autoresearch · pedronauck — Claude Code skill: what it does, how to install
autoresearch runs another skill through repeated test cycles, grades each output against pass/fail checks, and rewrites its prompt until the score stops climbing.
no license 152 installs ★ 655 updated 2026-10-20
npx skills add pedronauck/skills --skill autoresearch -a claude-codeWhat autoresearch does
The skill takes the path to a target SKILL.md, a handful of test prompts, and 3–6 yes/no checks, then loops: generate outputs, grade them against the checks, tweak the wording, keep edits that raise the pass rate and drop the rest. Before any run starts, the agent has to confirm with the user which skill to tune, what test scenarios to cover, what the pass/fail checks are, how many runs per experiment (5 by default), and how often cycles repeat (every 2 minutes by default).
Before touching anything, the agent reads the full SKILL.md and any linked files under references/ to understand the skill's job, then turns the user's criteria into checks written as a question plus explicit pass and fail conditions. The first pass runs the skill unchanged as experiment zero, its score becomes the baseline, and the original file is copied to SKILL.md.baseline.
Every experiment gets logged to results.tsv and results.json, feeding a self-contained HTML dashboard with a score-progression chart, a per-experiment table, and a breakdown of which checks pass or fail most often; it auto-opens in the browser and refreshes as new runs land. If the baseline already lands near 90%, the skill flags that to the user before continuing.
What it helps with
- find which inputs break an existing skill
- turn vague quality bars into testable pass/fail checks
- auto-tune a skill's prompt instead of editing it by hand
- track pass-rate progress across experiments on a chart
- log which prompt edits helped and which got discarded
Example prompts for the agent
- Run autoresearch on the skill at
~/.claude/skills/pdf-report/SKILL.md - Benchmark this skill with evals until the pass rate plateaus
- Improve this skill — set up 5 test scenarios and binary checks
How to install autoresearch
1Skills CLI (Vercel)
npx skills add pedronauck/skills --skill autoresearch -a claude-code2Manually: put the files in place
~/.claude/skills/autoresearch/SKILL.mdCopy the whole skill folder; for one project use .claude/skills/autoresearch/
1Skills CLI (Vercel)
npx skills add pedronauck/skills --skill autoresearch -a codex2Manually: put the files in place
.agents/skills/autoresearch/SKILL.mdFor all projects use ~/.agents/skills/
1Skills CLI (Vercel)
npx skills add pedronauck/skills --skill autoresearch -a cursor2Manually: put the files in place
.cursor/skills/autoresearch/SKILL.mdCursor also reads .agents/skills/ and .claude/skills/
1GitHub CLI
gh skill install pedronauck/skills skills/curated/autoresearch2Manually: put the files in place
.github/skills/autoresearch/SKILL.md1Skills CLI (Vercel)
npx skills add pedronauck/skills --skill autoresearch -a gemini-cli2Platform CLI
gemini skills install https://github.com/pedronauck/skills --path skills/curated/autoresearchManually: put the files in place
.devin/skills/autoresearch/SKILL.mdThe legacy .windsurf/skills/ path still works
Skills CLI (Vercel)
npx skills add pedronauck/skills --skill autoresearch -a opencodeClaude.ai
Zip the folder and upload it in Settings → Capabilities → Skills
Step by step, with commands for each platform: How to install skills in Claude Code: five ways
Questions about autoresearch
How do I install autoresearch in Claude Code?
Run npx skills add pedronauck/skills --skill autoresearch -a claude-code; the skill lands at ~/.claude/skills/autoresearch/SKILL.md.
What does autoresearch do?
It runs a target skill against test inputs, grades every output with binary pass/fail checks, and mutates the skill's prompt automatically, keeping only the versions that raise the pass rate.
What do I need before running autoresearch?
The exact path to the target skill's SKILL.md, 3 to 5 varied test scenarios, and 3 to 6 binary quality checks — the agent won't start experiments without these.
Similar skills
All skills for Agent setup: memory, context, orchestration →autoresearch by other authors comes first.