research-harness-template
AGENTS.md1 starsChanged 2 months agoMITView on GitHub
You are working in a research repo. The researcher owns the questions, the conclusions and the declared interests. You carry the work between those points. Read method/README.md: ten ground rules and a close-out check. Read method/templates.md before creating an investigation or a study. Read knowledge/README.md before touching knowledge/registry.json. Walk them through the first hour in README.md yourself: run the three commands, summarise method/README.md in a few sentences, then start .agents/skills/new-investigation/SKILL.md. Do not send them off to read; ask and build.
Where files are headingNew AGENTS.md files have more rules about keeping secrets out of the code, and fewer about code style.
Of the AGENTS.md files we read, the 33,315 first seen in the last six months, against the older ones: the kind of rule that gained most share, and the kind that lost most. About files like this in general, not this one, and an estimate, since a kind is a pattern in the words.
In this repository3 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- close-outSkill
- new-investigationSkill
The file
We split this AGENTS.md into the rules it gives an agent
8 rules. About 0 are common, 3 are distinctive, and 5 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈3
Distinctive · estimate
This author's own choices. Worth a look.
5
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 7
Read method/README.md: ten ground rules and a close-out check.
- Distinctive: line 23
Write the competing hypothesis and freeze the protocol before looking at data.
- Very specific: line 24
Never edit anything under a frozen/ directory.
- Distinctive: line 25
Never call a paid model or an external API from a study unless the researcher said so in this session.
- Very specific: line 26
Never commit identifiable personal data.
- Very specific: line 27
Every claim you add points at a preserved excerpt and has a concrete reconsider_if.
- Very specific: line 28
When a rule, template or register convention changes, write it into DECISIONS.md, CHANGELOG.md and the file that carries the rule, in the …
- Distinctive: line 36
Write in the language the researcher writes in.
≈ “Common” is an estimate: at least 1 in 5 of the 42,283 AGENTS.md files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Instructions for agents
You are working in a research repo. The researcher owns the questions, the conclusions and the declared interests. You carry the work between those points.
Before anything else
Read method/README.md: ten ground rules and a close-out check. Read method/templates.md before creating an investigation or a study. Read knowledge/README.md before touching knowledge/registry.json.
If the researcher is new here
Walk them through the first hour in README.md yourself: run the three commands, summarise method/README.md in a few sentences, then start .agents/skills/new-investigation/SKILL.md. Do not send them off to read; ask and build.
Skills
- To start an investigation, follow
.agents/skills/new-investigation/SKILL.md. - To close an investigation or a session, follow
.agents/skills/close-out/SKILL.md.
Codex discovers these as $new-investigation and $close-out. Other agents open the file and follow it.
How to work
- Drive the work without asking permission at every step. Stop and ask when a choice would change an investigation's direction: the question, the competing hypothesis, a threshold, what counts as evidence.
- Write the competing hypothesis and freeze the protocol before looking at data. If the researcher wants to look first, say that rule 3 is about to be broken and let them decide.
- Never edit anything under a
frozen/directory. If a protocol is wrong, write a new study. - Never call a paid model or an external API from a study unless the researcher said so in this session. The harness makes no calls.
- Never commit identifiable personal data. Interview material enters
knowledge/snapshots/only after the researcher has decided on de-identification. Example data is simulated and labelleddata_kind: simulated. - Every claim you add points at a preserved excerpt and has a concrete
reconsider_if. - When a rule, template or register convention changes, write it into
DECISIONS.md,CHANGELOG.mdand the file that carries the rule, in the same commit.
Changing the repo itself
Before every commit, python3 -m harness check --staged must pass. It refuses a commit that changes or deletes anything under frozen/, and a commit that changes method/, .agents/, AGENTS.md, harness/study.py or harness/registry.py without a line in CHANGELOG.md. Enable it as a hook once per clone: git config core.hooksPath .githooks. It also runs in GitHub Actions on every push. Never commit with --no-verify unless the researcher asked and the commit message says why.
Language
Write in the language the researcher writes in. Keep file names, JSON keys and commands in English so the harness and templates keep working.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

