qbpowers
AGENTS.md3 starsChanged 4 days agoUnlicenseView on GitHub
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.
The file
Show its rules and text
We split this AGENTS.md into the rules it gives an agent
20 rules. About 0 are common, 15 are distinctive, and 5 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈15
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 2
Run all Python, including one-off scripts, with Python 3.14 from the .venv at the repository root.
- Very specific: line 6
Put temporary, experimental and intermediate files whose location the user did not specify in .temp_files/YYYYMMDD_HHMMSS_<purpose>/ (for …
- Distinctive: line 8
Before deleting files or stopping processes or databases, check that they belong to the current work.
- Distinctive: line 11
Commit only when the user asks.
- Distinctive: line 14
If a request seems mistaken or a simpler approach exists, say so briefly, with the reason.
- Distinctive: line 15
Before answering about a symptom the user describes, check the relevant facts with tools.
- Distinctive: line 18
If a design defect within the task's scope would force a workaround, fix the defect instead, sized to its real impact.
- Distinctive: line 19
Keep designs as simple as the requirement allows.
- Distinctive: line 23
Write comments, in English, only for reasons, constraints or caveats the code cannot show.
- Distinctive: line 28
Run tests headless when possible (for example, Playwright headless).
- Distinctive: line 30
Read a worker's report that checks passed as covering only the checks that worker ran, and rerun the checks after merging the work.
- Distinctive: line 33
Use mypy in strict mode as the only type checker, and do not loosen its settings.
- Very specific: line 34
Prefer precise types (domain types, Pydantic models, TypedDict, dataclasses, generics, Protocol) over bare dict, dict[str, Any] or Any.
- Distinctive: line 35
Use Pydantic at external or untrusted boundaries: API payloads, LLM outputs, tool arguments, config, webhooks, database DTOs, and queue or …
- Distinctive: line 37
Use Ruff for Python formatting, linting and import sorting.
- Distinctive: line 40
Look for evidence in the repository first: code, configuration and project documentation.
- Very specific: line 41
If the built-in web tools cannot get a page (bot block or missing content), download it with curl -L -o and save the result under …
- Distinctive: line 42
Check dates and versions for time-sensitive or version-dependent facts.
- Distinctive: line 45
Keep in documents only what agents need to consult or work from.
- Very specific: line 47
When a change alters what a source-of-truth document describes, update that document in the same change.
≈ “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
Python Environment
- Run all Python, including one-off scripts, with Python 3.14 from the
.venvat the repository root. Never call the OS-globalpython3orpip; install packages withpython -m pip. - Manage dependencies in the root
pyproject.toml, never withrequirements.txt.
Files and Processes
- Put temporary, experimental and intermediate files whose location the user did not specify in
.temp_files/YYYYMMDD_HHMMSS_<purpose>/(for example,.temp_files/20260327_232323_translate_test/). - Edit files tracked by git only with the file-editing tool (Claude Code:
Edit/Write; Codex:apply_patch), not withsed -i, scripted replacement or shell redirection. Output from formatters and code generators is exempt. - Before deleting files or stopping processes or databases, check that they belong to the current work. Do not judge by names, prefixes or old PIDs alone; if another session or the user might be using them, ask first.
Git
- Commit only when the user asks.
Interpreting and Answering User Requests
- If a request seems mistaken or a simpler approach exists, say so briefly, with the reason.
- Before answering about a symptom the user describes, check the relevant facts with tools.
Code Changes
- If a design defect within the task's scope would force a workaround, fix the defect instead, sized to its real impact.
- Keep designs as simple as the requirement allows; do not write 200 lines where 50 will do.
- Do not add handling for cases that are very unlikely to happen, or just-in-case options, abstractions, fallbacks, exception handling or validation: over-engineering is itself a source of bugs. Validation at boundaries (user input, external APIs, file and queue messages) and handling of external failures that really occur are still required.
- Do not add production code that behaves differently under test (for example, checks of a test environment variable, or bypasses taken only in tests). If code is hard to test, inject dependencies such as clocks, clients and paths through parameters or a
Protocol, defaulting to production behavior. - Name tests and identifiers by behavior, not by review rounds or task IDs.
- Write comments, in English, only for reasons, constraints or caveats the code cannot show. Do not restate the code or record change history or task IDs.
Verification and Reporting
- After edits, run the formatter and linter configured for each language you changed (Python: Ruff, mypy and the relevant pytest tests). If a check fails, fix the root cause instead of loosening or ignoring the check. If a check cannot run, report exactly what blocked it.
- For a bug fix, first add a test that reproduces the bug, where practical.
- Run tests headless when possible (for example, Playwright headless). Open OS windows only when the user allows it.
- Report each check as passed, partially checked, failed or not run. Do not present a test on mock data as a real-environment test.
- Read a worker's report that checks passed as covering only the checks that worker ran, and rerun the checks after merging the work.
Python Types
- Use mypy in strict mode as the only type checker, and do not loosen its settings. Changed code must pass mypy; new or changed functions need explicit parameter and return types.
- Prefer precise types (domain types, Pydantic models,
TypedDict, dataclasses, generics,Protocol) over baredict,dict[str, Any]orAny. Keep unavoidableAny,cast(),# type: ignoreor# noqanarrow, with a specific code and a reason. - Use Pydantic at external or untrusted boundaries: API payloads, LLM outputs, tool arguments, config, webhooks, database DTOs, and queue or file messages.
- Let local variable types be inferred. Annotate a local only when it clarifies intent or mypy needs it: empty collections,
Noneinitialization,Anyquarantine, complex unions, public attributes or ambiguous inference. - Use Ruff for Python formatting, linting and import sorting; do not add other tools for these unless committed configuration already requires them.
Research and External Sources
- Look for evidence in the repository first: code, configuration and project documentation. Then search the web.
- If the built-in web tools cannot get a page (bot block or missing content), download it with
curl -L -oand save the result under.temp_files/. - Check dates and versions for time-sensitive or version-dependent facts.
Documentation
- Keep in documents only what agents need to consult or work from; delete the rest. Accumulate journal- or log-style records only when strictly necessary; otherwise delete or tightly compress them.
- Status documents and every
AGENTS.md/CLAUDE.md(root and subprojects) hold only the current state and stable rules, never run logs or incident-specific rules. - When a change alters what a source-of-truth document describes, update that document in the same change. Link to a rule's single source instead of copying it. Put conclusions in the relevant document, not only in
.temp_files/.
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.

