reframing a short segment:
+reframed = video.reframe(start=0, end=60, target="vertical", mode=ReframeMode.smart)
+
+# Async reframe for full-length videos (returns None, result via webhook):
+video.reframe(target="vertical", callback_url="https://example.com/webhook")
+
+# Presets: "vertical" (9:16), "square" (1:1), "landscape" (16:9)
+reframed = video.reframe(start=0, end=60, target="square")
+
+# Custom dimensions
+reframed = video.reframe(start=0, end=60, target={"width": 1280, "height": 720})
+```
+
+### Generative media
+
+```python
+image = coll.generate_image(
+ prompt="a sunset over mountains",
+ aspect_ratio="16:9",
+)
+```
+
+## Error handling
+
+```python
+from videodb.exceptions import AuthenticationError, InvalidRequestError
+
+try:
+ conn = videodb.connect()
+except AuthenticationError:
+ print("Check your VIDEO_DB_API_KEY")
-stream = timeline.generate_stream()
-print(stream) # Playable HLS link
+try:
+ video = coll.upload(url="https://example.com/video.mp4")
+except InvalidRequestError as e:
+ print(f"Upload failed: {e}")
```
-### AI Media Generation
+### Common pitfalls
+
+| Scenario | Error message | Solution |
+|----------|--------------|----------|
+| Indexing an already-indexed video | `Spoken word index for video already exists` | Use `video.index_spoken_words(force=True)` to skip if already indexed |
+| Scene index already exists | `Scene index with id XXXX already exists` | Extract the existing `scene_index_id` from the error with `re.search(r"id\s+([a-f0-9]+)", str(e))` |
+| Search finds no matches | `InvalidRequestError: No results found` | Catch the exception and treat as empty results (`shots = []`) |
+| Reframe times out | Blocks indefinitely on long videos | Use `start`/`end` to limit segment, or pass `callback_url` for async |
+| Negative timestamps on Timeline | Silently produces broken stream | Always validate `start >= 0` before creating `VideoAsset` |
+| `generate_video()` / `create_collection()` fails | `Operation not allowed` or `maximum limit` | Plan-gated features — inform the user about plan limits |
+
+## Additional docs
+
+Reference documentation is in the `reference/` directory adjacent to this SKILL.md file. Use the Glob tool to locate it if needed.
+
+- [reference/api-reference.md](reference/api-reference.md) - Complete VideoDB Python SDK API reference
+- [reference/search.md](reference/search.md) - In-depth guide to video search (spoken word and scene-based)
+- [reference/editor.md](reference/editor.md) - Timeline editing, assets, and composition
+- [reference/streaming.md](reference/streaming.md) - HLS streaming and instant playback
+- [reference/generative.md](reference/generative.md) - AI-powered media generation (images, video, audio)
+- [reference/rtstream.md](reference/rtstream.md) - Live stream ingestion workflow (RTSP/RTMP)
+- [reference/rtstream-reference.md](reference/rtstream-reference.md) - RTStream SDK methods and AI pipelines
+- [reference/capture.md](reference/capture.md) - Desktop capture workflow
+- [reference/capture-reference.md](reference/capture-reference.md) - Capture SDK and WebSocket events
+- [reference/use-cases.md](reference/use-cases.md) - Common video processing patterns and examples
+
+## Screen Recording (Desktop Capture)
+
+Use `ws_listener.py` to capture WebSocket events during recording sessions. Desktop capture supports **macOS** only.
+
+### Quick Start
+
+1. **Start listener**: `python scripts/ws_listener.py &`
+2. **Get WebSocket ID**: `cat /tmp/videodb_ws_id`
+3. **Run capture code** (see reference/capture.md for full workflow)
+4. **Events written to**: `/tmp/videodb_events.jsonl`
+
+### Query Events
```python
-audio = conn.generate_audio(text="Upbeat background music", duration=30)
-image = conn.generate_image(prompt="Title card: Welcome to the Demo")
+import json
+events = [json.loads(l) for l in open("/tmp/videodb_events.jsonl")]
+
+# Get all transcripts
+transcripts = [e["data"]["text"] for e in events if e.get("channel") == "transcript"]
+
+# Get visual descriptions from last 5 minutes
+import time
+cutoff = time.time() - 300
+recent_visual = [e for e in events
+ if e.get("channel") == "visual_index" and e["unix_ts"] > cutoff]
```
-## Capabilities
-
-| Capability | What It Does |
-|---|---|
-| Upload | YouTube, URLs, local files |
-| Search | Speech-based and scene-based |
-| Transcripts | Timestamped, multi-language |
-| Edit | Trim, combine, multi-timeline |
-| Subtitles | Auto-generate, custom styling |
-| AI Generate | Images, video, music, SFX, voiceover |
-| Capture | Screen + audio, real-time |
-| Transcode | Resolution, codec, aspect ratio |
-| Stream | HLS playable links |
-
-## Best Practices
-
-- Always verify SDK connection before operations: `conn.get_collection()`
-- Use `video.index_spoken_words()` before searching — indexing is required once per video
-- For scene search, use `video.index_scenes()` — this processes visual frames
-- Timeline edits produce new streams; the original video is never modified
-- AI generation is async — poll status or use callbacks for long operations
-- Store `VIDEO_DB_API_KEY` in `.env`, not hardcoded
+### Utility Scripts
+
+- [scripts/ws_listener.py](scripts/ws_listener.py) - WebSocket event listener (dumps to JSONL)
+
+For complete capture workflow, see [reference/capture.md](reference/capture.md).
+
+
+**Do not use ffmpeg, moviepy, or local encoding tools** when VideoDB supports the operation. The following are all handled server-side by VideoDB — trimming, combining clips, overlaying audio or music, adding subtitles, text/image overlays, transcoding, resolution changes, aspect-ratio conversion, resizing for platform requirements, transcription, and media generation. Only fall back to local tools for operations listed under Limitations in reference/editor.md (transitions, speed changes, crop/zoom, colour grading, volume mixing).
+
+### When to use what
+
+| Problem | VideoDB solution |
+|---------|-----------------|
+| Platform rejects video aspect ratio or resolution | `video.reframe()` or `conn.transcode()` with `VideoConfig` |
+| Need to resize video for Twitter/Instagram/TikTok | `video.reframe(target="vertical")` or `target="square"` |
+| Need to change resolution (e.g. 1080p → 720p) | `conn.transcode()` with `VideoConfig(resolution=720)` |
+| Need to overlay audio/music on video | `AudioAsset` on a `Timeline` |
+| Need to add subtitles | `video.add_subtitle()` or `CaptionAsset` |
+| Need to combine/trim clips | `VideoAsset` on a `Timeline` |
+| Need to generate voiceover, music, or SFX | `coll.generate_voice()`, `generate_music()`, `generate_sound_effect()` |
## Repository
https://github.com/video-db/skills
+
+**Maintained By:** [VideoDB](https://github.com/video-db)
diff --git a/skills/videodb-skills/reference/api-reference.md b/skills/videodb-skills/reference/api-reference.md
new file mode 100644
index 000000000..90f4ce10c
--- /dev/null
+++ b/skills/videodb-skills/reference/api-reference.md
@@ -0,0 +1,548 @@
+# Complete API Reference
+
+## Connection
+
+```python
+import videodb
+
+conn = videodb.connect(
+ api_key="your-api-key", # or set VIDEO_DB_API_KEY env var
+ base_url=None, # custom API endpoint (optional)
+)
+```
+
+**Returns:** `Connection` object
+
+### Connection Methods
+
+| Method | Returns | Description |
+|--------|---------|-------------|
+| `conn.get_collection(collection_id="default")` | `Collection` | Get collection (default if no ID) |
+| `conn.get_collections()` | `list[Collection]` | List all collections |
+| `conn.create_collection(name, description, is_public=False)` | `Collection` | Create new collection |
+| `conn.update_collection(id, name, description)` | `Collection` | Update a collection |
+| `conn.check_usage()` | `dict` | Get account usage stats |
+| `conn.upload(source, media_type, name, ...)` | `Video\|Audio\|Image` | Upload to default collection |
+| `conn.record_meeting(meeting_url, bot_name, ...)` | `Meeting` | Record a meeting |
+| `conn.create_capture_session(...)` | `CaptureSession` | Create a capture session (see [capture-reference.md](capture-reference.md)) |
+| `conn.youtube_search(query, result_threshold, duration)` | `list[dict]` | Search YouTube |
+| `conn.transcode(source, callback_url, mode, ...)` | `str` | Transcode video (returns job ID) |
+| `conn.get_transcode_details(job_id)` | `dict` | Get transcode job status and details |
+| `conn.connect_websocket(collection_id)` | `WebSocketConnection` | Connect to WebSocket (see [capture-reference.md](capture-reference.md)) |
+
+### Transcode
+
+Transcode a video from a URL with custom resolution, quality, and audio settings. Processing happens server-side — no local ffmpeg required.
+
+```python
+from videodb import TranscodeMode, VideoConfig, AudioConfig
+
+job_id = conn.transcode(
+ source="https://example.com/video.mp4",
+ callback_url="https://example.com/webhook",
+ mode=TranscodeMode.economy,
+ video_config=VideoConfig(resolution=720, quality=23),
+ audio_config=AudioConfig(mute=False),
+)
+```
+
+#### transcode Parameters
+
+| Parameter | Type | Default | Description |
+|-----------|------|---------|-------------|
+| `source` | `str` | required | URL of the video to transcode (preferably a downloadable URL) |
+| `callback_url` | `str` | required | URL to receive the callback when transcoding completes |
+| `mode` | `TranscodeMode` | `TranscodeMode.economy` | Transcoding speed: `economy` or `lightning` |
+| `video_config` | `VideoConfig` | `VideoConfig()` | Video encoding settings |
+| `audio_config` | `AudioConfig` | `AudioConfig()` | Audio encoding settings |
+
+Returns a job ID (`str`). Use `conn.get_transcode_details(job_id)` to check job status.
+
+```python
+details = conn.get_transcode_details(job_id)
+```
+
+#### VideoConfig
+
+```python
+from videodb import VideoConfig, ResizeMode
+
+config = VideoConfig(
+ resolution=720, # Target resolution height (e.g. 480, 720, 1080)
+ quality=23, # Encoding quality (lower = better, default 23)
+ framerate=30, # Target framerate
+ aspect_ratio="16:9", # Target aspect ratio
+ resize_mode=ResizeMode.crop, # How to fit: crop, fit, or pad
+)
+```
+
+| Field | Type | Default | Description |
+|-------|------|---------|-------------|
+| `resolution` | `int\|None` | `None` | Target resolution height in pixels |
+| `quality` | `int` | `23` | Encoding quality (lower = higher quality) |
+| `framerate` | `int\|None` | `None` | Target framerate |
+| `aspect_ratio` | `str\|None` | `None` | Target aspect ratio (e.g. `"16:9"`, `"9:16"`) |
+| `resize_mode` | `str` | `ResizeMode.crop` | Resize strategy: `crop`, `fit`, or `pad` |
+
+#### AudioConfig
+
+```python
+from videodb import AudioConfig
+
+config = AudioConfig(mute=False)
+```
+
+| Field | Type | Default | Description |
+|-------|------|---------|-------------|
+| `mute` | `bool` | `False` | Mute the audio track |
+
+## Collections
+
+```python
+coll = conn.get_collection()
+```
+
+### Collection Methods
+
+| Method | Returns | Description |
+|--------|---------|-------------|
+| `coll.get_videos()` | `list[Video]` | List all videos |
+| `coll.get_video(video_id)` | `Video` | Get specific video |
+| `coll.get_audios()` | `list[Audio]` | List all audios |
+| `coll.get_audio(audio_id)` | `Audio` | Get specific audio |
+| `coll.get_images()` | `list[Image]` | List all images |
+| `coll.get_image(image_id)` | `Image` | Get specific image |
+| `coll.upload(url=None, file_path=None, media_type=None, name=None)` | `Video\|Audio\|Image` | Upload media |
+| `coll.search(query, search_type, index_type, score_threshold, namespace, scene_index_id, ...)` | `SearchResult` | Search across collection (semantic only; keyword and scene search raise `NotImplementedError`) |
+| `coll.generate_image(prompt, aspect_ratio="1:1")` | `Image` | Generate image with AI |
+| `coll.generate_video(prompt, duration=5)` | `Video` | Generate video with AI |
+| `coll.generate_music(prompt, duration=5)` | `Audio` | Generate music with AI |
+| `coll.generate_sound_effect(prompt, duration=2)` | `Audio` | Generate sound effect |
+| `coll.generate_voice(text, voice_name="Default")` | `Audio` | Generate speech from text |
+| `coll.generate_text(prompt, model_name="basic", response_type="text")` | `dict` | LLM text generation — access result via `["output"]` |
+| `coll.dub_video(video_id, language_code)` | `Video` | Dub video into another language |
+| `coll.record_meeting(meeting_url, bot_name, ...)` | `Meeting` | Record a live meeting |
+| `coll.create_capture_session(...)` | `CaptureSession` | Create a capture session (see [capture-reference.md](capture-reference.md)) |
+| `coll.get_capture_session(...)` | `CaptureSession` | Retrieve capture session (see [capture-reference.md](capture-reference.md)) |
+| `coll.connect_rtstream(url, name, ...)` | `RTStream` | Connect to a live stream (see [rtstream-reference.md](rtstream-reference.md)) |
+| `coll.make_public()` | `None` | Make collection public |
+| `coll.make_private()` | `None` | Make collection private |
+| `coll.delete_video(video_id)` | `None` | Delete a video |
+| `coll.delete_audio(audio_id)` | `None` | Delete an audio |
+| `coll.delete_image(image_id)` | `None` | Delete an image |
+| `coll.delete()` | `None` | Delete the collection |
+
+### Upload Parameters
+
+```python
+video = coll.upload(
+ url=None, # Remote URL (HTTP, YouTube)
+ file_path=None, # Local file path
+ media_type=None, # "video", "audio", or "image" (auto-detected if omitted)
+ name=None, # Custom name for the media
+ description=None, # Description
+ callback_url=None, # Webhook URL for async notification
+)
+```
+
+## Video Object
+
+```python
+video = coll.get_video(video_id)
+```
+
+### Video Properties
+
+| Property | Type | Description |
+|----------|------|-------------|
+| `video.id` | `str` | Unique video ID |
+| `video.collection_id` | `str` | Parent collection ID |
+| `video.name` | `str` | Video name |
+| `video.description` | `str` | Video description |
+| `video.length` | `float` | Duration in seconds |
+| `video.stream_url` | `str` | Default stream URL |
+| `video.player_url` | `str` | Player embed URL |
+| `video.thumbnail_url` | `str` | Thumbnail URL |
+
+### Video Methods
+
+| Method | Returns | Description |
+|--------|---------|-------------|
+| `video.generate_stream(timeline=None)` | `str` | Generate stream URL (optional timeline of `[(start, end)]` tuples) |
+| `video.play()` | `str` | Open stream in browser, returns player URL |
+| `video.index_spoken_words(language_code=None, force=False)` | `None` | Index speech for search. Use `force=True` to skip if already indexed. |
+| `video.index_scenes(extraction_type, prompt, extraction_config, metadata, model_name, name, scenes, callback_url)` | `str` | Index visual scenes (returns scene_index_id) |
+| `video.index_visuals(prompt, batch_config, ...)` | `str` | Index visuals (returns scene_index_id) |
+| `video.index_audio(prompt, model_name, ...)` | `str` | Index audio with LLM (returns scene_index_id) |
+| `video.get_transcript(start=None, end=None)` | `list[dict]` | Get timestamped transcript |
+| `video.get_transcript_text(start=None, end=None)` | `str` | Get full transcript text |
+| `video.generate_transcript(force=None)` | `dict` | Generate transcript |
+| `video.translate_transcript(language, additional_notes)` | `list[dict]` | Translate transcript |
+| `video.search(query, search_type, index_type, filter, **kwargs)` | `SearchResult` | Search within video |
+| `video.add_subtitle(style=SubtitleStyle())` | `str` | Add subtitles (returns stream URL) |
+| `video.generate_thumbnail(time=None)` | `str\|Image` | Generate thumbnail |
+| `video.get_thumbnails()` | `list[Image]` | Get all thumbnails |
+| `video.extract_scenes(extraction_type, extraction_config)` | `SceneCollection` | Extract scenes |
+| `video.reframe(start, end, target, mode, callback_url)` | `Video\|None` | Reframe video aspect ratio |
+| `video.clip(prompt, content_type, model_name)` | `str` | Generate clip from prompt (returns stream URL) |
+| `video.insert_video(video, timestamp)` | `str` | Insert video at timestamp |
+| `video.download(name=None)` | `dict` | Download the video |
+| `video.delete()` | `None` | Delete the video |
+
+### Reframe
+
+Convert a video to a different aspect ratio with optional smart object tracking. Processing is server-side.
+
+> **Warning:** Reframe is a slow server-side operation. It can take several minutes for long videos and may time out. Always use `start`/`end` to limit the segment, or pass `callback_url` for async processing.
+
+```python
+from videodb import ReframeMode
+
+# Always prefer short segments to avoid timeouts:
+reframed = video.reframe(start=0, end=60, target="vertical", mode=ReframeMode.smart)
+
+# Async reframe for full-length videos (returns None, result via webhook):
+video.reframe(target="vertical", callback_url="https://example.com/webhook")
+
+# Custom dimensions
+reframed = video.reframe(start=0, end=60, target={"width": 1080, "height": 1080})
+```
+
+#### reframe Parameters
+
+| Parameter | Type | Default | Description |
+|-----------|------|---------|-------------|
+| `start` | `float\|None` | `None` | Start time in seconds (None = beginning) |
+| `end` | `float\|None` | `None` | End time in seconds (None = end of video) |
+| `target` | `str\|dict` | `"vertical"` | Preset string (`"vertical"`, `"square"`, `"landscape"`) or `{"width": int, "height": int}` |
+| `mode` | `str` | `ReframeMode.smart` | `"simple"` (centre crop) or `"smart"` (object tracking) |
+| `callback_url` | `str\|None` | `None` | Webhook URL for async notification |
+
+Returns a `Video` object when no `callback_url` is provided, `None` otherwise.
+
+## Audio Object
+
+```python
+audio = coll.get_audio(audio_id)
+```
+
+### Audio Properties
+
+| Property | Type | Description |
+|----------|------|-------------|
+| `audio.id` | `str` | Unique audio ID |
+| `audio.collection_id` | `str` | Parent collection ID |
+| `audio.name` | `str` | Audio name |
+| `audio.length` | `float` | Duration in seconds |
+
+### Audio Methods
+
+| Method | Returns | Description |
+|--------|---------|-------------|
+| `audio.generate_url()` | `str` | Generate signed URL for playback |
+| `audio.get_transcript(start=None, end=None)` | `list[dict]` | Get timestamped transcript |
+| `audio.get_transcript_text(start=None, end=None)` | `str` | Get full transcript text |
+| `audio.generate_transcript(force=None)` | `dict` | Generate transcript |
+| `audio.delete()` | `None` | Delete the audio |
+
+## Image Object
+
+```python
+image = coll.get_image(image_id)
+```
+
+### Image Properties
+
+| Property | Type | Description |
+|----------|------|-------------|
+| `image.id` | `str` | Unique image ID |
+| `image.collection_id` | `str` | Parent collection ID |
+| `image.name` | `str` | Image name |
+| `image.url` | `str\|None` | Image URL (may be `None` for generated images — use `generate_url()` instead) |
+
+### Image Methods
+
+| Method | Returns | Description |
+|--------|---------|-------------|
+| `image.generate_url()` | `str` | Generate signed URL |
+| `image.delete()` | `None` | Delete the image |
+
+## Timeline & Editor
+
+### Timeline
+
+```python
+from videodb.timeline import Timeline
+
+timeline = Timeline(conn)
+```
+
+| Method | Returns | Description |
+|--------|---------|-------------|
+| `timeline.add_inline(asset)` | `None` | Add `VideoAsset` sequentially on main track |
+| `timeline.add_overlay(start, asset)` | `None` | Overlay `AudioAsset`, `ImageAsset`, or `TextAsset` at timestamp |
+| `timeline.generate_stream()` | `str` | Compile and get stream URL |
+
+### Asset Types
+
+#### VideoAsset
+
+```python
+from videodb.asset import VideoAsset
+
+asset = VideoAsset(
+ asset_id=video.id,
+ start=0, # trim start (seconds)
+ end=None, # trim end (seconds, None = full)
+)
+```
+
+#### AudioAsset
+
+```python
+from videodb.asset import AudioAsset
+
+asset = AudioAsset(
+ asset_id=audio.id,
+ start=0,
+ end=None,
+ disable_other_tracks=True, # mute original audio when True
+ fade_in_duration=0, # seconds (max 5)
+ fade_out_duration=0, # seconds (max 5)
+)
+```
+
+#### ImageAsset
+
+```python
+from videodb.asset import ImageAsset
+
+asset = ImageAsset(
+ asset_id=image.id,
+ duration=None, # display duration (seconds)
+ width=100, # display width
+ height=100, # display height
+ x=80, # horizontal position (px from left)
+ y=20, # vertical position (px from top)
+)
+```
+
+#### TextAsset
+
+```python
+from videodb.asset import TextAsset, TextStyle
+
+asset = TextAsset(
+ text="Hello World",
+ duration=5,
+ style=TextStyle(
+ fontsize=24,
+ fontcolor="black",
+ boxcolor="white", # background box colour
+ alpha=1.0,
+ font="Sans",
+ text_align="T", # text alignment within box
+ ),
+)
+```
+
+#### CaptionAsset (Editor API)
+
+CaptionAsset belongs to the Editor API, which has its own Timeline, Track, and Clip system:
+
+```python
+from videodb.editor import CaptionAsset, FontStyling
+
+asset = CaptionAsset(
+ src="auto", # "auto" or base64 ASS string
+ font=FontStyling(name="Clear Sans", size=30),
+ primary_color="&H00FFFFFF",
+)
+```
+
+See [editor.md](editor.md#caption-overlays) for full CaptionAsset usage with the Editor API.
+
+## Video Search Parameters
+
+```python
+results = video.search(
+ query="your query",
+ search_type=SearchType.semantic, # semantic, keyword, or scene
+ index_type=IndexType.spoken_word, # spoken_word or scene
+ result_threshold=None, # max number of results
+ score_threshold=None, # minimum relevance score
+ dynamic_score_percentage=None, # percentage of dynamic score
+ scene_index_id=None, # target a specific scene index (pass via **kwargs)
+ filter=[], # metadata filters for scene search
+)
+```
+
+> **Note:** `filter` is an explicit named parameter in `video.search()`. `scene_index_id` is passed through `**kwargs` to the API.
+
+> **Important:** `video.search()` raises `InvalidRequestError` with message `"No results found"` when there are no matches. Always wrap search calls in try/except. For scene search, use `score_threshold=0.3` or higher to filter low-relevance noise.
+
+For scene search, use `search_type=SearchType.semantic` with `index_type=IndexType.scene`. Pass `scene_index_id` when targeting a specific scene index. See [search.md](search.md) for details.
+
+## SearchResult Object
+
+```python
+results = video.search("query", search_type=SearchType.semantic)
+```
+
+| Method | Returns | Description |
+|--------|---------|-------------|
+| `results.get_shots()` | `list[Shot]` | Get list of matching segments |
+| `results.compile()` | `str` | Compile all shots into a stream URL |
+| `results.play()` | `str` | Open compiled stream in browser |
+
+### Shot Properties
+
+| Property | Type | Description |
+|----------|------|-------------|
+| `shot.video_id` | `str` | Source video ID |
+| `shot.video_length` | `float` | Source video duration |
+| `shot.video_title` | `str` | Source video title |
+| `shot.start` | `float` | Start time (seconds) |
+| `shot.end` | `float` | End time (seconds) |
+| `shot.text` | `str` | Matched text content |
+| `shot.search_score` | `float` | Search relevance score |
+
+| Method | Returns | Description |
+|--------|---------|-------------|
+| `shot.generate_stream()` | `str` | Stream this specific shot |
+| `shot.play()` | `str` | Open shot stream in browser |
+
+## Meeting Object
+
+```python
+meeting = coll.record_meeting(
+ meeting_url="https://meet.google.com/...",
+ bot_name="Bot",
+ callback_url=None, # Webhook URL for status updates
+ callback_data=None, # Optional dict passed through to callbacks
+ time_zone="UTC", # Time zone for the meeting
+)
+```
+
+### Meeting Properties
+
+| Property | Type | Description |
+|----------|------|-------------|
+| `meeting.id` | `str` | Unique meeting ID |
+| `meeting.collection_id` | `str` | Parent collection ID |
+| `meeting.status` | `str` | Current status |
+| `meeting.video_id` | `str` | Recorded video ID (after completion) |
+| `meeting.bot_name` | `str` | Bot name |
+| `meeting.meeting_title` | `str` | Meeting title |
+| `meeting.meeting_url` | `str` | Meeting URL |
+| `meeting.speaker_timeline` | `dict` | Speaker timeline data |
+| `meeting.is_active` | `bool` | True if initializing or processing |
+| `meeting.is_completed` | `bool` | True if done |
+
+### Meeting Methods
+
+| Method | Returns | Description |
+|--------|---------|-------------|
+| `meeting.refresh()` | `Meeting` | Refresh data from server |
+| `meeting.wait_for_status(target_status, timeout=14400, interval=120)` | `bool` | Poll until status reached |
+
+## RTStream & Capture
+
+For RTStream (live ingestion, indexing, transcription), see [rtstream-reference.md](rtstream-reference.md).
+
+For capture sessions (desktop recording, CaptureClient, channels), see [capture-reference.md](capture-reference.md).
+
+## Enums & Constants
+
+### SearchType
+
+```python
+from videodb import SearchType
+
+SearchType.semantic # Natural language semantic search
+SearchType.keyword # Exact keyword matching
+SearchType.scene # Visual scene search (may require paid plan)
+SearchType.llm # LLM-powered search
+```
+
+### SceneExtractionType
+
+```python
+from videodb import SceneExtractionType
+
+SceneExtractionType.shot_based # Automatic shot boundary detection
+SceneExtractionType.time_based # Fixed time interval extraction
+SceneExtractionType.transcript # Transcript-based scene extraction
+```
+
+### SubtitleStyle
+
+```python
+from videodb import SubtitleStyle
+
+style = SubtitleStyle(
+ font_name="Arial",
+ font_size=18,
+ primary_colour="&H00FFFFFF",
+ bold=False,
+ # ... see SubtitleStyle for all options
+)
+video.add_subtitle(style=style)
+```
+
+### SubtitleAlignment & SubtitleBorderStyle
+
+```python
+from videodb import SubtitleAlignment, SubtitleBorderStyle
+```
+
+### TextStyle
+
+```python
+from videodb import TextStyle
+# or: from videodb.asset import TextStyle
+
+style = TextStyle(
+ fontsize=24,
+ fontcolor="black",
+ boxcolor="white",
+ font="Sans",
+ text_align="T",
+ alpha=1.0,
+)
+```
+
+### Other Constants
+
+```python
+from videodb import (
+ IndexType, # spoken_word, scene
+ MediaType, # video, audio, image
+ Segmenter, # word, sentence, time
+ SegmentationType, # sentence, llm
+ TranscodeMode, # economy, lightning
+ ResizeMode, # crop, fit, pad
+ ReframeMode, # simple, smart
+ RTStreamChannelType,
+)
+```
+
+## Exceptions
+
+```python
+from videodb.exceptions import (
+ AuthenticationError, # Invalid or missing API key
+ InvalidRequestError, # Bad parameters or malformed request
+ RequestTimeoutError, # Request timed out
+ SearchError, # Search operation failure (e.g. not indexed)
+ VideodbError, # Base exception for all VideoDB errors
+)
+```
+
+| Exception | Common Cause |
+|-----------|-------------|
+| `AuthenticationError` | Missing or invalid `VIDEO_DB_API_KEY` |
+| `InvalidRequestError` | Invalid URL, unsupported format, bad parameters |
+| `RequestTimeoutError` | Server took too long to respond |
+| `SearchError` | Searching before indexing, invalid search type |
+| `VideodbError` | Server errors, network issues, generic failures |
diff --git a/skills/videodb-skills/reference/capture-reference.md b/skills/videodb-skills/reference/capture-reference.md
new file mode 100644
index 000000000..aef55d41e
--- /dev/null
+++ b/skills/videodb-skills/reference/capture-reference.md
@@ -0,0 +1,386 @@
+# Capture Reference
+
+Code-level details for VideoDB capture sessions. For workflow guide, see [capture.md](capture.md).
+
+---
+
+## WebSocket Events
+
+Real-time events from capture sessions and AI pipelines. No webhooks or polling required.
+
+Use [scripts/ws_listener.py](../scripts/ws_listener.py) to connect and dump events to `/tmp/videodb_events.jsonl`.
+
+### Event Channels
+
+| Channel | Source | Content |
+|---------|--------|---------|
+| `capture_session` | Session lifecycle | Status changes |
+| `transcript` | `start_transcript()` | Speech-to-text |
+| `visual_index` / `scene_index` | `index_visuals()` | Visual analysis |
+| `audio_index` | `index_audio()` | Audio analysis |
+| `alert` | `create_alert()` | Alert notifications |
+
+### Session Lifecycle Events
+
+| Event | Status | Key Data |
+|-------|--------|----------|
+| `capture_session.created` | `created` | — |
+| `capture_session.starting` | `starting` | — |
+| `capture_session.active` | `active` | `rtstreams[]` |
+| `capture_session.stopping` | `stopping` | — |
+| `capture_session.stopped` | `stopped` | — |
+| `capture_session.exported` | `exported` | `exported_video_id`, `stream_url`, `player_url` |
+| `capture_session.failed` | `failed` | `error` |
+
+### Event Structures
+
+**Transcript event:**
+```json
+{
+ "channel": "transcript",
+ "rtstream_id": "rts-xxx",
+ "rtstream_name": "mic:default",
+ "data": {
+ "text": "Let's schedule the meeting for Thursday",
+ "is_final": true,
+ "start": 1710000001234,
+ "end": 1710000002345
+ }
+}
+```
+
+**Visual index event:**
+```json
+{
+ "channel": "visual_index",
+ "rtstream_id": "rts-xxx",
+ "rtstream_name": "display:1",
+ "data": {
+ "text": "User is viewing a Slack conversation with 3 unread messages",
+ "start": 1710000012340,
+ "end": 1710000018900
+ }
+}
+```
+
+**Audio index event:**
+```json
+{
+ "channel": "audio_index",
+ "rtstream_id": "rts-xxx",
+ "rtstream_name": "mic:default",
+ "data": {
+ "text": "Discussion about scheduling a team meeting",
+ "start": 1710000021500,
+ "end": 1710000029200
+ }
+}
+```
+
+**Session active event:**
+```json
+{
+ "event": "capture_session.active",
+ "capture_session_id": "cap-xxx",
+ "status": "active",
+ "data": {
+ "rtstreams": [
+ { "rtstream_id": "rts-1", "name": "mic:default", "media_types": ["audio"] },
+ { "rtstream_id": "rts-2", "name": "system_audio:default", "media_types": ["audio"] },
+ { "rtstream_id": "rts-3", "name": "display:1", "media_types": ["video"] }
+ ]
+ }
+}
+```
+
+**Session exported event:**
+```json
+{
+ "event": "capture_session.exported",
+ "capture_session_id": "cap-xxx",
+ "status": "exported",
+ "data": {
+ "exported_video_id": "v_xyz789",
+ "stream_url": "https://stream.videodb.io/...",
+ "player_url": "https://console.videodb.io/player?url=..."
+ }
+}
+```
+
+> For latest details, see https://docs.videodb.io/pages/ingest/capture-sdks/realtime-context.md
+
+---
+
+## Event Persistence
+
+Use `ws_listener.py` to dump all WebSocket events to a JSONL file for later analysis.
+
+### Start Listener and Get WebSocket ID
+
+```bash
+# Start with --clear to clear old events (recommended for new sessions)
+python scripts/ws_listener.py --clear &
+
+# Append to existing events (for reconnects)
+python scripts/ws_listener.py &
+```
+
+Or specify a custom output directory:
+
+```bash
+python scripts/ws_listener.py --clear /path/to/output &
+# Or via environment variable:
+VIDEODB_EVENTS_DIR=/path/to/output python scripts/ws_listener.py --clear &
+```
+
+The script outputs `WS_ID=<connection_id>` on the first line, then listens indefinitely.
+
+**Get the ws_id:**
+```bash
+cat /tmp/videodb_ws_id
+```
+
+**Stop the listener:**
+```bash
+kill $(cat /tmp/videodb_ws_pid)
+```
+
+**Functions that accept `ws_connection_id`:**
+
+| Function | Purpose |
+|----------|---------|
+| `conn.create_capture_session()` | Session lifecycle events |
+| RTStream methods | See [rtstream-reference.md](rtstream-reference.md) |
+
+**Output files** (in output directory, default `/tmp`):
+- `videodb_ws_id` - WebSocket connection ID
+- `videodb_events.jsonl` - All events
+- `videodb_ws_pid` - Process ID for easy termination
+
+**Features:**
+- `--clear` flag to clear events file on start (use for new sessions)
+- Auto-reconnect with exponential backoff on connection drops
+- Graceful shutdown on SIGINT/SIGTERM
+- Connection status logging
+
+### JSONL Format
+
+Each line is a JSON object with added timestamps:
+
+```json
+{"ts": "2026-03-02T10:15:30.123Z", "unix_ts": 1709374530.12, "channel": "visual_index", "data": {"text": "..."}}
+{"ts": "2026-03-02T10:15:31.456Z", "unix_ts": 1709374531.45, "event": "capture_session.active", "capture_session_id": "cap-xxx"}
+```
+
+### Reading Events
+
+```python
+import json
+events = [json.loads(l) for l in open("/tmp/videodb_events.jsonl")]
+
+# Filter by channel
+transcripts = [e for e in events if e.get("channel") == "transcript"]
+
+# Filter by time (last 10 minutes)
+import time
+cutoff = time.time() - 600
+recent = [e for e in events if e["unix_ts"] > cutoff]
+
+# Filter visual events containing keyword
+visual = [e for e in events
+ if e.get("channel") == "visual_index"
+ and "code" in e.get("data", {}).get("text", "").lower()]
+```
+
+---
+
+## WebSocket Connection
+
+Connect to receive real-time AI results from transcription and indexing pipelines.
+
+```python
+ws_wrapper = conn.connect_websocket()
+ws = await ws_wrapper.connect()
+ws_id = ws.connection_id
+```
+
+| Property / Method | Type | Description |
+|-------------------|------|-------------|
+| `ws.connection_id` | `str` | Unique connection ID (pass to AI pipeline methods) |
+| `ws.receive()` | `AsyncIterator[dict]` | Async iterator yielding real-time messages |
+
+---
+
+## CaptureSession
+
+### Connection Methods
+
+| Method | Returns | Description |
+|--------|---------|-------------|
+| `conn.create_capture_session(end_user_id, collection_id, ws_connection_id, metadata)` | `CaptureSession` | Create a new capture session |
+| `conn.get_capture_session(capture_session_id)` | `CaptureSession` | Retrieve an existing capture session |
+| `conn.generate_client_token()` | `str` | Generate a client-side authentication token |
+
+### Create a Capture Session
+
+```python
+ws_id = open("/tmp/videodb_ws_id").read().strip()
+
+session = conn.create_capture_session(
+ end_user_id="user-123", # required
+ collection_id="default",
+ ws_connection_id=ws_id,
+ metadata={"app": "my-app"},
+)
+print(f"Session ID: {session.id}")
+```
+
+> **Note:** `end_user_id` is required and identifies the user initiating the capture. For testing or demo purposes, any unique string identifier works (e.g., `"demo-user"`, `"test-123"`).
+
+### CaptureSession Properties
+
+| Property | Type | Description |
+|----------|------|-------------|
+| `session.id` | `str` | Unique capture session ID |
+
+### CaptureSession Methods
+
+| Method | Returns | Description |
+|--------|---------|-------------|
+| `session.get_rtstream(type)` | `list[RTStream]` | Get RTStreams by type: `"mic"`, `"screen"`, or `"system_audio"` |
+
+### Generate a Client Token
+
+```python
+token = conn.generate_client_token()
+```
+
+---
+
+## CaptureClient
+
+The client runs on the user's machine and handles permissions, channel discovery, and streaming.
+
+```python
+from videodb.capture import CaptureClient
+
+client = CaptureClient(client_token=token)
+```
+
+### CaptureClient Methods
+
+| Method | Returns | Description |
+|--------|---------|-------------|
+| `await client.request_permission(type)` | `None` | Request device permission (`"microphone"`, `"screen_capture"`) |
+| `await client.list_channels()` | `Channels` | Discover available audio/video channels |
+| `await client.start_capture_session(capture_session_id, channels, primary_video_channel_id)` | `None` | Start streaming selected channels |
+| `await client.stop_capture()` | `None` | Gracefully stop the capture session |
+| `await client.shutdown()` | `None` | Clean up client resources |
+
+### Request Permissions
+
+```python
+await client.request_permission("microphone")
+await client.request_permission("screen_capture")
+```
+
+### Start a Session
+
+```python
+selected_channels = [c for c in [mic, display, system_audio] if c]
+await client.start_capture_session(
+ capture_session_id=session.id,
+ channels=selected_channels,
+ primary_video_channel_id=display.id if display else None,
+)
+```
+
+### Stop a Session
+
+```python
+await client.stop_capture()
+await client.shutdown()
+```
+
+---
+
+## Channels
+
+Returned by `client.list_channels()`. Groups available devices by type.
+
+```python
+channels = await client.list_channels()
+for ch in channels.all():
+ print(f" {ch.id} ({ch.type}): {ch.name}")
+
+mic = channels.mics.default
+display = channels.displays.default
+system_audio = channels.system_audio.default
+```
+
+### Channel Groups
+
+| Property | Type | Description |
+|----------|------|-------------|
+| `channels.mics` | `ChannelGroup` | Available microphones |
+| `channels.displays` | `ChannelGroup` | Available screen displays |
+| `channels.system_audio` | `ChannelGroup` | Available system audio sources |
+
+### ChannelGroup Methods & Properties
+
+| Member | Type | Description |
+|--------|------|-------------|
+| `group.default` | `Channel` | Default channel in the group (or `None`) |
+| `group.all()` | `list[Channel]` | All channels in the group |
+
+### Channel Properties
+
+| Property | Type | Description |
+|----------|------|-------------|
+| `ch.id` | `str` | Unique channel ID |
+| `ch.type` | `str` | Channel type (`"mic"`, `"display"`, `"system_audio"`) |
+| `ch.name` | `str` | Human-readable channel name |
+| `ch.store` | `bool` | Whether to persist the recording (set to `True` to save) |
+
+Without `store = True`, streams are processed in real-time but not saved.
+
+---
+
+## RTStreams and AI Pipelines
+
+After session is active, retrieve RTStream objects with `session.get_rtstream()`.
+
+For RTStream methods (indexing, transcription, alerts, batch config), see [rtstream-reference.md](rtstream-reference.md).
+
+---
+
+## Session Lifecycle
+
+```
+ create_capture_session()
+ │
+ v
+ ┌───────────────┐
+ │ created │
+ └───────┬───────┘
+ │ client.start_capture_session()
+ v
+ ┌───────────────┐ WebSocket: capture_session.active
+ │ active │ ──> Start AI pipelines
+ └───────┬───────┘
+ │ client.stop_capture()
+ v
+ ┌───────────────┐ WebSocket: capture_session.stopping
+ │ stopping │ ──> Finalize streams
+ └───────┬───────┘
+ │
+ v
+ ┌───────────────┐ WebSocket: capture_session.stopped
+ │ stopped │ ──> All streams finalized
+ └───────┬───────┘
+ │ (if store=True)
+ v
+ ┌───────────────┐ WebSocket: capture_session.exported
+ │ exported │ ──> Access video_id, stream_url, player_url
+ └───────────────┘
+```
diff --git a/skills/videodb-skills/reference/capture.md b/skills/videodb-skills/reference/capture.md
new file mode 100644
index 000000000..8b6be2284
--- /dev/null
+++ b/skills/videodb-skills/reference/capture.md
@@ -0,0 +1,101 @@
+# Capture Guide
+
+## Overview
+
+VideoDB Capture enables real-time screen and audio recording with AI processing. Desktop capture currently supports **macOS** only.
+
+For code-level details (SDK methods, event structures, AI pipelines), see [capture-reference.md](capture-reference.md).
+
+## Quick Start
+
+1. **Start WebSocket listener**: `python scripts/ws_listener.py --clear &`
+2. **Run capture code** (see Complete Capture Workflow below)
+3. **Events written to**: `/tmp/videodb_events.jsonl`
+
+---
+
+## Complete Capture Workflow
+
+No webhooks or polling required. WebSocket delivers all events including session lifecycle.
+
+> **CRITICAL:** The `CaptureClient` must remain running for the entire duration of the capture. It runs the local recorder binary that streams screen/audio data to VideoDB. If the Python process that created the `CaptureClient` exits, the recorder binary is killed and capture stops silently. Always run the capture code as a **long-lived background process** (e.g. `nohup python capture_script.py &`) and use signal handling (`asyncio.Event` + `SIGINT`/`SIGTERM`) to keep it alive until you explicitly stop it.
+
+1. **Start WebSocket listener** in background with `--clear` flag to clear old events. Wait for it to create the WebSocket ID file.
+
+2. **Read the WebSocket ID**. This ID is required for capture session and AI pipelines.
+
+3. **Create a capture session** and generate a client token for the desktop client.
+
+4. **Initialize CaptureClient** with the token. Request permissions for microphone and screen capture.
+
+5. **List and select channels** (mic, display, system_audio). Set `store = True` on channels you want to persist as a video.
+
+6. **Start the session** with selected channels.
+
+7. **Wait for session active** by reading events until you see `capture_session.active`. This event contains the `rtstreams` array. Save session info (session ID, RTStream IDs) to a file (e.g. `/tmp/videodb_capture_info.json`) so other scripts can read it.
+
+8. **Keep the process alive.** Use `asyncio.Event` with signal handlers for `SIGINT`/`SIGTERM` to block until explicitly stopped. Write a PID file (e.g. `/tmp/videodb_capture_pid`) so the process can be stopped later with `kill $(cat /tmp/videodb_capture_pid)`. The PID file should be overwritten on every run so reruns always have the correct PID.
+
+9. **Start AI pipelines** (in a separate command/script) on each RTStream for audio indexing and visual indexing. Read the RTStream IDs from the saved session info file.
+
+10. **Write custom event processing logic** (in a separate command/script) to read real-time events based on your use case. Examples:
+ - Log Slack activity when `visual_index` mentions "Slack"
+ - Summarize discussions when `audio_index` events arrive
+ - Trigger alerts when specific keywords appear in `transcript`
+ - Track application usage from screen descriptions
+
+11. **Stop capture** when done — send SIGTERM to the capture process. It should call `client.stop_capture()` and `client.shutdown()` in its signal handler.
+
+12. **Wait for export** by reading events until you see `capture_session.exported`. This event contains `exported_video_id`, `stream_url`, and `player_url`. This may take several seconds after stopping capture.
+
+13. **Stop WebSocket listener** after receiving the export event. Use `kill $(cat /tmp/videodb_ws_pid)` to cleanly terminate it.
+
+---
+
+## Shutdown Sequence
+
+Proper shutdown order is important to ensure all events are captured:
+
+1. **Stop the capture session** — `client.stop_capture()` then `client.shutdown()`
+2. **Wait for export event** — poll `/tmp/videodb_events.jsonl` for `capture_session.exported`
+3. **Stop the WebSocket listener** — `kill $(cat /tmp/videodb_ws_pid)`
+
+Do NOT kill the WebSocket listener before receiving the export event, or you will miss the final video URLs.
+
+---
+
+## Scripts
+
+| Script | Description |
+|--------|-------------|
+| `scripts/ws_listener.py` | WebSocket event listener (dumps to JSONL) |
+
+### ws_listener.py Usage
+
+```bash
+# Start listener in background (append to existing events)
+python scripts/ws_listener.py &
+
+# Start listener with clear (new session, clears old events)
+python scripts/ws_listener.py --clear &
+
+# Custom output directory
+python scripts/ws_listener.py --clear /path/to/events &
+
+# Stop the listener
+kill $(cat /tmp/videodb_ws_pid)
+```
+
+**Options:**
+- `--clear`: Clear the events file before starting. Use when starting a new capture session.
+
+**Output files:**
+- `videodb_events.jsonl` - All WebSocket events
+- `videodb_ws_id` - WebSocket connection ID (for `ws_connection_id` parameter)
+- `videodb_ws_pid` - Process ID (for stopping the listener)
+
+**Features:**
+- Auto-reconnect with exponential backoff on connection drops
+- Graceful shutdown on SIGINT/SIGTERM
+- PID file for easy process management
+- Connection status logging
diff --git a/skills/videodb-skills/reference/editor.md b/skills/videodb-skills/reference/editor.md
new file mode 100644
index 000000000..4d9d7f5ed
--- /dev/null
+++ b/skills/videodb-skills/reference/editor.md
@@ -0,0 +1,434 @@
+# Timeline Editing Guide
+
+VideoDB provides a non-destructive timeline editor for composing videos from multiple assets, adding text and image overlays, mixing audio tracks, and trimming clips — all server-side without re-encoding or local tools. Use this for trimming, combining clips, overlaying audio/music on video, adding subtitles, and layering text or images.
+
+## Prerequisites
+
+Videos, audio, and images **must be uploaded** to a collection before they can be used as timeline assets. For caption overlays, the video must also be **indexed for spoken words**.
+
+## Core Concepts
+
+### Timeline
+
+A `Timeline` is a virtual composition layer. Assets are placed on it either **inline** (sequentially on the main track) or as **overlays** (layered at a specific timestamp). Nothing modifies the original media; the final stream is compiled on demand.
+
+```python
+from videodb.timeline import Timeline
+
+timeline = Timeline(conn)
+```
+
+### Assets
+
+Every element on a timeline is an **asset**. VideoDB provides five asset types:
+
+| Asset | Import | Primary Use |
+|-------|--------|-------------|
+| `VideoAsset` | `from videodb.asset import VideoAsset` | Video clips (trim, sequencing) |
+| `AudioAsset` | `from videodb.asset import AudioAsset` | Music, SFX, narration |
+| `ImageAsset` | `from videodb.asset import ImageAsset` | Logos, thumbnails, overlays |
+| `TextAsset` | `from videodb.asset import TextAsset, TextStyle` | Titles, captions, lower-thirds |
+| `CaptionAsset` | `from videodb.editor import CaptionAsset` | Auto-rendered subtitles (Editor API) |
+
+## Building a Timeline
+
+### Add Video Clips Inline
+
+Inline assets play one after another on the main video track. The `add_inline` method only accepts `VideoAsset`:
+
+```python
+from videodb.asset import VideoAsset
+
+video_a = coll.get_video(video_id_a)
+video_b = coll.get_video(video_id_b)
+
+timeline = Timeline(conn)
+timeline.add_inline(VideoAsset(asset_id=video_a.id))
+timeline.add_inline(VideoAsset(asset_id=video_b.id))
+
+stream_url = timeline.generate_stream()
+```
+
+### Trim / Sub-clip
+
+Use `start` and `end` on a `VideoAsset` to extract a portion:
+
+```python
+# Take only seconds 10–30 from the source video
+clip = VideoAsset(asset_id=video.id, start=10, end=30)
+timeline.add_inline(clip)
+```
+
+### VideoAsset Parameters
+
+| Parameter | Type | Default | Description |
+|-----------|------|---------|-------------|
+| `asset_id` | `str` | required | Video media ID |
+| `start` | `float` | `0` | Trim start (seconds) |
+| `end` | `float\|None` | `None` | Trim end (`None` = full) |
+
+> **Warning:** The SDK does not validate negative timestamps. Passing `start=-5` is silently accepted but produces broken or unexpected output. Always ensure `start >= 0`, `start < end`, and `end <= video.length` before creating a `VideoAsset`.
+
+## Text Overlays
+
+Add titles, lower-thirds, or captions at any point on the timeline:
+
+```python
+from videodb.asset import TextAsset, TextStyle
+
+title = TextAsset(
+ text="Welcome to the Demo",
+ duration=5,
+ style=TextStyle(
+ fontsize=36,
+ fontcolor="white",
+ boxcolor="black",
+ alpha=0.8,
+ font="Sans",
+ ),
+)
+
+# Overlay the title at the very start (t=0)
+timeline.add_overlay(0, title)
+```
+
+### TextStyle Parameters
+
+| Parameter | Type | Default | Description |
+|-----------|------|---------|-------------|
+| `fontsize` | `int` | `24` | Font size in pixels |
+| `fontcolor` | `str` | `"black"` | CSS colour name or hex |
+| `fontcolor_expr` | `str` | `""` | Dynamic font colour expression |
+| `alpha` | `float` | `1.0` | Text opacity (0.0–1.0) |
+| `font` | `str` | `"Sans"` | Font family |
+| `box` | `bool` | `True` | Enable background box |
+| `boxcolor` | `str` | `"white"` | Background box colour |
+| `boxborderw` | `str` | `"10"` | Box border width |
+| `boxw` | `int` | `0` | Box width override |
+| `boxh` | `int` | `0` | Box height override |
+| `line_spacing` | `int` | `0` | Line spacing |
+| `text_align` | `str` | `"T"` | Text alignment within the box |
+| `y_align` | `str` | `"text"` | Vertical alignment reference |
+| `borderw` | `int` | `0` | Text border width |
+| `bordercolor` | `str` | `"black"` | Text border colour |
+| `expansion` | `str` | `"normal"` | Text expansion mode |
+| `basetime` | `int` | `0` | Base time for time-based expressions |
+| `fix_bounds` | `bool` | `False` | Fix text bounds |
+| `text_shaping` | `bool` | `True` | Enable text shaping |
+| `shadowcolor` | `str` | `"black"` | Shadow colour |
+| `shadowx` | `int` | `0` | Shadow X offset |
+| `shadowy` | `int` | `0` | Shadow Y offset |
+| `tabsize` | `int` | `4` | Tab size in spaces |
+| `x` | `str` | `"(main_w-text_w)/2"` | Horizontal position expression |
+| `y` | `str` | `"(main_h-text_h)/2"` | Vertical position expression |
+
+## Audio Overlays
+
+Layer background music, sound effects, or voiceover on top of the video track:
+
+```python
+from videodb.asset import AudioAsset
+
+music = coll.get_audio(music_id)
+
+audio_layer = AudioAsset(
+ asset_id=music.id,
+ disable_other_tracks=False,
+ fade_in_duration=2,
+ fade_out_duration=2,
+)
+
+# Start the music at t=0, overlaid on the video track
+timeline.add_overlay(0, audio_layer)
+```
+
+### AudioAsset Parameters
+
+| Parameter | Type | Default | Description |
+|-----------|------|---------|-------------|
+| `asset_id` | `str` | required | Audio media ID |
+| `start` | `float` | `0` | Trim start (seconds) |
+| `end` | `float\|None` | `None` | Trim end (`None` = full) |
+| `disable_other_tracks` | `bool` | `True` | When True, mutes other audio tracks |
+| `fade_in_duration` | `float` | `0` | Fade-in seconds (max 5) |
+| `fade_out_duration` | `float` | `0` | Fade-out seconds (max 5) |
+
+## Image Overlays
+
+Add logos, watermarks, or generated images as overlays:
+
+```python
+from videodb.asset import ImageAsset
+
+logo = coll.get_image(logo_id)
+
+logo_overlay = ImageAsset(
+ asset_id=logo.id,
+ duration=10,
+ width=120,
+ height=60,
+ x=20,
+ y=20,
+)
+
+timeline.add_overlay(0, logo_overlay)
+```
+
+### ImageAsset Parameters
+
+| Parameter | Type | Default | Description |
+|-----------|------|---------|-------------|
+| `asset_id` | `str` | required | Image media ID |
+| `width` | `int\|str` | `100` | Display width |
+| `height` | `int\|str` | `100` | Display height |
+| `x` | `int` | `80` | Horizontal position (px from left) |
+| `y` | `int` | `20` | Vertical position (px from top) |
+| `duration` | `float\|None` | `None` | Display duration (seconds) |
+
+## Caption Overlays
+
+There are two ways to add captions to video.
+
+### Method 1: Subtitle Workflow (simplest)
+
+Use `video.add_subtitle()` to burn subtitles directly onto a video stream. This uses the `videodb.timeline.Timeline` internally:
+
+```python
+from videodb import SubtitleStyle
+
+# Video must have spoken words indexed first (force=True skips if already done)
+video.index_spoken_words(force=True)
+
+# Add subtitles with default styling
+stream_url = video.add_subtitle()
+
+# Or customise the subtitle style
+stream_url = video.add_subtitle(style=SubtitleStyle(
+ font_name="Arial",
+ font_size=22,
+ primary_colour="&H00FFFFFF",
+ bold=True,
+))
+```
+
+### Method 2: Editor API (advanced)
+
+The Editor API (`videodb.editor`) provides a track-based composition system with `CaptionAsset`, `Clip`, `Track`, and its own `Timeline`. This is a separate API from the `videodb.timeline.Timeline` used above.
+
+```python
+from videodb.editor import (
+ CaptionAsset,
+ Clip,
+ Track,
+ Timeline as EditorTimeline,
+ FontStyling,
+ BorderAndShadow,
+ Positioning,
+ CaptionAnimation,
+)
+
+# Video must have spoken words indexed first (force=True skips if already done)
+video.index_spoken_words(force=True)
+
+# Create a caption asset
+caption = CaptionAsset(
+ src="auto",
+ font=FontStyling(name="Clear Sans", size=30),
+ primary_color="&H00FFFFFF",
+ back_color="&H00000000",
+ border=BorderAndShadow(outline=1),
+ position=Positioning(margin_v=30),
+ animation=CaptionAnimation.box_highlight,
+)
+
+# Build an editor timeline with tracks and clips
+editor_tl = EditorTimeline(conn)
+track = Track()
+track.add_clip(start=0, clip=Clip(asset=caption, duration=video.length))
+editor_tl.add_track(track)
+stream_url = editor_tl.generate_stream()
+```
+
+### CaptionAsset Parameters
+
+| Parameter | Type | Default | Description |
+|-----------|------|---------|-------------|
+| `src` | `str` | `"auto"` | Caption source (`"auto"` or base64 ASS string) |
+| `font` | `FontStyling\|None` | `FontStyling()` | Font styling (name, size, bold, italic, etc.) |
+| `primary_color` | `str` | `"&H00FFFFFF"` | Primary text colour (ASS format) |
+| `secondary_color` | `str` | `"&H000000FF"` | Secondary text colour (ASS format) |
+| `back_color` | `str` | `"&H00000000"` | Background colour (ASS format) |
+| `border` | `BorderAndShadow\|None` | `BorderAndShadow()` | Border and shadow styling |
+| `position` | `Positioning\|None` | `Positioning()` | Caption alignment and margins |
+| `animation` | `CaptionAnimation\|None` | `None` | Animation effect (e.g., `box_highlight`, `reveal`, `karaoke`) |
+
+## Compiling & Streaming
+
+After assembling a timeline, compile it into a streamable URL. Streams are generated instantly - no render wait times.
+
+```python
+stream_url = timeline.generate_stream()
+print(f"Stream: {stream_url}")
+```
+
+For more streaming options (segment streams, search-to-stream, audio playback), see [streaming.md](streaming.md).
+
+## Complete Workflow Examples
+
+### Highlight Reel with Title Card
+
+```python
+import videodb
+from videodb import SearchType
+from videodb.timeline import Timeline
+from videodb.asset import VideoAsset, TextAsset, TextStyle
+
+conn = videodb.connect()
+coll = conn.get_collection()
+video = coll.get_video("your-video-id")
+
+# 1. Search for key moments
+video.index_spoken_words(force=True)
+results = video.search("product announcement", search_type=SearchType.semantic)
+shots = results.get_shots() # may be empty if no results
+
+# 2. Build timeline
+timeline = Timeline(conn)
+
+# Title card
+title = TextAsset(
+ text="Product Launch Highlights",
+ duration=4,
+ style=TextStyle(fontsize=48, fontcolor="white", boxcolor="#1a1a2e", alpha=0.95),
+)
+timeline.add_overlay(0, title)
+
+# Append each matching clip
+for shot in shots:
+ asset = VideoAsset(asset_id=shot.video_id, start=shot.start, end=shot.end)
+ timeline.add_inline(asset)
+
+# 3. Generate stream
+stream_url = timeline.generate_stream()
+print(f"Highlight reel: {stream_url}")
+```
+
+### Picture-in-Picture with Background Music
+
+```python
+import videodb
+from videodb.timeline import Timeline
+from videodb.asset import VideoAsset, AudioAsset, ImageAsset
+
+conn = videodb.connect()
+coll = conn.get_collection()
+
+main_video = coll.get_video(main_video_id)
+music = coll.get_audio(music_id)
+logo = coll.get_image(logo_id)
+
+timeline = Timeline(conn)
+
+# Main video track
+timeline.add_inline(VideoAsset(asset_id=main_video.id))
+
+# Background music — disable_other_tracks=False to mix with video audio
+timeline.add_overlay(
+ 0,
+ AudioAsset(asset_id=music.id, disable_other_tracks=False, fade_in_duration=3),
+)
+
+# Logo in top-right corner for first 10 seconds
+timeline.add_overlay(
+ 0,
+ ImageAsset(asset_id=logo.id, duration=10, x=1140, y=20, width=120, height=60),
+)
+
+stream_url = timeline.generate_stream()
+print(f"Final video: {stream_url}")
+```
+
+### Multi-Clip Montage from Multiple Videos
+
+```python
+import videodb
+from videodb.timeline import Timeline
+from videodb.asset import VideoAsset, TextAsset, TextStyle
+
+conn = videodb.connect()
+coll = conn.get_collection()
+
+clips = [
+ {"video_id": "vid_001", "start": 5, "end": 15, "label": "Scene 1"},
+ {"video_id": "vid_002", "start": 0, "end": 20, "label": "Scene 2"},
+ {"video_id": "vid_003", "start": 30, "end": 45, "label": "Scene 3"},
+]
+
+timeline = Timeline(conn)
+
+for clip in clips:
+ # Add a label as an overlay on each clip
+ label = TextAsset(
+ text=clip["label"],
+ duration=2,
+ style=TextStyle(fontsize=32, fontcolor="white", boxcolor="#333333"),
+ )
+ timeline.add_inline(
+ VideoAsset(asset_id=clip["video_id"], start=clip["start"], end=clip["end"])
+ )
+ timeline.add_overlay(0, label)
+
+stream_url = timeline.generate_stream()
+print(f"Montage: {stream_url}")
+```
+
+## Two Timeline APIs
+
+VideoDB has two separate timeline systems. They are **not interchangeable**:
+
+| | `videodb.timeline.Timeline` | `videodb.editor.Timeline` (Editor API) |
+|---|---|---|
+| **Import** | `from videodb.timeline import Timeline` | `from videodb.editor import Timeline as EditorTimeline` |
+| **Assets** | `VideoAsset`, `AudioAsset`, `ImageAsset`, `TextAsset` | `CaptionAsset`, `Clip`, `Track` |
+| **Methods** | `add_inline()`, `add_overlay()` | `add_track()` with `Track` / `Clip` |
+| **Best for** | Video composition, overlays, multi-clip editing | Caption/subtitle styling with animations |
+
+Do not mix assets from one API into the other. `CaptionAsset` only works with the Editor API. `VideoAsset` / `AudioAsset` / `ImageAsset` / `TextAsset` only work with `videodb.timeline.Timeline`.
+
+## Limitations & Constraints
+
+The timeline editor is designed for **non-destructive linear composition**. The following operations are **not supported**:
+
+### Not Possible
+
+| Limitation | Detail |
+|---|---|
+| **No transitions or effects** | No crossfades, wipes, dissolves, or transitions between clips. All cuts are hard cuts. |
+| **No video-on-video (picture-in-picture)** | `add_inline()` only accepts `VideoAsset`. You cannot overlay one video stream on top of another. Image overlays can approximate static PiP but not live video. |
+| **No speed or playback control** | No slow-motion, fast-forward, reverse playback, or time remapping. `VideoAsset` has no `speed` parameter. |
+| **No crop, zoom, or pan** | Cannot crop a region of a video frame, apply zoom effects, or pan across a frame. `video.reframe()` is for aspect-ratio conversion only. |
+| **No video filters or color grading** | No brightness, contrast, saturation, hue, or color correction adjustments. |
+| **No animated text** | `TextAsset` is static for its full duration. No fade-in/out, movement, or animation. For animated captions, use `CaptionAsset` with the Editor API. |
+| **No mixed text styling** | A single `TextAsset` has one `TextStyle`. Cannot mix bold, italic, or colors within a single text block. |
+| **No blank or solid-color clips** | Cannot create a solid color frame, black screen, or standalone title card. Text and image overlays require a `VideoAsset` beneath them on the inline track. |
+| **No audio volume control** | `AudioAsset` has no `volume` parameter. Audio is either full volume or muted via `disable_other_tracks`. Cannot mix at a reduced level. |
+| **No keyframe animation** | Cannot change overlay properties over time (e.g., move an image from position A to B). |
+
+### Constraints
+
+| Constraint | Detail |
+|---|---|
+| **Audio fade max 5 seconds** | `fade_in_duration` and `fade_out_duration` are capped at 5 seconds each. |
+| **Overlay positioning is absolute** | Overlays use absolute timestamps from the timeline start. Rearranging inline clips does not move their overlays. |
+| **Inline track is video only** | `add_inline()` only accepts `VideoAsset`. Audio, image, and text must use `add_overlay()`. |
+| **No overlay-to-clip binding** | Overlays are placed at a fixed timeline timestamp. There is no way to attach an overlay to a specific inline clip so it moves with it. |
+
+## Tips
+
+- **Non-destructive**: Timelines never modify source media. You can create multiple timelines from the same assets.
+- **Overlay stacking**: Multiple overlays can start at the same timestamp. Audio overlays mix together; image/text overlays layer in add-order.
+- **Inline is VideoAsset only**: `add_inline()` only accepts `VideoAsset`. Use `add_overlay()` for `AudioAsset`, `ImageAsset`, and `TextAsset`.
+- **Trim precision**: `start`/`end` on `VideoAsset` and `AudioAsset` are in seconds.
+- **Muting video audio**: Set `disable_other_tracks=True` on `AudioAsset` to mute the original video audio when overlaying music or narration.
+- **Fade limits**: `fade_in_duration` and `fade_out_duration` on `AudioAsset` have a maximum of 5 seconds.
+- **Generated media**: Use `coll.generate_music()`, `coll.generate_sound_effect()`, `coll.generate_voice()`, and `coll.generate_image()` to create media that can be used as timeline assets immediately.
diff --git a/skills/videodb-skills/reference/generative.md b/skills/videodb-skills/reference/generative.md
new file mode 100644
index 000000000..e3a6f5935
--- /dev/null
+++ b/skills/videodb-skills/reference/generative.md
@@ -0,0 +1,321 @@
+# Generative Media Guide
+
+VideoDB provides AI-powered generation of images, videos, music, sound effects, voice, and text content. All generation methods are on the **Collection** object.
+
+## Prerequisites
+
+You need a connection and a collection reference before calling any generation method:
+
+```python
+import videodb
+
+conn = videodb.connect()
+coll = conn.get_collection()
+```
+
+## Image Generation
+
+Generate images from text prompts:
+
+```python
+image = coll.generate_image(
+ prompt="a futuristic cityscape at sunset with flying cars",
+ aspect_ratio="16:9",
+)
+
+# Access the generated image
+print(image.id)
+print(image.generate_url()) # returns a signed download URL
+```
+
+### generate_image Parameters
+
+| Parameter | Type | Default | Description |
+|-----------|------|---------|-------------|
+| `prompt` | `str` | required | Text description of the image to generate |
+| `aspect_ratio` | `str` | `"1:1"` | Aspect ratio: `"1:1"`, `"9:16"`, `"16:9"`, `"4:3"`, or `"3:4"` |
+| `callback_url` | `str\|None` | `None` | URL to receive async callback |
+
+Returns an `Image` object with `.id`, `.name`, and `.collection_id`. The `.url` property may be `None` for generated images — always use `image.generate_url()` to get a reliable signed download URL.
+
+> **Note:** Unlike `Video` objects (which use `.generate_stream()`), `Image` objects use `.generate_url()` to retrieve the image URL. The `.url` property is only populated for some image types (e.g. thumbnails).
+
+## Video Generation
+
+Generate short video clips from text prompts:
+
+```python
+video = coll.generate_video(
+ prompt="a timelapse of a flower blooming in a garden",
+ duration=5,
+)
+
+stream_url = video.generate_stream()
+video.play()
+```
+
+### generate_video Parameters
+
+| Parameter | Type | Default | Description |
+|-----------|------|---------|-------------|
+| `prompt` | `str` | required | Text description of the video to generate |
+| `duration` | `float` | `5` | Duration in seconds (must be integer value, 5-8) |
+| `callback_url` | `str\|None` | `None` | URL to receive async callback |
+
+Returns a `Video` object. Generated videos are automatically added to the collection and can be used in timelines, searches, and compilations like any uploaded video.
+
+## Audio Generation
+
+VideoDB provides three separate methods for different audio types.
+
+### Music
+
+Generate background music from text descriptions:
+
+```python
+music = coll.generate_music(
+ prompt="upbeat electronic music with a driving beat, suitable for a tech demo",
+ duration=30,
+)
+
+print(music.id)
+```
+
+| Parameter | Type | Default | Description |
+|-----------|------|---------|-------------|
+| `prompt` | `str` | required | Text description of the music |
+| `duration` | `int` | `5` | Duration in seconds |
+| `callback_url` | `str\|None` | `None` | URL to receive async callback |
+
+### Sound Effects
+
+Generate specific sound effects:
+
+```python
+sfx = coll.generate_sound_effect(
+ prompt="thunderstorm with heavy rain and distant thunder",
+ duration=10,
+)
+```
+
+| Parameter | Type | Default | Description |
+|-----------|------|---------|-------------|
+| `prompt` | `str` | required | Text description of the sound effect |
+| `duration` | `int` | `2` | Duration in seconds |
+| `config` | `dict` | `{}` | Additional configuration |
+| `callback_url` | `str\|None` | `None` | URL to receive async callback |
+
+### Voice (Text-to-Speech)
+
+Generate speech from text:
+
+```python
+voice = coll.generate_voice(
+ text="Welcome to our product demo. Today we'll walk through the key features.",
+ voice_name="Default",
+)
+```
+
+| Parameter | Type | Default | Description |
+|-----------|------|---------|-------------|
+| `text` | `str` | required | Text to convert to speech |
+| `voice_name` | `str` | `"Default"` | Voice to use |
+| `config` | `dict` | `{}` | Additional configuration |
+| `callback_url` | `str\|None` | `None` | URL to receive async callback |
+
+All three audio methods return an `Audio` object with `.id`, `.name`, `.length`, and `.collection_id`.
+
+## Text Generation (LLM Integration)
+
+Use `coll.generate_text()` to run LLM analysis. This is a **Collection-level** method -- pass any context (transcripts, descriptions) directly in the prompt string.
+
+```python
+# Get transcript from a video first
+transcript_text = video.get_transcript_text()
+
+# Generate analysis using collection LLM
+result = coll.generate_text(
+ prompt=f"Summarize the key points discussed in this video:\n{transcript_text}",
+ model_name="pro",
+)
+
+print(result["output"])
+```
+
+### generate_text Parameters
+
+| Parameter | Type | Default | Description |
+|-----------|------|---------|-------------|
+| `prompt` | `str` | required | Prompt with context for the LLM |
+| `model_name` | `str` | `"basic"` | Model tier: `"basic"`, `"pro"`, or `"ultra"` |
+| `response_type` | `str` | `"text"` | Response format: `"text"` or `"json"` |
+
+Returns a `dict` with an `output` key. When `response_type="text"`, `output` is a `str`. When `response_type="json"`, `output` is a `dict`.
+
+```python
+result = coll.generate_text(prompt="Summarize this", model_name="pro")
+print(result["output"]) # access the actual text/dict
+```
+
+### Analyze Scenes with LLM
+
+Combine scene extraction with text generation:
+
+```python
+from videodb import SceneExtractionType
+
+# First index scenes
+video.index_scenes(
+ extraction_type=SceneExtractionType.time_based,
+ extraction_config={"time": 10},
+ prompt="Describe the visual content in this scene.",
+)
+
+# Get transcript for spoken context
+transcript_text = video.get_transcript_text()
+
+# Analyze with collection LLM
+result = coll.generate_text(
+ prompt=(
+ f"Given this video transcript:\n{transcript_text}\n\n"
+ "Based on the spoken and visual content, describe the main topics covered."
+ ),
+ model_name="pro",
+)
+print(result["output"])
+```
+
+## Dubbing and Translation
+
+### Dub a Video
+
+Dub a video into another language using the collection method:
+
+```python
+dubbed_video = coll.dub_video(
+ video_id=video.id,
+ language_code="es", # Spanish
+)
+
+dubbed_video.play()
+```
+
+### dub_video Parameters
+
+| Parameter | Type | Default | Description |
+|-----------|------|---------|-------------|
+| `video_id` | `str` | required | ID of the video to dub |
+| `language_code` | `str` | required | Target language code (e.g., `"es"`, `"fr"`, `"de"`) |
+| `callback_url` | `str\|None` | `None` | URL to receive async callback |
+
+Returns a `Video` object with the dubbed content.
+
+### Translate Transcript
+
+Translate a video's transcript without dubbing:
+
+```python
+translated = video.translate_transcript(
+ language="Spanish",
+ additional_notes="Use formal tone",
+)
+
+for entry in translated:
+ print(entry)
+```
+
+**Supported languages** include: `en`, `es`, `fr`, `de`, `it`, `pt`, `ja`, `ko`, `zh`, `hi`, `ar`, and more.
+
+## Complete Workflow Examples
+
+### Generate Narration for a Video
+
+```python
+import videodb
+
+conn = videodb.connect()
+coll = conn.get_collection()
+video = coll.get_video("your-video-id")
+
+# Get transcript
+transcript_text = video.get_transcript_text()
+
+# Generate narration script using collection LLM
+result = coll.generate_text(
+ prompt=(
+ f"Write a professional narration script for this video content:\n"
+ f"{transcript_text[:2000]}"
+ ),
+ model_name="pro",
+)
+script = result["output"]
+
+# Convert script to speech
+narration = coll.generate_voice(text=script)
+print(f"Narration audio: {narration.id}")
+```
+
+### Generate Thumbnail from Prompt
+
+```python
+thumbnail = coll.generate_image(
+ prompt="professional video thumbnail showing data analytics dashboard, modern design",
+ aspect_ratio="16:9",
+)
+print(f"Thumbnail URL: {thumbnail.generate_url()}")
+```
+
+### Add Generated Music to Video
+
+```python
+import videodb
+from videodb.timeline import Timeline
+from videodb.asset import VideoAsset, AudioAsset
+
+conn = videodb.connect()
+coll = conn.get_collection()
+video = coll.get_video("your-video-id")
+
+# Generate background music
+music = coll.generate_music(
+ prompt="calm ambient background music for a tutorial video",
+ duration=60,
+)
+
+# Build timeline with video + music overlay
+timeline = Timeline(conn)
+timeline.add_inline(VideoAsset(asset_id=video.id))
+timeline.add_overlay(0, AudioAsset(asset_id=music.id, disable_other_tracks=False))
+
+stream_url = timeline.generate_stream()
+print(f"Video with music: {stream_url}")
+```
+
+### Structured JSON Output
+
+```python
+transcript_text = video.get_transcript_text()
+
+result = coll.generate_text(
+ prompt=(
+ f"Given this transcript:\n{transcript_text}\n\n"
+ "Return a JSON object with keys: summary, topics (array), action_items (array)."
+ ),
+ model_name="pro",
+ response_type="json",
+)
+
+# result["output"] is a dict when response_type="json"
+print(result["output"]["summary"])
+print(result["output"]["topics"])
+```
+
+## Tips
+
+- **Generated media is persistent**: All generated content is stored in your collection and can be reused.
+- **Three audio methods**: Use `generate_music()` for background music, `generate_sound_effect()` for SFX, and `generate_voice()` for text-to-speech. There is no unified `generate_audio()` method.
+- **Text generation is collection-level**: `coll.generate_text()` does not have access to video content automatically. Fetch the transcript with `video.get_transcript_text()` and pass it in the prompt.
+- **Model tiers**: `"basic"` is fastest, `"pro"` is balanced, `"ultra"` is highest quality. Use `"pro"` for most analysis tasks.
+- **Combine generation types**: Generate images for overlays, music for backgrounds, and voice for narration, then compose using timelines (see [editor.md](editor.md)).
+- **Prompt quality matters**: Descriptive, specific prompts produce better results across all generation types.
+- **Aspect ratios for images**: Choose from `"1:1"`, `"9:16"`, `"16:9"`, `"4:3"`, or `"3:4"`.
diff --git a/skills/videodb-skills/reference/rtstream-reference.md b/skills/videodb-skills/reference/rtstream-reference.md
new file mode 100644
index 000000000..ad6de9706
--- /dev/null
+++ b/skills/videodb-skills/reference/rtstream-reference.md
@@ -0,0 +1,551 @@
+# RTStream Reference
+
+Code-level details for RTStream operations. For workflow guide, see [rtstream.md](rtstream.md).
+
+Based on [docs.videodb.io](https://docs.videodb.io/pages/ingest/live-streams/realtime-apis.md).
+
+---
+
+## Collection RTStream Methods
+
+Methods on `Collection` for managing RTStreams:
+
+| Method | Returns | Description |
+|--------|---------|-------------|
+| `coll.connect_rtstream(url, name, ...)` | `RTStream` | Create new RTStream from RTSP/RTMP URL |
+| `coll.get_rtstream(id)` | `RTStream` | Get existing RTStream by ID |
+| `coll.list_rtstreams(limit, offset, status, name, ordering)` | `List[RTStream]` | List all RTStreams in collection |
+| `coll.search(query, namespace="rtstream")` | `RTStreamSearchResult` | Search across all RTStreams |
+
+### Connect RTStream
+
+```python
+import videodb
+
+conn = videodb.connect()
+coll = conn.get_collection()
+
+rtstream = coll.connect_rtstream(
+ url="rtmp://your-stream-server/live/stream-key",
+ name="My Live Stream",
+ media_types=["video"], # or ["audio", "video"]
+ sample_rate=30, # optional
+ store=True, # enable recording storage for export
+ enable_transcript=True, # optional
+ ws_connection_id=ws_id, # optional, for real-time events
+)
+```
+
+### Get Existing RTStream
+
+```python
+rtstream = coll.get_rtstream("rts-xxx")
+```
+
+### List RTStreams
+
+```python
+rtstreams = coll.list_rtstreams(
+ limit=10,
+ offset=0,
+ status="connected", # optional filter
+ name="meeting", # optional filter
+ ordering="-created_at",
+)
+
+for rts in rtstreams:
+ print(f"{rts.id}: {rts.name} - {rts.status}")
+```
+
+### From Capture Session
+
+After a capture session is active, retrieve RTStream objects:
+
+```python
+session = conn.get_capture_session(session_id)
+
+mics = session.get_rtstream("mic")
+displays = session.get_rtstream("screen")
+system_audios = session.get_rtstream("system_audio")
+```
+
+Or use the `rtstreams` data from the `capture_session.active` WebSocket event:
+
+```python
+for rts in rtstreams:
+ rtstream = coll.get_rtstream(rts["rtstream_id"])
+```
+
+---
+
+## RTStream Methods
+
+| Method | Returns | Description |
+|--------|---------|-------------|
+| `rtstream.start()` | `None` | Begin ingestion |
+| `rtstream.stop()` | `None` | Stop ingestion |
+| `rtstream.generate_stream(start, end)` | `str` | Stream recorded segment (Unix timestamps) |
+| `rtstream.export(name=None)` | `RTStreamExportResult` | Export to permanent video |
+| `rtstream.index_visuals(prompt, ...)` | `RTStreamSceneIndex` | Create visual index with AI analysis |
+| `rtstream.index_audio(prompt, ...)` | `RTStreamSceneIndex` | Create audio index with LLM summarization |
+| `rtstream.list_scene_indexes()` | `List[RTStreamSceneIndex]` | List all scene indexes on the stream |
+| `rtstream.get_scene_index(index_id)` | `RTStreamSceneIndex` | Get a specific scene index |
+| `rtstream.search(query, ...)` | `RTStreamSearchResult` | Search indexed content |
+| `rtstream.start_transcript(ws_connection_id, engine)` | `dict` | Start live transcription |
+| `rtstream.get_transcript(page, page_size, start, end, since)` | `dict` | Get transcript pages |
+| `rtstream.stop_transcript(engine)` | `dict` | Stop transcription |
+
+---
+
+## Starting and Stopping
+
+```python
+# Begin ingestion
+rtstream.start()
+
+# ... stream is being recorded ...
+
+# Stop ingestion
+rtstream.stop()
+```
+
+---
+
+## Generating Streams
+
+Use Unix timestamps (not seconds offsets) to generate a playback stream from recorded content:
+
+```python
+import time
+
+start_ts = time.time()
+rtstream.start()
+
+# Let it record for a while...
+time.sleep(60)
+
+end_ts = time.time()
+rtstream.stop()
+
+# Generate a stream URL for the recorded segment
+stream_url = rtstream.generate_stream(start=start_ts, end=end_ts)
+print(f"Recorded stream: {stream_url}")
+```
+
+---
+
+## Exporting to Video
+
+Export the recorded stream to a permanent video in the collection:
+
+```python
+export_result = rtstream.export(name="Meeting Recording 2024-01-15")
+
+print(f"Video ID: {export_result.video_id}")
+print(f"Stream URL: {export_result.stream_url}")
+print(f"Player URL: {export_result.player_url}")
+print(f"Duration: {export_result.duration}s")
+```
+
+### RTStreamExportResult Properties
+
+| Property | Type | Description |
+|----------|------|-------------|
+| `video_id` | `str` | ID of the exported video |
+| `stream_url` | `str` | HLS stream URL |
+| `player_url` | `str` | Web player URL |
+| `name` | `str` | Video name |
+| `duration` | `float` | Duration in seconds |
+
+---
+
+## AI Pipelines
+
+AI pipelines process live streams and send results via WebSocket.
+
+### RTStream AI Pipeline Methods
+
+| Method | Returns | Description |
+|--------|---------|-------------|
+| `rtstream.index_audio(prompt, batch_config, ...)` | `RTStreamSceneIndex` | Start audio indexing with LLM summarization |
+| `rtstream.index_visuals(prompt, batch_config, ...)` | `RTStreamSceneIndex` | Start visual indexing of screen content |
+
+### Audio Indexing
+
+Generate LLM summaries of audio content at intervals:
+
+```python
+audio_index = rtstream.index_audio(
+ prompt="Summarize what is being discussed",
+ batch_config={"type": "word", "value": 50},
+ model_name=None, # optional
+ name="meeting_audio", # optional
+ ws_connection_id=ws_id,
+)
+```
+
+**Audio batch_config options:**
+
+| Type | Value | Description |
+|------|-------|-------------|
+| `"word"` | count | Segment every N words |
+| `"sentence"` | count | Segment every N sentences |
+| `"time"` | seconds | Segment every N seconds |
+
+Examples:
+```python
+{"type": "word", "value": 50} # every 50 words
+{"type": "sentence", "value": 5} # every 5 sentences
+{"type": "time", "value": 30} # every 30 seconds
+```
+
+Results arrive on the `audio_index` WebSocket channel.
+
+### Visual Indexing
+
+Generate AI descriptions of visual content:
+
+```python
+scene_index = rtstream.index_visuals(
+ prompt="Describe what is happening on screen",
+ batch_config={"type": "time", "value": 2, "frame_count": 5},
+ model_name="basic",
+ name="screen_monitor", # optional
+ ws_connection_id=ws_id,
+)
+```
+
+**Parameters:**
+
+| Parameter | Type | Description |
+|-----------|------|-------------|
+| `prompt` | `str` | Instructions for the AI model (supports structured JSON output) |
+| `batch_config` | `dict` | Controls frame sampling (see below) |
+| `model_name` | `str` | Model tier: `"mini"`, `"basic"`, `"pro"`, `"ultra"` |
+| `name` | `str` | Name for the index (optional) |
+| `ws_connection_id` | `str` | WebSocket connection ID for receiving results |
+
+**Visual batch_config:**
+
+| Key | Type | Description |
+|-----|------|-------------|
+| `type` | `str` | Only `"time"` is supported for visuals |
+| `value` | `int` | Window size in seconds |
+| `frame_count` | `int` | Number of frames to extract per window |
+
+Example: `{"type": "time", "value": 2, "frame_count": 5}` samples 5 frames every 2 seconds and sends them to the model.
+
+**Structured JSON output:**
+
+Use a prompt that requests JSON format for structured responses:
+
+```python
+scene_index = rtstream.index_visuals(
+ prompt="""Analyze the screen and return a JSON object with:
+{
+ "app_name": "name of the active application",
+ "activity": "what the user is doing",
+ "ui_elements": ["list of visible UI elements"],
+ "contains_text": true/false,
+ "dominant_colors": ["list of main colors"]
+}
+Return only valid JSON.""",
+ batch_config={"type": "time", "value": 3, "frame_count": 3},
+ model_name="pro",
+ ws_connection_id=ws_id,
+)
+```
+
+Results arrive on the `scene_index` WebSocket channel.
+
+---
+
+## Batch Config Summary
+
+| Indexing Type | `type` Options | `value` | Extra Keys |
+|---------------|----------------|---------|------------|
+| **Audio** | `"word"`, `"sentence"`, `"time"` | words/sentences/seconds | - |
+| **Visual** | `"time"` only | seconds | `frame_count` |
+
+Examples:
+```python
+# Audio: every 50 words
+{"type": "word", "value": 50}
+
+# Audio: every 30 seconds
+{"type": "time", "value": 30}
+
+# Visual: 5 frames every 2 seconds
+{"type": "time", "value": 2, "frame_count": 5}
+```
+
+---
+
+## Transcription
+
+Real-time transcription via WebSocket:
+
+```python
+# Start live transcription
+rtstream.start_transcript(
+ ws_connection_id=ws_id,
+ engine=None, # optional, defaults to "assemblyai"
+)
+
+# Get transcript pages (with optional filters)
+transcript = rtstream.get_transcript(
+ page=1,
+ page_size=100,
+ start=None, # optional: start timestamp filter
+ end=None, # optional: end timestamp filter
+ since=None, # optional: for polling, get transcripts after this timestamp
+ engine=None,
+)
+
+# Stop transcription
+rtstream.stop_transcript(engine=None)
+```
+
+Transcript results arrive on the `transcript` WebSocket channel.
+
+---
+
+## RTStreamSceneIndex
+
+When you call `index_audio()` or `index_visuals()`, the method returns an `RTStreamSceneIndex` object. This object represents the running index and provides methods for managing scenes and alerts.
+
+```python
+# index_visuals returns an RTStreamSceneIndex
+scene_index = rtstream.index_visuals(
+ prompt="Describe what is on screen",
+ ws_connection_id=ws_id,
+)
+
+# index_audio also returns an RTStreamSceneIndex
+audio_index = rtstream.index_audio(
+ prompt="Summarize the discussion",
+ ws_connection_id=ws_id,
+)
+```
+
+### RTStreamSceneIndex Properties
+
+| Property | Type | Description |
+|----------|------|-------------|
+| `rtstream_index_id` | `str` | Unique ID of the index |
+| `rtstream_id` | `str` | ID of the parent RTStream |
+| `extraction_type` | `str` | Type of extraction (`time` or `transcript`) |
+| `extraction_config` | `dict` | Extraction configuration |
+| `prompt` | `str` | The prompt used for analysis |
+| `name` | `str` | Name of the index |
+| `status` | `str` | Status (`connected`, `stopped`) |
+
+### RTStreamSceneIndex Methods
+
+| Method | Returns | Description |
+|--------|---------|-------------|
+| `index.get_scenes(start, end, page, page_size)` | `dict` | Get indexed scenes |
+| `index.start()` | `None` | Start/resume the index |
+| `index.stop()` | `None` | Stop the index |
+| `index.create_alert(event_id, callback_url, ws_connection_id)` | `str` | Create alert for event detection |
+| `index.list_alerts()` | `list` | List all alerts on this index |
+| `index.enable_alert(alert_id)` | `None` | Enable an alert |
+| `index.disable_alert(alert_id)` | `None` | Disable an alert |
+
+### Getting Scenes
+
+Poll indexed scenes from the index:
+
+```python
+result = scene_index.get_scenes(
+ start=None, # optional: start timestamp
+ end=None, # optional: end timestamp
+ page=1,
+ page_size=100,
+)
+
+for scene in result["scenes"]:
+ print(f"[{scene['start']}-{scene['end']}] {scene['text']}")
+
+if result["next_page"]:
+ # fetch next page
+ pass
+```
+
+### Managing Scene Indexes
+
+```python
+# List all indexes on the stream
+indexes = rtstream.list_scene_indexes()
+
+# Get a specific index by ID
+scene_index = rtstream.get_scene_index(index_id)
+
+# Stop an index
+scene_index.stop()
+
+# Restart an index
+scene_index.start()
+```
+
+---
+
+## Events
+
+Events are reusable detection rules. Create them once, attach to any index via alerts.
+
+### Connection Event Methods
+
+| Method | Returns | Description |
+|--------|---------|-------------|
+| `conn.create_event(event_prompt, label)` | `str` (event_id) | Create detection event |
+| `conn.list_events()` | `list` | List all events |
+
+### Creating an Event
+
+```python
+event_id = conn.create_event(
+ event_prompt="User opened Slack application",
+ label="slack_opened",
+)
+```
+
+### Listing Events
+
+```python
+events = conn.list_events()
+for event in events:
+ print(f"{event['event_id']}: {event['label']}")
+```
+
+---
+
+## Alerts
+
+Alerts wire events to indexes for real-time notifications. When the AI detects content matching the event description, an alert is sent.
+
+### Creating an Alert
+
+```python
+# Get the RTStreamSceneIndex from index_visuals
+scene_index = rtstream.index_visuals(
+ prompt="Describe what application is open on screen",
+ ws_connection_id=ws_id,
+)
+
+# Create an alert on the index
+alert_id = scene_index.create_alert(
+ event_id=event_id,
+ callback_url="https://your-backend.com/alerts", # for webhook delivery
+ ws_connection_id=ws_id, # for WebSocket delivery (optional)
+)
+```
+
+**Note:** `callback_url` is required. Pass an empty string `""` if only using WebSocket delivery.
+
+### Managing Alerts
+
+```python
+# List all alerts on an index
+alerts = scene_index.list_alerts()
+
+# Enable/disable alerts
+scene_index.disable_alert(alert_id)
+scene_index.enable_alert(alert_id)
+```
+
+### Alert Delivery
+
+| Method | Latency | Use Case |
+|--------|---------|----------|
+| WebSocket | Real-time | Dashboards, live UI |
+| Webhook | < 1 second | Server-to-server, automation |
+
+### WebSocket Alert Event
+
+```json
+{
+ "channel": "alert",
+ "rtstream_id": "rts-xxx",
+ "data": {
+ "event_label": "slack_opened",
+ "timestamp": 1710000012340,
+ "text": "User opened Slack application"
+ }
+}
+```
+
+### Webhook Payload
+
+```json
+{
+ "event_id": "event-xxx",
+ "label": "slack_opened",
+ "confidence": 0.95,
+ "explanation": "User opened the Slack application",
+ "timestamp": "2024-01-15T10:30:45Z",
+ "start_time": 1234.5,
+ "end_time": 1238.0,
+ "stream_url": "https://stream.videodb.io/v3/...",
+ "player_url": "https://console.videodb.io/player?url=..."
+}
+```
+
+---
+
+## WebSocket Integration
+
+All real-time AI results are delivered via WebSocket. Pass `ws_connection_id` to:
+- `rtstream.start_transcript()`
+- `rtstream.index_audio()`
+- `rtstream.index_visuals()`
+- `scene_index.create_alert()`
+
+### WebSocket Channels
+
+| Channel | Source | Content |
+|---------|--------|---------|
+| `transcript` | `start_transcript()` | Real-time speech-to-text |
+| `scene_index` | `index_visuals()` | Visual analysis results |
+| `audio_index` | `index_audio()` | Audio analysis results |
+| `alert` | `create_alert()` | Alert notifications |
+
+For WebSocket event structures and ws_listener usage, see [capture-reference.md](capture-reference.md).
+
+---
+
+## Complete Workflow
+
+```python
+import time
+import videodb
+
+conn = videodb.connect()
+coll = conn.get_collection()
+
+# 1. Connect and start recording
+rtstream = coll.connect_rtstream(
+ url="rtmp://your-stream-server/live/stream-key",
+ name="Weekly Standup",
+)
+rtstream.start()
+
+# 2. Record for the duration of the meeting
+start_ts = time.time()
+time.sleep(1800) # 30 minutes
+end_ts = time.time()
+rtstream.stop()
+
+# 3. Export to a permanent video
+export_result = rtstream.export(name="Weekly Standup Recording")
+print(f"Exported video: {export_result.video_id}")
+
+# 4. Index the exported video for search
+video = coll.get_video(export_result.video_id)
+video.index_spoken_words(force=True)
+
+# 5. Search for action items
+results = video.search("action items and next steps")
+stream_url = results.compile()
+print(f"Action items clip: {stream_url}")
+```
diff --git a/skills/videodb-skills/reference/rtstream.md b/skills/videodb-skills/reference/rtstream.md
new file mode 100644
index 000000000..0d7984335
--- /dev/null
+++ b/skills/videodb-skills/reference/rtstream.md
@@ -0,0 +1,65 @@
+# RTStream Guide
+
+## Overview
+
+RTStream enables real-time ingestion of live video streams (RTSP/RTMP) and desktop capture sessions. Once connected, you can record, index, search, and export content from live sources.
+
+For code-level details (SDK methods, parameters, examples), see [rtstream-reference.md](rtstream-reference.md).
+
+## Use Cases
+
+- **Security & Monitoring**: Connect RTSP cameras, detect events, trigger alerts
+- **Live Broadcasts**: Ingest RTMP streams, index in real-time, enable instant search
+- **Meeting Recording**: Capture desktop screen and audio, transcribe live, export recordings
+- **Event Processing**: Monitor live feeds, run AI analysis, respond to detected content
+
+## Quick Start
+
+1. **Connect to a live stream** (RTSP/RTMP URL) or get RTStream from a capture session
+
+2. **Start ingestion** to begin recording the live content
+
+3. **Start AI pipelines** for real-time indexing (audio, visual, transcription)
+
+4. **Monitor events** via WebSocket for live AI results and alerts
+
+5. **Stop ingestion** when done
+
+6. **Export to video** for permanent storage and further processing
+
+7. **Search the recording** to find specific moments
+
+## RTStream Sources
+
+### From RTSP/RTMP Streams
+
+Connect directly to a live video source:
+
+```python
+rtstream = coll.connect_rtstream(
+ url="rtmp://your-stream-server/live/stream-key",
+ name="My Live Stream",
+)
+```
+
+### From Capture Sessions
+
+Get RTStreams from desktop capture (mic, screen, system audio):
+
+```python
+session = conn.get_capture_session(session_id)
+
+mics = session.get_rtstream("mic")
+displays = session.get_rtstream("screen")
+system_audios = session.get_rtstream("system_audio")
+```
+
+For capture session workflow, see [capture.md](capture.md).
+
+---
+
+## Scripts
+
+| Script | Description |
+|--------|-------------|
+| `scripts/ws_listener.py` | WebSocket event listener for real-time AI results |
diff --git a/skills/videodb-skills/reference/search.md b/skills/videodb-skills/reference/search.md
new file mode 100644
index 000000000..7fff8c872
--- /dev/null
+++ b/skills/videodb-skills/reference/search.md
@@ -0,0 +1,230 @@
+# Search & Indexing Guide
+
+Search allows you to find specific moments inside videos using natural language queries, exact keywords, or visual scene descriptions.
+
+## Prerequisites
+
+Videos **must be indexed** before they can be searched. Indexing is a one-time operation per video per index type.
+
+## Indexing
+
+### Spoken Word Index
+
+Index the transcribed speech content of a video for semantic and keyword search:
+
+```python
+video = coll.get_video(video_id)
+
+# force=True makes indexing idempotent — skips if already indexed
+video.index_spoken_words(force=True)
+```
+
+This transcribes the audio track and builds a searchable index over the spoken content. Required for semantic search and keyword search.
+
+**Parameters:**
+
+| Parameter | Type | Default | Description |
+|-----------|------|---------|-------------|
+| `language_code` | `str\|None` | `None` | Language code of the video |
+| `segmentation_type` | `SegmentationType` | `SegmentationType.sentence` | Segmentation type (`sentence` or `llm`) |
+| `force` | `bool` | `False` | Set to `True` to skip if already indexed (avoids "already exists" error) |
+| `callback_url` | `str\|None` | `None` | Webhook URL for async notification |
+
+### Scene Index
+
+Index visual content by generating AI descriptions of scenes. Like spoken word indexing, this raises an error if a scene index already exists. Extract the existing `scene_index_id` from the error message.
+
+```python
+import re
+from videodb import SceneExtractionType
+
+try:
+ scene_index_id = video.index_scenes(
+ extraction_type=SceneExtractionType.shot_based,
+ prompt="Describe the visual content, objects, actions, and setting in this scene.",
+ )
+except Exception as e:
+ match = re.search(r"id\s+([a-f0-9]+)", str(e))
+ if match:
+ scene_index_id = match.group(1)
+ else:
+ raise
+```
+
+**Extraction types:**
+
+| Type | Description | Best For |
+|------|-------------|----------|
+| `SceneExtractionType.shot_based` | Splits on visual shot boundaries | General purpose, action content |
+| `SceneExtractionType.time_based` | Splits at fixed intervals | Uniform sampling, long static content |
+| `SceneExtractionType.transcript` | Splits based on transcript segments | Speech-driven scene boundaries |
+
+**Parameters for `time_based`:**
+
+```python
+video.index_scenes(
+ extraction_type=SceneExtractionType.time_based,
+ extraction_config={"time": 5, "select_frames": ["first", "last"]},
+ prompt="Describe what is happening in this scene.",
+)
+```
+
+## Search Types
+
+### Semantic Search
+
+Natural language queries matched against spoken content:
+
+```python
+from videodb import SearchType
+
+results = video.search(
+ query="explaining the benefits of machine learning",
+ search_type=SearchType.semantic,
+)
+```
+
+Returns ranked segments where the spoken content semantically matches the query.
+
+### Keyword Search
+
+Exact term matching in transcribed speech:
+
+```python
+results = video.search(
+ query="artificial intelligence",
+ search_type=SearchType.keyword,
+)
+```
+
+Returns segments containing the exact keyword or phrase.
+
+### Scene Search
+
+Visual content queries matched against indexed scene descriptions. Requires a prior `index_scenes()` call.
+
+`index_scenes()` returns a `scene_index_id`. Pass it to `video.search()` to target a specific scene index (especially important when a video has multiple scene indexes):
+
+```python
+from videodb import SearchType, IndexType
+from videodb.exceptions import InvalidRequestError
+
+# Search using semantic search against the scene index.
+# Use score_threshold to filter low-relevance noise (recommended: 0.3+).
+try:
+ results = video.search(
+ query="person writing on a whiteboard",
+ search_type=SearchType.semantic,
+ index_type=IndexType.scene,
+ scene_index_id=scene_index_id,
+ score_threshold=0.3,
+ )
+ shots = results.get_shots()
+except InvalidRequestError as e:
+ if "No results found" in str(e):
+ shots = []
+ else:
+ raise
+```
+
+**Important notes:**
+
+- Use `SearchType.semantic` with `index_type=IndexType.scene` — this is the most reliable combination and works on all plans.
+- `SearchType.scene` exists but may not be available on all plans (e.g. Free tier). Prefer `SearchType.semantic` with `IndexType.scene`.
+- The `scene_index_id` parameter is optional. If omitted, the search runs against all scene indexes on the video. Pass it to target a specific index.
+- You can create multiple scene indexes per video (with different prompts or extraction types) and search them independently using `scene_index_id`.
+
+### Scene Search with Metadata Filtering
+
+When indexing scenes with custom metadata, you can combine semantic search with metadata filters:
+
+```python
+from videodb import SearchType, IndexType
+
+results = video.search(
+ query="a skillful chasing scene",
+ search_type=SearchType.semantic,
+ index_type=IndexType.scene,
+ scene_index_id=scene_index_id,
+ filter=[{"camera_view": "road_ahead"}, {"action_type": "chasing"}],
+)
+```
+
+See the [scene_level_metadata_indexing cookbook](https://github.com/video-db/videodb-cookbook/blob/main/quickstart/scene_level_metadata_indexing.ipynb) for a full example of custom metadata indexing and filtered search.
+
+## Working with Results
+
+### Get Shots
+
+Access individual result segments:
+
+```python
+results = video.search("your query")
+
+for shot in results.get_shots():
+ print(f"Video: {shot.video_id}")
+ print(f"Start: {shot.start:.2f}s")
+ print(f"End: {shot.end:.2f}s")
+ print(f"Text: {shot.text}")
+ print("---")
+```
+
+### Play Compiled Results
+
+Stream all matching segments as a single compiled video:
+
+```python
+results = video.search("your query")
+stream_url = results.compile()
+results.play() # opens compiled stream in browser
+```
+
+### Extract Clips
+
+Download or stream specific result segments:
+
+```python
+for shot in results.get_shots():
+ stream_url = shot.generate_stream()
+ print(f"Clip: {stream_url}")
+```
+
+## Cross-Collection Search
+
+Search across all videos in a collection:
+
+```python
+coll = conn.get_collection()
+
+# Search across all videos in the collection
+results = coll.search(
+ query="product demo",
+ search_type=SearchType.semantic,
+)
+
+for shot in results.get_shots():
+ print(f"Video: {shot.video_id} [{shot.start:.1f}s - {shot.end:.1f}s]")
+```
+
+> **Note:** Collection-level search only supports `SearchType.semantic`. Using `SearchType.keyword` or `SearchType.scene` with `coll.search()` will raise `NotImplementedError`. For keyword or scene search, use `video.search()` on individual videos instead.
+
+## Search + Compile
+
+Index, search, and compile matching segments into a single playable stream:
+
+```python
+video.index_spoken_words(force=True)
+results = video.search(query="your query", search_type=SearchType.semantic)
+stream_url = results.compile()
+print(stream_url)
+```
+
+## Tips
+
+- **Index once, search many times**: Indexing is the expensive operation. Once indexed, searches are fast.
+- **Combine index types**: Index both spoken words and scenes to enable all search types on the same video.
+- **Refine queries**: Semantic search works best with descriptive, natural language phrases rather than single keywords.
+- **Use keyword search for precision**: When you need exact term matches, keyword search avoids semantic drift.
+- **Handle "No results found"**: `video.search()` raises `InvalidRequestError` when no results match. Always wrap search calls in try/except and treat `"No results found"` as an empty result set.
+- **Filter scene search noise**: Semantic scene search can return low-relevance results for vague queries. Use `score_threshold=0.3` (or higher) to filter noise.
+- **Idempotent indexing**: Use `index_spoken_words(force=True)` to safely re-index. `index_scenes()` has no `force` parameter — wrap it in try/except and extract the existing `scene_index_id` from the error message with `re.search(r"id\s+([a-f0-9]+)", str(e))`.
diff --git a/skills/videodb-skills/reference/streaming.md b/skills/videodb-skills/reference/streaming.md
new file mode 100644
index 000000000..59b39dadc
--- /dev/null
+++ b/skills/videodb-skills/reference/streaming.md
@@ -0,0 +1,339 @@
+# Streaming & Playback
+
+VideoDB generates streams on-demand, returning HLS-compatible URLs that play instantly in any standard video player. No render times or export waits - edits, searches, and compositions stream immediately.
+
+## Prerequisites
+
+Videos **must be uploaded** to a collection before streams can be generated. For search-based streams, the video must also be **indexed** (spoken words and/or scenes). See [search.md](search.md) for indexing details.
+
+## Core Concepts
+
+### Stream Generation
+
+Every video, search result, and timeline in VideoDB can produce a **stream URL**. This URL points to an HLS (HTTP Live Streaming) manifest that is compiled on demand.
+
+```python
+# From a video
+stream_url = video.generate_stream()
+
+# From a timeline
+stream_url = timeline.generate_stream()
+
+# From search results
+stream_url = results.compile()
+```
+
+## Streaming a Single Video
+
+### Basic Playback
+
+```python
+import videodb
+
+conn = videodb.connect()
+coll = conn.get_collection()
+video = coll.get_video("your-video-id")
+
+# Generate stream URL
+stream_url = video.generate_stream()
+print(f"Stream: {stream_url}")
+
+# Open in default browser
+video.play()
+```
+
+### With Subtitles
+
+```python
+# Index and add subtitles first
+video.index_spoken_words(force=True)
+video.add_subtitle()
+
+# Stream now includes subtitles
+stream_url = video.generate_stream()
+```
+
+### Specific Segments
+
+Stream only a portion of a video by passing a timeline of timestamp ranges:
+
+```python
+# Stream seconds 10-30 and 60-90
+stream_url = video.generate_stream(timeline=[(10, 30), (60, 90)])
+print(f"Segment stream: {stream_url}")
+```
+
+## Streaming Timeline Compositions
+
+Build a multi-asset composition and stream it in real time:
+
+```python
+import videodb
+from videodb.timeline import Timeline
+from videodb.asset import VideoAsset, AudioAsset, ImageAsset, TextAsset, TextStyle
+
+conn = videodb.connect()
+coll = conn.get_collection()
+
+video = coll.get_video(video_id)
+music = coll.get_audio(music_id)
+
+timeline = Timeline(conn)
+
+# Main video content
+timeline.add_inline(VideoAsset(asset_id=video.id))
+
+# Background music overlay (starts at second 0)
+timeline.add_overlay(0, AudioAsset(asset_id=music.id))
+
+# Text overlay at the beginning
+timeline.add_overlay(0, TextAsset(
+ text="Live Demo",
+ duration=3,
+ style=TextStyle(fontsize=48, fontcolor="white", boxcolor="#000000"),
+))
+
+# Generate the composed stream
+stream_url = timeline.generate_stream()
+print(f"Composed stream: {stream_url}")
+```
+
+**Important:** `add_inline()` only accepts `VideoAsset`. Use `add_overlay()` for `AudioAsset`, `ImageAsset`, and `TextAsset`.
+
+For detailed timeline editing, see [editor.md](editor.md).
+
+## Streaming Search Results
+
+Compile search results into a single stream of all matching segments:
+
+```python
+from videodb import SearchType
+
+video.index_spoken_words(force=True)
+results = video.search("key announcement", search_type=SearchType.semantic)
+
+# Compile all matching shots into one stream
+stream_url = results.compile()
+print(f"Search results stream: {stream_url}")
+
+# Or play directly
+results.play()
+```
+
+### Stream Individual Search Hits
+
+```python
+results = video.search("product demo", search_type=SearchType.semantic)
+
+for i, shot in enumerate(results.get_shots()):
+ stream_url = shot.generate_stream()
+ print(f"Hit {i+1} [{shot.start:.1f}s-{shot.end:.1f}s]: {stream_url}")
+```
+
+## Audio Playback
+
+Get a signed playback URL for audio content:
+
+```python
+audio = coll.get_audio(audio_id)
+playback_url = audio.generate_url()
+print(f"Audio URL: {playback_url}")
+```
+
+## Complete Workflow Examples
+
+### Search-to-Stream Pipeline
+
+Combine search, timeline composition, and streaming in one workflow:
+
+```python
+import videodb
+from videodb import SearchType
+from videodb.timeline import Timeline
+from videodb.asset import VideoAsset, TextAsset, TextStyle
+
+conn = videodb.connect()
+coll = conn.get_collection()
+video = coll.get_video("your-video-id")
+
+video.index_spoken_words(force=True)
+
+# Search for key moments
+queries = ["introduction", "main demo", "Q&A"]
+timeline = Timeline(conn)
+
+for query in queries:
+ # Find matching segments
+ results = video.search(query, search_type=SearchType.semantic)
+ for shot in results.get_shots():
+ timeline.add_inline(
+ VideoAsset(asset_id=shot.video_id, start=shot.start, end=shot.end)
+ )
+
+ # Add section label as overlay on the first shot
+ timeline.add_overlay(0, TextAsset(
+ text=query.title(),
+ duration=2,
+ style=TextStyle(fontsize=36, fontcolor="white", boxcolor="#222222"),
+ ))
+
+stream_url = timeline.generate_stream()
+print(f"Dynamic compilation: {stream_url}")
+```
+
+### Multi-Video Stream
+
+Combine clips from different videos into a single stream:
+
+```python
+import videodb
+from videodb.timeline import Timeline
+from videodb.asset import VideoAsset
+
+conn = videodb.connect()
+coll = conn.get_collection()
+
+video_clips = [
+ {"id": "vid_001", "start": 0, "end": 15},
+ {"id": "vid_002", "start": 10, "end": 30},
+ {"id": "vid_003", "start": 5, "end": 25},
+]
+
+timeline = Timeline(conn)
+for clip in video_clips:
+ timeline.add_inline(
+ VideoAsset(asset_id=clip["id"], start=clip["start"], end=clip["end"])
+ )
+
+stream_url = timeline.generate_stream()
+print(f"Multi-video stream: {stream_url}")
+```
+
+### Conditional Stream Assembly
+
+Build a stream dynamically based on search availability:
+
+```python
+import videodb
+from videodb import SearchType
+from videodb.timeline import Timeline
+from videodb.asset import VideoAsset, TextAsset, TextStyle
+
+conn = videodb.connect()
+coll = conn.get_collection()
+video = coll.get_video("your-video-id")
+
+video.index_spoken_words(force=True)
+
+timeline = Timeline(conn)
+
+# Try to find specific content; fall back to full video
+topics = ["opening remarks", "technical deep dive", "closing"]
+
+found_any = False
+for topic in topics:
+ results = video.search(topic, search_type=SearchType.semantic)
+ shots = results.get_shots()
+ if shots:
+ found_any = True
+ for shot in shots:
+ timeline.add_inline(
+ VideoAsset(asset_id=shot.video_id, start=shot.start, end=shot.end)
+ )
+ # Add a label overlay for the section
+ timeline.add_overlay(0, TextAsset(
+ text=topic.title(),
+ duration=2,
+ style=TextStyle(fontsize=32, fontcolor="white", boxcolor="#1a1a2e"),
+ ))
+
+if found_any:
+ stream_url = timeline.generate_stream()
+ print(f"Curated stream: {stream_url}")
+else:
+ # Fall back to full video stream
+ stream_url = video.generate_stream()
+ print(f"Full video stream: {stream_url}")
+```
+
+### Live Event Recap
+
+Process an event recording into a streamable recap with multiple sections:
+
+```python
+import videodb
+from videodb import SearchType
+from videodb.timeline import Timeline
+from videodb.asset import VideoAsset, AudioAsset, ImageAsset, TextAsset, TextStyle
+
+conn = videodb.connect()
+coll = conn.get_collection()
+
+# Upload event recording
+event = coll.upload(url="https://example.com/event-recording.mp4")
+event.index_spoken_words(force=True)
+
+# Generate background music
+music = coll.generate_music(
+ prompt="upbeat corporate background music",
+ duration=120,
+)
+
+# Generate title image
+title_img = coll.generate_image(
+ prompt="modern event recap title card, dark background, professional",
+ aspect_ratio="16:9",
+)
+
+# Build the recap timeline
+timeline = Timeline(conn)
+
+# Main video segments from search
+keynote = event.search("keynote announcement", search_type=SearchType.semantic)
+if keynote.get_shots():
+ for shot in keynote.get_shots()[:5]:
+ timeline.add_inline(
+ VideoAsset(asset_id=shot.video_id, start=shot.start, end=shot.end)
+ )
+
+demo = event.search("product demo", search_type=SearchType.semantic)
+if demo.get_shots():
+ for shot in demo.get_shots()[:5]:
+ timeline.add_inline(
+ VideoAsset(asset_id=shot.video_id, start=shot.start, end=shot.end)
+ )
+
+# Overlay title card image
+timeline.add_overlay(0, ImageAsset(
+ asset_id=title_img.id, width=100, height=100, x=80, y=20, duration=5
+))
+
+# Overlay section labels
+timeline.add_overlay(5, TextAsset(
+ text="Keynote Highlights",
+ duration=3,
+ style=TextStyle(fontsize=40, fontcolor="white", boxcolor="#0d1117"),
+))
+
+# Overlay background music
+timeline.add_overlay(0, AudioAsset(
+ asset_id=music.id, fade_in_duration=3
+))
+
+# Stream the final recap
+stream_url = timeline.generate_stream()
+print(f"Event recap: {stream_url}")
+```
+
+---
+
+## Tips
+
+- **HLS compatibility**: Stream URLs return HLS manifests (`.m3u8`). They work in Safari natively, and in other browsers via hls.js or similar libraries.
+- **On-demand compilation**: Streams are compiled server-side when requested. The first play may have a brief compilation delay; subsequent plays of the same composition are cached.
+- **Caching**: Calling `video.generate_stream()` a second time without arguments returns the cached stream URL rather than recompiling.
+- **Segment streams**: `video.generate_stream(timeline=[(start, end)])` is the fastest way to stream a specific clip without building a full `Timeline` object.
+- **Inline vs overlay**: `add_inline()` only accepts `VideoAsset` and places assets sequentially on the main track. `add_overlay()` accepts `AudioAsset`, `ImageAsset`, and `TextAsset` and layers them on top at a given start time.
+- **TextStyle defaults**: `TextStyle` defaults to `font='Sans'`, `fontcolor='black'`. Use `boxcolor` (not `bgcolor`) for background color on text.
+- **Combine with generation**: Use `coll.generate_music(prompt, duration)` and `coll.generate_image(prompt, aspect_ratio)` to create assets for timeline compositions.
+- **Playback**: `.play()` opens the stream URL in the default system browser. For programmatic use, work with the URL string directly.
diff --git a/skills/videodb-skills/reference/use-cases.md b/skills/videodb-skills/reference/use-cases.md
new file mode 100644
index 000000000..d86f8470f
--- /dev/null
+++ b/skills/videodb-skills/reference/use-cases.md
@@ -0,0 +1,118 @@
+# Use Cases
+
+Common workflows and what VideoDB enables. For code details, see [api-reference.md](api-reference.md), [capture.md](capture.md), [editor.md](editor.md), and [search.md](search.md).
+
+---
+
+## Video Search & Highlights
+
+### Create Highlight Reels
+Upload a long video (conference talk, lecture, meeting recording), search for key moments by topic ("product announcement", "Q&A session", "demo"), and automatically compile matching segments into a shareable highlight reel.
+
+### Build Searchable Video Libraries
+Batch upload videos to a collection, index them for spoken word search, then query across the entire library. Find specific topics across hundreds of hours of content instantly.
+
+### Extract Specific Clips
+Search for moments matching a query ("budget discussion", "action items") and extract each matching segment as an individual clip with its own stream URL.
+
+---
+
+## Video Enhancement
+
+### Add Professional Polish
+Take raw footage and enhance it with:
+- Auto-generated subtitles from speech
+- Custom thumbnails at specific timestamps
+- Background music overlays
+- Intro/outro sequences with generated images
+
+### AI-Enhanced Content
+Combine existing video with generative AI:
+- Generate text summaries from transcript
+- Create background music matching video duration
+- Generate title cards and overlay images
+- Mix all elements into a polished final output
+
+---
+
+## Real-Time Capture (Desktop/Meeting)
+
+### Screen + Audio Recording with AI
+Capture screen, microphone, and system audio simultaneously. Get real-time:
+- **Live transcription** - Speech to text as it happens
+- **Audio summaries** - Periodic AI-generated summaries of discussions
+- **Visual indexing** - AI descriptions of screen activity
+
+### Meeting Capture with Summarization
+Record meetings with live transcription of all participants. Get periodic summaries with key discussion points, decisions, and action items delivered in real-time.
+
+### Screen Activity Tracking
+Track what's happening on screen with AI-generated descriptions:
+- "User is browsing a spreadsheet in Google Sheets"
+- "User switched to a code editor with a Python file"
+- "Video call with screen sharing enabled"
+
+### Post-Session Processing
+After capture ends, the recording is exported as a permanent video. Then:
+- Generate searchable transcript
+- Search for specific topics within the recording
+- Extract clips of important moments
+- Share via stream URL or player link
+
+---
+
+## Live Stream Intelligence (RTSP/RTMP)
+
+### Connect External Streams
+Ingest live video from RTSP/RTMP sources (security cameras, encoders, broadcasts). Process and index content in real-time.
+
+### Real-Time Event Detection
+Define events to detect in live streams:
+- "Person entering restricted area"
+- "Traffic violation at intersection"
+- "Product visible on shelf"
+
+Get alerts via WebSocket or webhook when events occur.
+
+### Live Stream Search
+Search across recorded live stream content. Find specific moments and generate clips from hours of continuous footage.
+
+---
+
+## Content Moderation & Safety
+
+### Automated Content Review
+Index video scenes with AI and search for problematic content. Flag videos containing violence, inappropriate content, or policy violations.
+
+### Profanity Detection
+Detect and locate profanity in audio. Optionally overlay beep sounds at detected timestamps.
+
+---
+
+## Platform Integration
+
+### Social Media Formatting
+Reframe videos for different platforms:
+- Vertical (9:16) for TikTok, Reels, Shorts
+- Square (1:1) for Instagram feed
+- Landscape (16:9) for YouTube
+
+### Transcode for Delivery
+Change resolution, bitrate, or quality for different delivery targets. Output optimized streams for web, mobile, or broadcast.
+
+### Generate Shareable Links
+Every operation produces playable stream URLs. Embed in web players, share directly, or integrate with existing platforms.
+
+---
+
+## Workflow Summary
+
+| Goal | VideoDB Approach |
+|------|------------------|
+| Find moments in video | Index spoken words/scenes → Search → Compile clips |
+| Create highlights | Search multiple topics → Build timeline → Generate stream |
+| Add subtitles | Index spoken words → Add subtitle overlay |
+| Record screen + AI | Start capture → Run AI pipelines → Export video |
+| Monitor live streams | Connect RTSP → Index scenes → Create alerts |
+| Reformat for social | Reframe to target aspect ratio |
+| Combine clips | Build timeline with multiple assets → Generate stream |
diff --git a/skills/videodb-skills/scripts/ws_listener.py b/skills/videodb-skills/scripts/ws_listener.py
new file mode 100644
index 000000000..e62fb72b1
--- /dev/null
+++ b/skills/videodb-skills/scripts/ws_listener.py
@@ -0,0 +1,204 @@
+#!/usr/bin/env python3
+"""
+WebSocket event listener for VideoDB with auto-reconnect and graceful shutdown.
+
+Usage:
+ python scripts/ws_listener.py [OPTIONS] [output_dir]
+
+Arguments:
+ output_dir Directory for output files (default: /tmp or VIDEODB_EVENTS_DIR env var)
+
+Options:
+ --clear Clear the events file before starting (use when starting a new session)
+
+Output files:
+ <output_dir>/videodb_events.jsonl - All WebSocket events (JSONL format)
+ <output_dir>/videodb_ws_id - WebSocket connection ID
+ <output_dir>/videodb_ws_pid - Process ID for easy termination
+
+Output (first line, for parsing):
+ WS_ID=<connection_id>
+
+Examples:
+ python scripts/ws_listener.py & # Run in background
+ python scripts/ws_listener.py --clear # Clear events and start fresh
+ python scripts/ws_listener.py --clear /tmp/mydir # Custom dir with clear
+ kill $(cat /tmp/videodb_ws_pid) # Stop the listener
+"""
+import os
+import sys
+import json
+import signal
+import asyncio
+from datetime import datetime, timezone
+from pathlib import Path
+
+from dotenv import load_dotenv
+load_dotenv()
+
+import videodb
+
+# Retry config
+MAX_RETRIES = 10
+INITIAL_BACKOFF = 1 # seconds
+MAX_BACKOFF = 60 # seconds
+
+# Parse arguments
+def parse_args():
+ clear = False
+ output_dir = None
+
+ args = sys.argv[1:]
+ for arg in args:
+ if arg == "--clear":
+ clear = True
+ elif not arg.startswith("-"):
+ output_dir = arg
+
+ if output_dir is None:
+ output_dir = os.environ.get("VIDEODB_EVENTS_DIR", "/tmp")
+
+ return clear, Path(output_dir)
+
+CLEAR_EVENTS, OUTPUT_DIR = parse_args()
+EVENTS_FILE = OUTPUT_DIR / "videodb_events.jsonl"
+WS_ID_FILE = OUTPUT_DIR / "videodb_ws_id"
+PID_FILE = OUTPUT_DIR / "videodb_ws_pid"
+
+# Track if this is the first connection (for clearing events)
+_first_connection = True
+
+
+def log(msg: str):
+ """Log with timestamp."""
+ ts = datetime.now().strftime("%H:%M:%S")
+ print(f"[{ts}] {msg}", flush=True)
+
+
+def append_event(event: dict):
+ """Append event to JSONL file with timestamps."""
+ event["ts"] = datetime.now(timezone.utc).isoformat()
+ event["unix_ts"] = datetime.now(timezone.utc).timestamp()
+ with open(EVENTS_FILE, "a") as f:
+ f.write(json.dumps(event) + "\n")
+
+
+def write_pid():
+ """Write PID file for easy process management."""
+ OUTPUT_DIR.mkdir(parents=True, exist_ok=True)
+ PID_FILE.write_text(str(os.getpid()))
+
+
+def cleanup_pid():
+ """Remove PID file on exit."""
+ try:
+ PID_FILE.unlink(missing_ok=True)
+ except Exception:
+ pass
+
+
+async def listen_with_retry():
+ """Main listen loop with auto-reconnect and exponential backoff."""
+ global _first_connection
+
+ retry_count = 0
+ backoff = INITIAL_BACKOFF
+
+ while retry_count < MAX_RETRIES:
+ try:
+ conn = videodb.connect()
+ ws_wrapper = conn.connect_websocket()
+ ws = await ws_wrapper.connect()
+ ws_id = ws.connection_id
+
+ # Ensure output directory exists
+ OUTPUT_DIR.mkdir(parents=True, exist_ok=True)
+
+ # Clear events file only on first connection if --clear flag is set
+ if _first_connection and CLEAR_EVENTS:
+ EVENTS_FILE.unlink(missing_ok=True)
+ log("Cleared events file")
+ _first_connection = False
+
+ # Write ws_id to file for easy retrieval
+ WS_ID_FILE.write_text(ws_id)
+
+ # Print ws_id (parseable format for LLM)
+ if retry_count == 0:
+ print(f"WS_ID={ws_id}", flush=True)
+ log(f"Connected (ws_id={ws_id})")
+
+ # Reset retry state on successful connection
+ retry_count = 0
+ backoff = INITIAL_BACKOFF
+
+ # Listen for messages
+ async for msg in ws.receive():
+ append_event(msg)
+ channel = msg.get("channel", msg.get("event", "unknown"))
+ text = msg.get("data", {}).get("text", "")
+ if text:
+ print(f"[{channel}] {text[:80]}", flush=True)
+
+ # If we exit the loop normally, connection was closed
+ log("Connection closed by server")
+
+ except asyncio.CancelledError:
+ log("Shutdown requested")
+ raise
+ except Exception as e:
+ retry_count += 1
+ log(f"Connection error: {e}")
+
+ if retry_count >= MAX_RETRIES:
+ log(f"Max retries ({MAX_RETRIES}) exceeded, exiting")
+ break
+
+ log(f"Reconnecting in {backoff}s (attempt {retry_count}/{MAX_RETRIES})...")
+ await asyncio.sleep(backoff)
+ backoff = min(backoff * 2, MAX_BACKOFF)
+
+
+async def main_async():
+ """Async main with signal handling."""
+ loop = asyncio.get_running_loop()
+ shutdown_event = asyncio.Event()
+
+ def handle_signal():
+ log("Received shutdown signal")
+ shutdown_event.set()
+
+ # Register signal handlers
+ for sig in (signal.SIGINT, signal.SIGTERM):
+ loop.add_signal_handler(sig, handle_signal)
+
+ # Run listener with cancellation support
+ listen_task = asyncio.create_task(listen_with_retry())
+ shutdown_task = asyncio.create_task(shutdown_event.wait())
+
+ done, pending = await asyncio.wait(
+ [listen_task, shutdown_task],
+ return_when=asyncio.FIRST_COMPLETED,
+ )
+
+ # Cancel remaining tasks
+ for task in pending:
+ task.cancel()
+ try:
+ await task
+ except asyncio.CancelledError:
+ pass
+
+ log("Shutdown complete")
+
+
+def main():
+ write_pid()
+ try:
+ asyncio.run(main_async())
+ finally:
+ cleanup_pid()
+
+
+if __name__ == "__main__":
+ main()