oa-skill-watcher

name: oa-skill-watcher description: "Detects newly created or recently modified skills in ~/.claude/skills/ and auto-generates matching agent templates in the Open-Agents library. Use when a new skill has been created, when skills have been updated, or when the user says 'generate template for this skill', 'scan new skills', 'update agent library from skills', or 'watch skill changes'." disable-model-invocation: true user-invocable: true allowed-tools: Glob, Read, Write, Grep, Bash argument-hint: "[skill-name or 'all' for full scan]"

Purpose

Bridges the gap between skill creation and agent library population. When a new Claude Code skill is added to ~/.claude/skills/, this skill detects it and generates a corresponding oa run-compatible agent template in agents/library/open-agents-meta/.

Critical Rules

ALWAYS read the full SKILL.md before generating a template — the skill content determines the systemPrompt.

NEVER generate a template if a functionally identical one already exists in agents/library/. Check first.

NEVER modify the original SKILL.md — this skill is read-only with respect to skill files.

Step 1: Identify Target Skills

If $ARGUMENTS is provided and not all, focus on that specific skill:

ls ~/.claude/skills/$ARGUMENTS/SKILL.md

If $ARGUMENTS is all or empty, scan all skills and detect recent changes:

find ~/.claude/skills/ -name "SKILL.md" -newer ~/.claude/skills/.last-watcher-run 2>/dev/null || \
find ~/.claude/skills/ -name "SKILL.md" -mtime -7

Also scan project-level skills if in an Open-Agents workspace:

find .claude/skills/ -name "SKILL.md" -mtime -7 2>/dev/null

Step 2: Read and Parse Each Skill

For each identified SKILL.md:

1. Read the frontmatter to extract: name, description, allowed-tools, disable-model-invocation, user-invocable
2. Read the markdown body to understand: task type, steps, inputs/outputs, domain
3. Classify the skill:
   • Meta/orchestration: manages agents, spawning, pipelines → open-agents-meta
   • Code/dev: software development, refactoring, testing → code-dev
   • Research: web search, analysis, knowledge gathering → research
   • Review/quality: validation, QA, auditing → review-quality
   • Data: transforms, conversions, parsing → data-transform
   • Communication: messages, reports, docs → communication
   • File system: file management, directory ops → file-system

Step 3: Check for Existing Templates

Before generating, search the agent library for duplicates:

grep -rl "<skill-name>" /path/to/Open-Agents/agents/library/
grep -r "\"<keyword>\"" /path/to/Open-Agents/agents/library/ --include="*.json" -l

Skip generation if a template with >70% conceptual overlap already exists.

Step 4: Determine Model Hint

Select the minimal capable model:

Default to Haiku unless the skill clearly requires reasoning or multi-step judgment.

Step 5: Generate the Agent Template

Create a JSON file following this exact structure:

{
  "id": "<category>-<kebab-case-name>",
  "name": "<Title Case Name>",
  "description": "<One sentence: what it does and when to use it.>",
  "atomic": true,
  "category": "<category-slug>",
  "tags": ["<tag1>", "<tag2>", "<tag3>"],
  "maturity": "tool-capable",
  "modelHint": "<full-model-id>",
  "tools": ["<Tool1>", "<Tool2>"],
  "systemPrompt": "<Self-contained prompt derived from the skill's purpose and steps. Include: role, task, output format, numbered rules. Max 400 words.>"
}

systemPrompt construction rules:

1. Start with: "You are a <role derived from skill name>."
2. Add: "Task: <one sentence from skill description>"
3. Add: "Input: <what the agent receives>" (derive from skill's $ARGUMENTS or context)
4. Add: "Output: <what the agent produces>" (derive from skill's output sections)
5. Add: "Rules:\n1. <First rule>\n2. <Second rule>..." (derive from skill's Critical Rules)
6. Keep under 400 words. Must be self-contained — no references to external files.

tools selection:

• Read-only analysis: ["Read", "Glob", "Grep"]
• Writes files: add "Write", "Edit"
• Runs shell commands: add "Bash"
• Web research: add "WebFetch", "WebSearch"

Step 6: Write the Template

Save to:

<open-agents-repo>/agents/library/<category>/<skill-name>.json

Example paths:

agents/library/open-agents-meta/oa-skill-watcher.json
agents/library/code-dev/my-new-skill.json
agents/library/research/web-analyzer.json

Detect the Open-Agents repo path by checking common locations:

• /mnt/c/Users/Freek Heijting/Documents/GitHub/Open-Agents
• ~/Documents/GitHub/Open-Agents
• $OA_REPO environment variable

Step 7: Update the Watcher Timestamp

After a successful run, update the timestamp file so future runs only check newer skills:

touch ~/.claude/skills/.last-watcher-run

Step 8: Report

Output a summary:

Skill Watcher Report — <date>
──────────────────────────────
Skills scanned: N
New/modified:   N
Templates generated: N
Skipped (duplicate): N

Generated:
  - <skill-name> → agents/library/<category>/<file>.json
  - ...

Skipped:
  - <skill-name> — duplicate of <existing-template>

Quality Gate

Before writing any template, verify:

• id is unique in the library
• systemPrompt is self-contained (no file references)
• tools list is minimal — no tools the agent won't use
• category matches an existing directory in agents/library/
• description starts with a verb and fits one sentence

Reference

• See oa-agent-library-builder — for manual template extraction from agent runs
• See oa-library-templates — for template field specifications
• Agent library categories: agents/library/ in Open-Agents repo