Skill Creator

Problem

Building Claude skills usually fails in three practical places: trigger conditions are vague, results lack a comparable baseline, and description optimization overfits or misses real queries. skill-creator-team turns skill creation into an engineering loop instead of a one-shot prompt.

How It Works

• Capture intent: clarify what the skill does, when it triggers, expected output, edge cases, dependencies, and success criteria.
• Draft SKILL.md: organize name, description, compatibility, and body, with trigger context and reasoning in the description.
• Test and evaluate: run with-skill and baseline variants, draft verifiable assertions, aggregate benchmark.json, and use generate_review.py to inspect qualitative output and quantitative metrics.
• Iterate: generalize user feedback, remove ineffective instructions, explain rationale, and re-run tests until results are satisfactory.
• Optimize description: create should-trigger and should-not-trigger queries, then select a stable description using train and holdout sets.

Boundaries

It works best in environments such as Claude Code or Cowork where subagents and file systems are available. In Claude.ai, it degrades to serial runs, manual review, and limited baseline comparison. Skill content must be safe and non-misleading, and evaluation queries should be concrete enough to test real triggering behavior.