Skill

skill-lab-autoresearch

Iteratively refines copied SKILL.md candidates in private labs with explicit evaluation goals, rubrics, one-at-a-time revisions, notes, and manual promotion recommendations.

developer-tools

code-quality

npx claudepluginhub junjslee/episteme --plugin episteme

Tool Access

This skill uses the workspace's default tool permissions.

Preview

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

SKILL.md

Similar Skills

autoresearch

Autonomously optimizes SKILL.md prompts via Karpathy-style mutate/score/keep loop using test inputs, checklists, and git-based versioning.

darkroom

auto-optimize

Autonomously optimizes Claude Code skills by iteratively running them on test inputs, scoring against binary evals, reflecting on failures to mutate prompts, and archiving improvements. Invoke via /auto-optimize for skill enhancement or autoresearch.

1 file

skill-creator-pro

darwin-skill — Autonomous SKILL.md Optimizer

Optimizes SKILL.md files for Claude Code by scoring across 8 dimensions (YAML completeness, triggers, structure, clarity, tests), proposing improvements, testing changes, and git-reverting non-improvements.

aradotso-trending-skills-37

Stats

Stars16

Forks2

Last CommitApr 20, 2026

Actions

View Source View Plugin View on GitHub View README

Help us improve

Share bugs, ideas, or general feedback.

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

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

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

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.