Skip to content

Self-improving AI Employees for everyone

feat(skills): data-skill and vision-skill for the new data + vision tools

trohitgc1e5cf98a325
diff --git a/server/skills/assistant/data-skill/SKILL.md b/server/skills/assistant/data-skill/SKILL.md
new file mode 100644
index 00000000..a8368120
--- /dev/null
+++ b/server/skills/assistant/data-skill/SKILL.md
@@ -0,0 +1,67 @@
+---
+name: data-skill
+description: Read, search, and write raw local data (files, CSV, JSON, PDF, HTML, XLSX, images) across the workspace and operator-mounted folders. Use when the user asks about local files, documents, spreadsheets, logs, or data on disk.
+allowed-tools: data
+metadata:
+ author: opencompany
+ version: "1.0"
+ category: data
+ icon: "πŸ—‚οΈ"
+ color: "#50fa7b"
+
+---
+
+# Data Access Skill
+
+Work with raw local data through the `data` tool. Check the tool in your
+tool list for the full schema; this skill covers strategy, not parameters.
+
+## The two path namespaces
+
+- `reports/q3.csv` β€” workspace-relative (the workflow's own scratch space,
+ always readable and writable)
+- `mnt/<mount_name>/file.csv` β€” an external folder the operator approved;
+ read-only unless its writable flag is on
+
+You MUST start with `list` on an empty path: it returns the workspace root
+AND the mounts enabled on this node (with their writable flags). Never guess
+mount names.
+
+## Reading β€” pick the honest tier, page deliberately
+
+`read` auto-detects text / csv / json / pdf / html / xlsx / image / binary
+from the extension; `as_type` forces a tier when the extension lies.
+
+- Everything is paged with `offset` / `limit`. Check `truncated` and
+ `rows_total` / `lines_total` / `pages_total` before summarizing β€” a
+ truncated read is a sample, not the file.
+- Skim first, then zoom: read with the default limit to see the shape, then
+ page into the region that matters. Do not crank `limit` to the max on the
+ first read of an unknown file.
+- CSV/XLSX return `columns` + `rows`; JSON deep structures are pruned at
+ depth 8 (`pruned_paths` tells you where β€” read a subpath as text if you
+ need what was pruned).
+- Images return metadata plus a file reference β€” if the host model supports
+ vision the image itself becomes visible; otherwise use the `vision` tool.
+- Binary files return a reference and checksum, never content. Pass the
+ reference's path to a tool that can handle the format.
+
+## Writing β€” bounded and non-destructive
+
+- `write` replaces, `append` extends; both are UTF-8 text, workspace or
+ writable mounts only.
+- There is NO delete. Do not try to empty files as a substitute; tell the
+ user deletion is a manual action in the gallery panel.
+- `copy_to_workspace` imports a mount file into the workspace (never
+ overwrites β€” it suffixes). Use it before operations that need a workspace
+ reference: previews, drags, vision on mount images.
+
+## Anti-patterns
+
+- Guessing paths instead of `list`/`search` first.
+- Re-reading a whole file when you already have the page you need β€” results
+ persist in your context; reread only new offsets.
+- Reading a huge JSON as `json` when you need one field β€” read as text and
+ page, or search first.
+- Asking for a mount that is not in your `list` result: mounts are
+ operator-controlled; you cannot add them.
diff --git a/server/skills/assistant/vision-skill/SKILL.md b/server/skills/assistant/vision-skill/SKILL.md
new file mode 100644
index 00000000..e2de8f2a
--- /dev/null
+++ b/server/skills/assistant/vision-skill/SKILL.md
@@ -0,0 +1,58 @@
+---
+name: vision-skill
+description: Understand images (describe content, answer questions about them, extract text/OCR) via a vision model. Use when the user asks what an image, screenshot, chart, scan, or photo contains.
+allowed-tools: vision data
+metadata:
+ author: opencompany
+ version: "1.0"
+ category: vision
+ icon: "πŸ‘οΈ"
+ color: "#bd93f9"
+
+---
+
+# Vision Skill
+
+See images through the `vision` tool (a vision-capable model looks at the
+image and answers in text). Check the tools in your tool list for full
+schemas; this skill covers strategy.
+
+## Workflow
+
+1. Locate the image with the `data` tool first (`list` / `search`) β€” the
+ `vision` tool takes a workspace-relative path, and guessing paths wastes
+ a paid model call.
+2. An image on an external mount (`mnt/...`) must be imported first:
+ `data copy_to_workspace`, then pass the returned workspace path.
+3. Choose the operation:
+ - `describe` β€” what is in the image; pass `question` for something
+ specific ("what is the y-axis peak?") instead of asking for a general
+ description and hoping.
+ - `extract_text` β€” OCR-style text extraction (documents, screenshots,
+ signs). Returns only the text.
+
+## The budget ladder
+
+`budget` controls resolution and therefore cost: `small` (~256 visual
+tokens), `normal` (~1024, default), `large` (~2048, fine detail).
+
+- Skim at `normal`; escalate to `large` only when the answer needs fine
+ detail (dense charts, small print, UI screenshots).
+- Never open with `large` on an image you haven't seen at `normal` β€” each
+ call bills the vision provider.
+
+## When NOT to use this tool
+
+- The host model may already see images natively: if an image you read via
+ the `data` tool is visible in the conversation, answer from it directly β€”
+ a delegate call would be a second opinion at extra cost and latency.
+- Extracting text from a PDF: use `data read` (text extraction) first;
+ vision OCR is for rasterized/scanned content the parser cannot read.
+
+## Anti-patterns
+
+- Calling `vision` on a path that was never confirmed via `data` list/read.
+- Repeating `describe` with rephrased questions β€” ask one specific
+ `question` with the detail you actually need.
+- Using `extract_text` on charts to infer numbers β€” ask `describe` with a
+ targeted question instead; OCR on plot labels loses structure.