Claude Code skill that makes launch videos, ads and music videos from one line: idea, script, visuals, voice, music, edit and review, then a review room where your notes make the next take
Toolbox: use the best tool for each part, and offer it
The goal is the most mind-blowing result, not "everything drawn by Opus". Opus is the director and the renderer: it writes the plan, the timing, the code, the motion, the type, the compositing and the self-review. It is weak at photoreal images, faces, painterly likeness detail, music and voice. For those, bring in specialist tools.
Illustrated styles are code-only by nature. Real-looking films are not.
Pitch the route that makes the best film first, with its price, then the best free one (SKILL.md step 3). The table below says where code alone is the best look, not where it is the cheap fallback.
Opus 5.5 is an excellent illustrator: watercolor characters, flat and stylized-3D cartoons, typography, diagrams, UI, particles, light, motion. Most of the strongest references used no images at all (a P(doom) music video, a kinetic lyric video, a title-sequence reel, a product explainer, a wordless short), and "no images, every frame drawn in code" is part of the wow. What it can't do well is photorealism: real photos, real faces, realistic footage.
| Style | Default | When to add a model |
|---|---|---|
| Watercolor character | code-only | only if the user wants a specific real person's likeness beyond a cartoon |
| Clean kinetic explainer | code-only (the mascot can be drawn in code and dithered) | a photo-based mascot or real product photos |
| Dark premium explainer | code-only | rarely |
| Broadcast title sequence | code-only | real logos or photos of people in the lineup |
| Kinetic lyric video | code-only | never needed |
| Cinematic animated short | code-only | rarely |
| Image-to-game cinematic | model images (that's the premise) | always |
| Editorial HUD film | model footage (photoreal is the premise) | always |
Music is different: code can't make it, so Suno is worth offering for most films (or the user brings a track). Sound effects ARE code now: scripts/sfx.mjs synthesizes hits, whooshes, UI ticks, risers and impacts deterministically and places them frame-exactly (audio.md); ElevenLabs stays for voice, announcers and organic one-offs.
Understand what the user wants before proposing tools: "a cartoon of my family" is code; "my family, looking like us, in a Pixar-style film" may want model-made character sheets. When unsure, pitch both with their ceilings and let the person choose.
Offer, never assume. Not every user has these subscriptions, and credits cost money. At the pitch step, show an asset plan: for each asset the film needs, the options (code-only, free route, paid route with a cost estimate), and your recommendation. The user picks. Never spend credits without an explicit yes on the number.
What each tool is for
| Need | Best tool | Cost notes |
|---|---|---|
| Photoreal or painterly images, character sheets, pose sheets, sets, moodboards, a real person's likeness | Nano Banana (Pro / 2.5) in the Higgsfield web UI | Free and unlimited on a Higgsfield plan only in the web UI with the Unlimited toggle ON. The Higgsfield MCP/API bills ~1 credit per image. If the user is logged in, you can drive the web UI through Chrome (see below). |
| Same, alternative | GPT image through Codex, or ChatGPT | Covered by the user's ChatGPT plan, no charge per picture. If the user has Codex (its command-line tool, or the ChatGPT app on a Mac, which ships it), make the pictures yourself with codex exec (see below). Otherwise the user pastes your prompt into ChatGPT and you read the downloaded result. Strong at characters kept the same across poses, matching icon sets, text-in-image and UI screens. |
| Moving footage, reference motion, lip-sync to a song | Seedance 2.5 / Kling / Midjourney video (via Higgsfield or direct) | Always credits. Read the real price before every render and say it (references/real-shots.md). Use for reference motion that code redraws (rotoscoping, see overnight-run.md) or for editorial footage under a code HUD. |
| Music | Suno | User generates from your prompt (see audio.md). Ask for full length. |
| Sound effects (hits, whooshes, UI ticks, risers, impacts, chimes) | scripts/sfx.mjs: code-synthesized, free, deterministic |
Default. 10 kinds x variants, placed by exact film frame over the music bed with ducking, a -1 dBTP ceiling and an audibility check (see audio.md). |
| Voice, announcer, organic one-off SFX (animals, crowds, real-world foley) | ElevenLabs (or Higgsfield generate_audio) |
User's subscription or credits. Place on exact hit frames. |
| Upscale, background removal, reframe | Higgsfield tools | Credits; or do keying in code for flat backgrounds (free). |
| Lyrics, word timestamps | whisper.cpp local: whisper-cli -m <model.bin> -f audio.wav -l <lang> -ml 1 -sow (whisper.cpp with the large-v3-turbo model) |
Free. Transcripts of sung lyrics are rough; ask the user for the real lyrics and use the timestamps. |
| Isolate vocals to find exact hits | demucs local: python -m demucs --two-stems=vocals -o sep song.mp3 (demucs in a Python environment) |
Free. Then read vocal onsets from the vocals stem's loudness. |
| Tempo | onset autocorrelation (see audio.md) | Free. Lock the beat grid to the vocal onsets. |
(On any machine, check what is installed and offer to install what is missing.)
Making pictures through Codex (when the user has ChatGPT or Codex)
Codex's command-line tool has image generation built in, so with a ChatGPT plan you can make the pictures yourself, without asking the user to paste anything. It runs on the plan's usage, with no charge per picture; a big batch counts toward the plan's Codex limits and can pause it for a few hours. Check for it before the pitch (codex --version, and the path below). When it is there, it is usually the best free picture route, so recommend it. Ask before the first Codex run, in the pitch with the other services (services.md): say how many pictures, and that they come out of the ChatGPT plan's Codex allowance, not money. Run nothing until they say yes. The yes covers that number of pictures; a batch beyond it is a new question.
- Find it.
codex --version. On a Mac the ChatGPT app ships it inside the app even when nothing is on the PATH:/Applications/ChatGPT.app/Contents/Resources/codex-cli/bin/codexin newer versions,.../Resources/codexin older ones (find /Applications/ChatGPT.app -name codex -type ffinds it).codex features listshould showimage_generationas true. - Set it up, when they have a ChatGPT plan but no Codex. Walk them through it in their own terminal, one step at a time: on a Mac, the ChatGPT desktop app already has it inside (above); otherwise
npm install -g @openai/codex(needs Node), thencodex loginand they choose "Sign in with ChatGPT" in the browser window it opens. They sign in themselves. Then checkcodex --versionandcodex features list(image_generation true) and make one test picture before the batch. - One picture folder per project, named after it, and every run happens there. Give it an
AGENTS.md(Codex reads that file on its own in the folder it runs in) holding the project's picture rules and its style paragraphs (characters, icon set, photos), arefs/folder with the approved pictures, and anout/folder for new ones. Then each prompt only names what to make and which references to attach, the look stays the same across months and batches, and the user finds every run in Codex's history under the project's name. Run:cd <project>-pictures && codex exec --skip-git-repo-check -s workspace-write -c model_reasoning_effort="low" -i refs/<ref>.png - < prompt.txt. When a new picture is approved, copy it intorefs/and name it inAGENTS.md. - What the prompt says. "Use your image generation tool to make N separate images, one at a time, and save each as a PNG in the current folder with the exact file name given (copy each generated file here; do not write any code to draw them). When done, reply with only the saved file paths." Then one shared style paragraph and a numbered list of pictures with their file names. Without "do not write any code" it may draw the picture with code instead.
- Keep a character or a set the same.
-i picture.png(repeat it for several) attaches references: the character's first picture keeps the same face, colours and outfit across new poses; two members of an icon set keep a new icon in the set's style. - Speed. About a minute a picture. Four
codex execruns side by side worked (sixteen pictures in about four minutes). - Cut-outs. Ask for a plain flat white background and cut the subject out with
swift scripts/matte.swift in.png out.png(Mac); it kept white sea foam and a white butter dish whole. A transparent background straight from Codex was not tested. - Check the set as one sheet. Tile every picture on one sheet and look for the one drawn in another style (a hairy, realistic fly in a glossy cartoon set); a viewer reads it as a mistake. Redo it with two members of the set attached as references.
- Pictures that look real are still made by AI. Say so in the delivery note, and tell the person to declare them where the platform asks (YouTube's "altered or synthetic content" box).
Driving the Higgsfield web UI for free images (when the user is logged in)
Chrome MCP on higgsfield.ai/ai/image: pick Nano Banana Pro (badged UNLIMITED), toggle Unlimited ON in the prompt bar (the button then reads "Unlimited ✦"; never toggle Unlimited on GPT Image 2, it opens a paid upsell). Set long prompts via the Lexical editor state in JS instead of typing (typing 2k chars freezes the tab). Upload references through the uploads panel's file input. Queue 2 at a time (the web UI runs ~4 concurrently and the user may be using it too). Fetch results with the free MCP show_generations read and curl them down. Video is not covered by Unlimited.
Seedance lip-sync at scale (public production record, Sep 2026)
Proven on 47 clips + 30 continuations, 86% of sung seconds verified in sync ($93 at 720p via OpenRouter):
- Condition every singing clip on both the plate image and the exact audio window of the song, and quote the words in the prompt:
@Image1 sings @Audio1. She lip-syncs to @Audio1 exactly, every word in time: "…". - The key fact: Seedance copies the reference audio into its own soundtrack, and the mouth follows THAT soundtrack, which drifts from the real vocal partway through. So sync is measurable, not eyeballable: compare the clip's returned soundtrack against the real vocal (spectrogram/alignment) to find the divergence point.
- Cut on the last eighth note before the drift and cover the rest with a continuation clip starting from that frame. In the final mix, replace all clip audio with the real song.
- Dedicated lip-sync models (LatentSync class) are the fallback, not the plan.
- Keep generation inputs in color: plates graded to mono come back grey and lose the image-model look; grade in post.
Characters: stylized in code, realistic from a model. Lesson from a song clip (Sep 2026): pushing a code-drawn human toward realism (7-head proportions, anatomy, hands, close-ups) looked fine in stills but twisted in motion (necks, hands, legs), and the user said the animation quality failed. Code characters should stay stylized (picture-book proportions, simple joints, emotion through posture, light, silhouette and back views). If the user wants realistic people, offer an image/video model for the character with code for everything else.
Seek-compatible libraries (and the ones never to use)
film.html renders frames out of order, so any animation library must be driven by t:
- Anime.js: create timelines with
autoplay: false, thenanim.seek(t * 1000)inrenderFrame. - Theatre.js: hand-tune camera paths and keyframes in its studio UI, then
sequence.position = tinrenderFrame(ship the exported state JSON, not the studio). - Our own
springs.jsfor everything else (closed-form,timing.md). - Never a real-time-only library (GSAP tickers left running, CSS transitions/animations, physics engines stepped by the clock, Lottie in autoplay): they drift from t and fail the gate's determinism check. Motion ideas that need them get rebuilt as functions of t (
motion-kit.md).
Free asset sources
- Mixkit (mixkit.co): free music and SFX, no attribution required; check each item's licence line.
- svgl.app: clean SVG logos of real brands and tools, for inline icons and "works with" rows (real integrations only; brand marks stay the brand's).
- 3dicons.co: CC0 3D icon set, several styles and angles; good for explainer cards.
Deciding code vs. model assets
- Characters that must move a lot (dance, fight, act): code-drawn (watercolor engine) or model-made pose sheets that code animates. Single stills can only slide.
- Recognizable real people: model-made images from their photo (keep kids stylized unless the parent explicitly wants realistic).
- Type, UI, diagrams, HUDs, charts, counters: always code. Crisper and perfectly timed.
- Photoreal footage: a video model, with a code layer on top for intelligence and timing.
- A mix is usually best: model for beauty, code for motion, timing and type.
Web UI mechanics learned the hard way (Sep 2026):
- A page reload switches Unlimited back off. Re-check the toggle before every Generate.
- Reference images from earlier sessions stay attached. Hover each thumbnail and click its top-right X before generating.
- Reference slots clear after every Generate. Uploads take 5-15 s, and a second upload can replace the first instead of adding one; zoom-check the slots before clicking Generate.
- Never click the slot tile itself: it opens the native file picker (Escape closes it). Use
findfor the file input and upload through it. - To reuse a finished image as the reference for the next one (burst pairs, a character's identity), open the + panel, go to the Generations tab and click its tile.
- The window size can change mid-session; re-find elements instead of reusing old coordinates.