Skill Creator

Use this skill as the owner for skill design, cleanup, and evaluation.

Core workflow

1. Classify the request:
   • new skill
   • existing skill cleanup
   • trigger-description optimization
   • eval / benchmark request
2. Keep the scope concrete:
   • what the skill should do
   • when it should trigger
   • which files are canonical
   • whether scripts / references / assets are actually needed
3. Prefer lean SKILL.md files:
   • keep only non-obvious workflow rules
   • move detailed procedures into references/
   • keep name and description as the trigger-defining layer
4. Validate before claiming completion:
   • run scripts/quick_validate.py with the install-time core-skills Python runtime available
   • use representative prompts or comparisons when the change is meaningful
5. Optimize trigger wording last:
   • first make the workflow correct
   • then tighten the description so the skill fires reliably

Reference routing

Skill schema and benchmark data

Read references/schemas.md when you need:

• eval input schema
• benchmark output shape
• grading data structure

Evaluation helpers

Read these only when the task includes measurement:

• agents/grader.md
• agents/analyzer.md

Execution helpers

Use these when the task reaches validation or packaging:

• scripts/quick_validate.py
• scripts/aggregate_benchmark.py
• scripts/improve_description.py
• eval-viewer/generate_review.py

Working rules

• Keep one canonical skill file per real capability.
• Remove stale aliases, duplicated wording, and dead examples when they no longer help triggering.
• Do not leave detailed operational docs inside SKILL.md if a reference file can carry them.
• Prefer realistic eval prompts over toy examples.
• Explain results in plain language, not benchmark jargon alone.