skills
AGENTS.md1 starsChanged 6 days agoMITView on GitHub
Read this before adding or editing a skill. It applies to humans and to agents. One hard rule: the skill must contain no executable code. That is the whole admission test and it is the only thing every skill here has in common. Topics vary on purpose. A skill for scaffolding a Mirth plugin sits next to one for writing a postmortem because both are useful and both are prose. Do not look for a unifying theme, and do not…
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 32,671 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 repository33 other files it gives its agents.
- hands-offSkill
- hill-climbSkill
- optimize-agent-instructionsSkill
- asset-gardenSkill
- clean-code-stewardSkill
- css-override-censusSkill
- debt-ledgerSkill
- find-component-reinventionSkill
- find-contract-driftSkill
- find-dead-codeSkill
- find-dead-cssSkill
- find-dependency-overlapSkill
- find-reinventionSkill
- find-token-driftSkill
- optimize-commentsSkill
- optimize-docsSkill
- pattern-censusSkill
- test-gardenSkill
- theme-paritySkill
- type-rhythmSkill
- vocabulary-driftSkill
- ui-feature-specSkill
- oie-channel-code-reviewSkill
- oie-plugin-developmentSkill
- batch-workSkill
- deep-code-reviewSkill
- triage-cardSkill
- write-testsSkill
- postmortemSkill
- brainstormSkill
- debate-planSkill
- steelmanSkill
- unslopSkill
The file
We split this AGENTS.md into the rules it gives an agent
7 rules. About 0 are common, 6 are distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈6
Distinctive · estimate
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 3
Read this before adding or editing a skill.
- Distinctive: line 15
If a stranger cannot use it, it belongs in your own dotfiles.
- Distinctive: line 33
If a skill genuinely needs a deterministic tool, see generated tools below.
- Very specific: line 89
Keep SKILL.md under roughly 150 lines.
- Distinctive: line 93
When a skill genuinely needs a deterministic tool, ship a specification rather than an implementation.
- Distinctive: line 120
No em dashes. Use a period or a comma.
- Distinctive: line 128
Read the skill as if you use a different agent than the author.
≈ “Common” is an estimate: at least 1 in 5 of the 41,581 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
Contributing
Read this before adding or editing a skill. It applies to humans and to agents.
What belongs here
One hard rule: the skill must contain no executable code. That is the whole admission test and it is the only thing every skill here has in common.
Topics vary on purpose. A skill for scaffolding a Mirth plugin sits next to one for writing a postmortem because both are useful and both are prose. Do not look for a unifying theme, and do not reject a good skill for lacking one.
What does get rejected: anything that only makes sense inside one company, one codebase, or one machine. If a stranger cannot use it, it belongs in your own dotfiles.
The rules
Rule 1 is what the repo is. Rules 2 to 4 are quality bars.
1. No executable files, ever
No .mjs, no .ps1, no .sh, no .py. Not in a skill folder, not in scripts/, not anywhere.
npx skills add copies a skill folder into your agent's skills directory, and the agent then does what
the skill body says. If the body says to run the script sitting next to it, the agent runs it. One
command should never be able to do that, so the guarantee this repo makes is that there is nothing here
to run.
This is not a style preference. It is the guarantee the repo exists to make, and it is the reason someone can install everything here without auditing it first.
If a skill genuinely needs a deterministic tool, see generated tools below.
2. No agent-specific names
Banned from skill bodies:
- Model names. Not
Opus,GPT-5,Gemini, or a version number. - Tool names. Not "the Task tool", "the Skill tool", "call Read".
- Agent file names as if universal. Say "your agent's context file", not
CLAUDE.md. - Absolute paths, and anything under
~/.claude,~/.cursoror equivalent. - Harness concepts. Turn semantics, background tasks and subagent lifecycles differ per agent.
Say what capability you need, not which product provides it. "If you can run work in parallel" beats "spawn subagents".
3. Degrade gracefully
State the fallback explicitly. A skill that fans out to eight workers must say what to do on an agent with none, and the fallback has to actually work rather than being a shrug.
4. Every skill ships its own check
A skill that tells the agent to do something must tell it how to know it worked. "Looks right" is not a check.
A check that cannot run must fail, not pass. Never write a step that skips or returns early because a dependency is missing. A procedure that completes having verified nothing is worse than no procedure, and it hides for months.
Format
skills/<name>/
SKILL.md required
reference/*.md optional, for detail that would bloat the bodyTwo fields are required:
---
name: write-tests
description: >-
What it does, then when to use it, then when not to.
---Agent-specific frontmatter keys are allowed on top of those two. An agent that does not recognise a key
ignores it, so disable-model-invocation: true costs nothing on an agent that has no such concept. This
is the one place agent-specific naming is fine, because it degrades gracefully by construction. The rule
in rule 2 is about the body, where a name the reader has to translate is a
real bug.
The description is a retrieval surface, not a summary. An agent decides whether to load the skill from
this text alone, so lead with the trigger conditions and include the phrases a user would actually say.
Say what it is NOT for when the boundary is easy to get wrong.
Keep SKILL.md under roughly 150 lines. Push detail into reference/.
Generated tools
When a skill genuinely needs a deterministic tool, ship a specification rather than an implementation. The agent generates the tool on the user's machine and proves it correct before using it.
skills/<name>/
SKILL.md
reference/
ALGORITHM.md pseudocode, precise enough to implement from
CONTRACT.md invocation signature, exit codes, output shape
fixtures/
01-basic.in / .out byte-exact expected output
02-empty.in / .out the edge cases that actually bite
03-malformed.in / .outThe skill body instructs: detect the runtime, implement ALGORITHM.md to satisfy CONTRACT.md, run it
against every fixture, and stop and report if any fixture fails. Never proceed with a
partially-passing tool.
The fixtures are not optional and they are not decoration. Without them this is worse than shipping a script, because a hand-written script is wrong the same way every time while a generated one is wrong differently on every machine, silently. The fixtures are also the regression suite for the spec: a reported bug becomes a fixture, and every future generated implementation inherits the fix.
Writing style
- Sentence case headings. No decorative emoji.
- No em dashes. Use a period or a comma.
- Say what a step does, not how it feels. If a sentence could appear unchanged in another repo's docs, it says nothing and should be cut.
- Active voice, and name the actor.
- Second person imperative for instructions. "Read the file", not "the file should be read".
Before you open a PR
- Read the skill as if you use a different agent than the author. Every name you had to translate is a bug.
- Confirm the skill's own check is one you could actually run.
- Confirm no executable file entered the repo.
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.

