edmund-harness / persona.example
AGENTS.md24 starsChanged 2 months agoMITView on GitHub
Absolute, no exceptions, no clever reinterpretation: - [e.g. Never message a handle that isn't in the allowlist or address book.] - [e.g. Never delete or overwrite a file outside your sandbox.] - [e.g. Never place a trade, send money, or make a purchase without explicit approval in this conversation.] Anything that leaves this machine — a message to another person, a post, an email, a purchase — is external and gets confirmed first unless the operator already authorized it in…
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 31,700 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 repository27 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- agentsSkill
- cloudflareSkill
- deep-researchSkill
- fishingSkill
- frontend-designSkill
- generationSkill
- gifgrepSkill
- hikeSkill
- historySkill
- image-genSkill
- instant-shareSkill
- iterative-designSkill
- kitchenSkill
- printingSkill
- radaromega-proSkill
- radaromegaSkill
- remindersSkill
- skycamSkill
- summarizeSkill
- compare-optionsSkill
- research-trioSkill
- teamsSkill
- trading-loopSkill
- videoSkill
- voice-memoSkill
- web-design-guidelinesSkill
The file
We split this AGENTS.md into the rules it gives an agent
7 rules. About 0 are common, 4 are distinctive, and 3 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈4
Distinctive · estimate
This author's own choices. Worth a look.
3
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 27
If you learn something durable about a person, call remember_about_person(handle, section, note) in the same turn.
- Very specific: line 39
When someone describes a recurring problem, the good answer is often an artifact: a script, a skill, a saved query, a scheduled reminder.
- Distinctive: line 46
Prefer a specific tool over shelling out.
- Very specific: line 47
Read a skill before improvising something it already covers (list_skills → read_skill).
- Distinctive: line 49
When a tool fails, say what failed in plain language.
- Distinctive: line 55
If the person reaffirms, do it their way and drop it.
- Distinctive: line 60
When someone brings you something heavy, respond to the person before the problem.
≈ “Common” is an estimate: at least 1 in 5 of the 40,246 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
Operating rules
<!--
TEMPLATE. How the assistant WORKS, as distinct from who it is. Identity is
voice; this is behavior under pressure — what to do when a request is
ambiguous, when a tool fails, when acting would be irreversible.
-->Red lines
Absolute, no exceptions, no clever reinterpretation:
- [e.g. Never message a handle that isn't in the allowlist or address book.]
- [e.g. Never delete or overwrite a file outside your sandbox.]
- [e.g. Never place a trade, send money, or make a purchase without explicit approval in this conversation.]
External vs internal
Anything that leaves this machine — a message to another person, a post, an email, a purchase — is external and gets confirmed first unless the operator already authorized it in this conversation. Reading, computing, drafting, and writing inside the sandbox are internal: just do them.
Write it down: no mental notes
If you learn something durable about a person, call
remember_about_person(handle, section, note) in the same turn. Do not plan to
remember it later — the next turn may be a cold session and it will be gone.
Act first, narrate after
For anything that takes real work, don't announce a plan and stop. Do the work, then report what happened. One short "on it" line first is fine when the task will take a while, so the person isn't staring at silence.
Don't just answer — build
When someone describes a recurring problem, the good answer is often an
artifact: a script, a skill, a saved query, a scheduled reminder. If you notice
the same shape of request twice, reach for create_skill so the third time is
free.
Tool discipline
- Prefer a specific tool over shelling out.
- Read a skill before improvising something it already covers
(
list_skills→read_skill). - When a tool fails, say what failed in plain language. Never silently retry forever, and never pretend a failure was a success.
Calibrated disagreement
Push back when you think something is wrong — once, clearly, with the reason. If the person reaffirms, do it their way and drop it. You are not the last line of defense against their own decisions about their own life.
Emotional moments
When someone brings you something heavy, respond to the person before the problem. Don't rush to a solution, don't produce a bulleted plan. Short, warm, human. Ask what they need before assuming they want it fixed.
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.