One file. Your AI coding agent becomes a scientist.
Install as a Claude Code plugin, or drop skills/researcher/SKILL.md into Codex, Cursor, or any agent that reads markdown skills. The agent designs experiments, tests hypotheses, discards what fails, keeps what works — 30+ experiments overnight while you sleep.
Install
Claude Code plugin (recommended)
Two slash commands inside Claude Code — first registers this repo as a marketplace, second installs the plugin from it:
Restart Claude Code (or run /plugin reload) and trigger the skill with /researcher or by asking the agent to run a research loop on something.
To upgrade later: /plugin marketplace update researcher-marketplace then /plugin install researcher@researcher-marketplace again.
Single-file drop-in (any agent)
The canonical skill body is skills/researcher/SKILL.md in this repo (one file, ~300 lines, frontmatter-tagged). Copy it into your agent's skill directory:
The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond.
First-time install walkthrough for Claude Code, Codex CLI, and ChatGPT.
~/.claude/skills/researcher/SKILL.md
Claude Code (project-level):.claude/skills/researcher/SKILL.md in your repo
Codex / other agents: wherever your tool reads skills or instructions from (consult its docs)
Trigger with /researcher (Claude Code) or by asking the agent to enter "researcher mode".
What it looks like running
Experiment b4 — READ/WRITE phase separation
Branch: research/graph-protocol-optimization · Parent: #b1 · Type: real
Hypothesis: Agents read architectural rules but treat them as optional. Separating the instruction into a READ phase ("load constraints first") and a WRITE phase ("now implement") with a guard ("if you haven't done READ, stop") should improve compliance.
Changes: restructured agent rules into explicit READ/WRITE phases, added structural guard
Result: 7.04/10 (was 1.82 baseline, 5.91 best) — new bestStatus: keep
Insight: Every attempt to add verification checklists regressed. What worked was changing the structure, not adding steps. Agents respond to framing, not policing.
b0: baseline (no special instructions): 1.82/10. keep.
b1: reframe rules as "constraints, not suggestions": 5.91. keep.
Real experiment from optimizing Yggdrasil agent rules. The skill works on any codebase.
Same loop, different problems:
npm run build takes 40s → agent gets it to 18s
prompt returns wrong format 30% of the time → agent gets it to 3%
API p99 is 200ms → agent finds the bottleneck and cuts it to 80ms
document parser misses edge cases → agent improves match rate from 74% to 91%
How it works
The agent interviews you about what to optimize, sets up a lab on a git branch, and works autonomously. Thinks, tests, reflects. Commits before every experiment, reverts on failure, logs everything.
It detects when it's stuck and changes strategy. Forks branches to explore different approaches. Keeps going until you stop it or it hits a target. Resume where you left off across sessions.
Generalizes autoresearch beyond ML. Works on any problem where you can measure a result — code, configs, prompts, documents.
All experiment history lives in an untracked .lab/ directory. Git manages code. .lab/ manages knowledge.
Want the full walkthrough? Read the guide. It walks through a complete example from start to finish.
FAQ
How is this different from autoresearch?
Autoresearch's core loop is universal, but the repo is wired to train.py, val_bpb, and GPU training. To use it on something else you'd rewrite the setup. This gives you that loop ready to go for any codebase.
When would I use this instead of ML?
It's not instead of ML. ML is one possible domain. This works on anything where the agent can try things, measure, and iterate. Code, scripts, documents, configs. Slow builds, flaky tests, API latency, prompt accuracy.
How does it measure success for non-ML code?
Whatever you can measure. Test pass rate, benchmark output, type check errors, build time. You set it up in the discovery phase. The agent asks what to measure and how. If you can run a command and get a number, that's your metric. For cases where there's no command to run, the agent scores against a qualitative rubric you define together.
How does convergence detection work?
The agent checks a table of signals after every experiment. If it sees 5+ failures in a row, a metric plateau, or the same area modified too many times, it knows to change approach. Some signals are advisory (consider pivoting), others are hard guardrails (you must pivot). Details in the guide.
Can it improve itself?
Sort of. The skill was optimized using the skill itself. A research document about how LLMs process instructions (attention decay, primacy/recency, instruction budgets) was used as criteria, and the agent ran the loop against its own prompt. Not fully recursive, but the loop was: research → skill → use skill to improve skill.
Can't I just ask Claude to build this from the autoresearch repo?
You can try. This saves you the work and includes things autoresearch doesn't have: thought experiments, non-linear branching, convergence detection, qualitative metrics, and session resume.
License
MIT
See also
Yggdrasil — the agent experiments on your code. But does it understand what it's working on? Semantic memory for repositories.
Claude Code is an agentic coding tool that lives in your terminal, understands your codebase, and helps you code faster by executing routine tasks, explaining complex code, and handling git workflows - all through natural language commands.
