skill-lab-autoresearch

name	skill-lab-autoresearch
description	Private Claude-only lab skill for iteratively improving a copied skill candidate with explicit evaluation goals and manual promotion.

Use this skill only for private skill-lab work inside ~/episteme.

This wrapper is inspired by olelehmann100kMRR/autoresearch-skill, but it is adapted for a private lab workflow where live skills are never edited directly.

Required Inputs

Before doing any work, collect or confirm:

• target_candidate_path: path to a copied candidate SKILL.md under labs/skills/<skill-name>/candidate/
• eval_goal: what improvement is being optimized
• iteration_cap: maximum number of iterations
• output_dir: path under labs/skills/<skill-name>/runs/

If any of these are missing, stop and ask for them.

Hard Refusals

Refuse the task if target_candidate_path points anywhere inside:

• skills/custom/
• skills/vendor/
• ~/.claude/skills/
• ~/.codex/skills/

Refuse the task if output_dir is not inside labs/skills/<skill-name>/runs/.

Refuse any request to overwrite a live skill automatically.

Workflow

1. Validate the target path is a lab candidate copy.
2. Read the candidate SKILL.md.
3. Clarify the evaluation objective and the iteration limit.
4. Propose a compact evaluation rubric tied to the stated goal.
5. Generate one revised candidate at a time.
6. Record observations, tradeoffs, and reasons for each revision under output_dir.
7. Stop when:

• the iteration cap is reached
• the objective appears satisfied
• further changes are speculative

8. Finish with:

• the current best candidate path
• an evaluation summary
• a promotion recommendation

Outputs

Produce:

• revised candidate content or patch guidance for the lab copy
• concise evaluation notes
• a recommendation: keep iterating, promote manually, or discard

Promotion Rule

Promotion into the live skill set is always manual.

Do not copy changes into skills/house or skills/vendor.