← ClaudeAtlas

skill-autoresearchlisted

Validates and improves agentic skills by running isolated executor agents against test briefs, scoring confusion, fixing unclear instructions, and rerunning until stop criteria pass. Use when asked to test or evaluate a skill, find skill failure modes, or harden a skill before reuse. Inspired by Karpathy's autoresearch.
souljorje/agent-skills · ★ 4 · AI & Automation · score 68
Install: claude install-skill souljorje/agent-skills
# Skill: Skill Autoresearch ## When to Use - After building or significantly updating a skill, to validate it works end-to-end - When a skill has known confusion points or failure modes - Before promoting a skill from draft to production use - When onboarding a skill written by someone else (stress-test before trusting) - Periodically, to catch drift between SKILL.md instructions and actual tool behavior ## How It Works ``` ┌──────────────────────────┐ │ SETUP (one-time) │ │ │ │ 1. Choose target skill │ │ 2. Create test brief │ │ 3. Define stop criteria │ └──────────┬───────────────┘ │ ┌─────────────────────▼─────────────────────┐ │ │ │ ITERATION LOOP │ │ │ │ ┌──────────────────────────────────┐ │ │ │ Executor Agent (worktree) │ │ │ │ - Reads target SKILL.md │ │ │ │ - Follows procedure step by step │ │ │ │ - Uses test brief as input │ │ │ │ - Reports confusion log per step │ │ │ └──────────────┬──