skill-craft
Installation
SKILL.md
skill-craft
Create skills through baseline-first evals, or audit existing ones. Judgment work (intake, design, eval authoring, interpretation, edits) happens here; everything checkable (validation, trigger probes, run spawning, grading, verdicts, stats, packaging) runs through the skill-eval MCP tools — never fill in compliance verdicts by hand.
Mode router
| Intent | Mode | Load |
|---|---|---|
| Build something new ("create/write a skill for X") | CREATE | Pipeline below; references per phase |
| Audit an existing skill (path to a SKILL.md or skill directory) | REVIEW | references/review-skill.md in this skill directory |
| Audit the prompt body wording inside an agent definition, or a dispatch brief | Redirect | Invoke oberskills:prompt REVIEW (the file as an artifact — structure, frontmatter, evals — stays here) |
| Improve an existing skill | CREATE from phase 3 (BASELINE), with an old_skill snapshot as the baseline config |
references/eval.md in this skill directory |
| Unclear or ambiguous request | Ask, or invoke oberskills:clarify for structured decomposition |
— |