autoresearch-loop
Domain-agnostic metric-driven improvement loop, generalizing Karpathy's autoresearch.
Connecting the steps of a workflow
What it does
Generalize Karpathy's autoresearch into a domain-adaptive improvement loop. The agent discovers what to measure, then runs a disciplined propose → trial → keep-or-revert loop, maintaining an explicit ledger of what was tried, kept, discarded, and implemented.
Use when you want an agent to discover what to measure for a project/goal, then run a keep-or-revert experiment loop that proposes changes,…
You don’t need to name it. Ask for the job in your own words — your AI loads the skill when the request matches.
What it can touch
Includes 7 scripts
Your AI can run these when the skill is used — on your computer in a coding tool (which asks first unless you have told it not to), or in the AI’s own sandbox in Claude and ChatGPT.
How we know
Found in the skill’s folder: examples/code-perf-demo/.auto/checks.sh, examples/code-perf-demo/.auto/measure.sh, examples/code-perf-demo/bench.js, examples/code-perf-demo/gen.js….
No known risk pattern found
Launch7’s automated check read all 23 text files in this skill and matched none of its risk patterns. This is not a guarantee: a check like this can miss new or disguised tricks, and the author can change the files after the check.
How we know
Read 2026-10-05, commit 439cd3a by pipeline/trust/analyze.py — 40+ patterns for hidden text, prompt injection, downloads that run, credential access and more. It reads; it never runs the skill.
Recently maintained
Its repository has had changes recently.
How we know
Last push to jdrhyne/agent-skills: 2026-08-30 (36 days before the scan).
Read automatically from the skill’s own files on GitHub on 5 Oct 2026. Launch7 does not run or manually review skills: this lists what the check saw, and nothing here means “safe”. How we check
Add it to your AI
Partly. Upload it in Claude’s settings — the instructions work; the scripts need the full folder.
Built in your browser from the SKILL.md on GitHub — nothing passes through Launch7.
In Claude, open Customize → Skills, press +, then Create skill → Upload a skill, and choose the .zip.
Make sure code execution is switched on (Settings → Capabilities). Skills need it.
Start a new chat and ask for the job in your own words — Claude picks the skill up when it fits.
This skill ships 7 scripts beside its instructions. The download here is the SKILL.md alone — enough for the instructions, not for the parts that run code. Use the command route, or take the full folder from the repository.
How Claude does this: support.claude.com · checked 28 Sep 2026 · “Supported” means documented by the vendor; Launch7 has not run it end to end
More from jdrhyne
All from this publisher →- clawddocs
OpenClaw documentation expert with decision tree navigation, search scripts, doc fetching, version tracking, and config snippets for all…
- design-audit
Premium UI/UX design auditor with Jobs/Ive philosophy.
- elegant-reports
Generate beautifully designed PDF reports with a Nordic/Scandinavian aesthetic.
- gallery-scraper
Bulk download images from login-protected gallery websites using an attached browser session.
- gsc
Query Google Search Console for SEO data - search queries, top pages, CTR opportunities, URL inspection, and sitemaps.
