v2: Skill doc rewriter - generate improved docs from model feedback

[EmZod]

Feature

After running a focus group test, generate an improved version of the skill document that incorporates model feedback.

Current Behavior

Models report:

• Confusions
• Potential failures
• Suggested improvements

User manually updates their skill doc based on this feedback.

Proposed Behavior

After collecting model feedback, synthesize it into a rewritten skill document that:

1. Fixes identified confusions - Clarify ambiguous sections that multiple models flagged
2. Optimizes progressive disclosure - Restructure so agents get essential info first, details later
3. Improves examples - Add/improve examples where models indicated confusion
4. Better references - Add cross-references and context where models needed more info
5. Clearer structure - Reorganize based on how models actually parse the document

Output

focus-group test ./SKILL.md "task" --rewrite

# Outputs:
# 1. Normal model feedback (current behavior)
# 2. NEW: ./SKILL.improved.md - rewritten skill doc

Implementation Notes

• Use a capable model (e.g., claude-3-5-sonnet) for synthesis
• Feed it: original doc + all model feedback + rewrite instructions
• May need structured output from test models to make synthesis reliable
• Consider iterative refinement (run focus group on improved doc)

Success Criteria

• Improved doc scores better on subsequent focus group runs
• Reduced "confusions" count across models
• Maintains original intent while improving clarity