mill
AGENTS.md1 starsChanged 2 months agoApache-2.0View on GitHub
Read CLAUDE.md first. It remains the shared project contract. Then read every unconditional rule below before working: Before editing or reviewing relevant files, also read these scoped rules: Inspect .claude/rules/ frontmatter for additions: no paths: means unconditional; otherwise load when its existing path patterns match the work. Read only relevant docs/SPEC.md sections, active goals/briefs, role bodies and skills. Do not preload the full SPEC, backlog or historical transcripts. After compaction, re-read the active goal and written contract before continuing. .agents/skills/…
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 repository10 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- userdocs/llms-full.txtllms.txt
- userdocs/llms.txtllms.txt
- briefSkill
- build-briefSkill
- drive-installed-appSkill
- manual-checksSkill
- run-millSkill
- mill-devSkill
- mill-useSkill
The file
We split this AGENTS.md into the rules it gives an agent
10 rules. About 0 are common, 6 are distinctive, and 4 are 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.
4
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 3
Read CLAUDE.md first. It remains the shared project contract.
- Distinctive: line 13
Before editing or reviewing relevant files, also read these scoped rules:
- Very specific: line 23
Read only relevant docs/SPEC.md sections, active goals/briefs, role bodies and skills.
- Distinctive: line 25
After compaction, re-read the active goal and written contract before continuing.
- Distinctive: line 43
Leave the orchestrator model unchanged.
- Very specific: line 45
When agent_type or named roles are exposed, use them and their runtime model assignments.
- Distinctive: line 50
Create each builder's git worktree explicitly.
- Very specific: line 57
Use current CLAUDE layout (internal/services/<ctx>svc) when the explorer's older root-service map differs.
- Distinctive: line 60
Do not fabricate Claude/Fable coauthors or Claude session URLs.
- Very specific: line 67
Read .codex/README.md for native hook activation and lifecycle translations.
≈ “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
Mill agent context
Read CLAUDE.md first. It remains the shared project contract. Then read every unconditional rule below before working:
.claude/rules/adopt-converged-patterns.md.claude/rules/architecture.md.claude/rules/comments.md.claude/rules/delivery-discipline.md.claude/rules/testing.md.claude/rules/ux-writing.md
Before editing or reviewing relevant files, also read these scoped rules:
| File paths | Rule |
**/*.go | .claude/rules/backend.md |
frontend/src/**/*.tsx, frontend/src/**/*.ts | .claude/rules/frontend.md |
internal/domain/composition/** | .claude/rules/node-standard.md |
Inspect .claude/rules/ frontmatter for additions: no paths: means
unconditional; otherwise load when its existing path patterns match the work.
Read only relevant docs/SPEC.md sections, active goals/briefs, role bodies and
skills. Do not preload the full SPEC, backlog or historical transcripts.
After compaction, re-read the active goal and written contract before continuing.
Skills and roles
.agents/skills/ links to the canonical .claude/skills/ folders. Read the
applicable SKILL.md there (or through its canonical path) when using a skill.
Native role profiles in .codex/agents/ point to .claude/agents/<role>.md;
read that body before executing a role. Preserve its scope, never-list, docs
ownership/draft requirement, review format and owner PR verification duties.
For a primary orchestration task, also read .codex/CONDUCTOR.md.
Codex translations
These translations replace Claude-specific transport, model and attribution examples, not the shared working discipline:
- Native role model/effort defaults live in
.codex/agents/<role>.toml. They are operational assignments, not claims of Claude cost/quality equivalence. Leave the orchestrator model unchanged. State the actual model on dispatch; use a permitted runtime choice and disclose unavailable configured models. - When
agent_typeor named roles are exposed, use them and their runtime model assignments; otherwise pass the canonical role path, brief and boundaries explicitly. Neither mode implies a tool allowlist or enforced sandbox; inspect the active runtime and report unavailable capabilities. - A reviewer starts fresh with
fork_turns=none; never same-context review. Create each builder's git worktree explicitly. At most three concurrent children; honor any lower runtime limit and CLAUDE's heavy-gate lock. - Track canonical
maxTurnsand brief token ceilings when metrics exist; disclose unavailable metrics, never invent usage or budget compliance. - Continue foreground commands using
exec_command'ssession_idandwrite_stdin, with bounded waits. Nonohupor detached gates. A tool wait returning a running session is foreground continuation, not abandoned work. - Use current CLAUDE layout (
internal/services/<ctx>svc) when the explorer's older root-service map differs. The command guard's rebase prohibition wins over older agent README guidance: merge or fix forward, never rewrite history. - Do not fabricate Claude/Fable coauthors or Claude session URLs. Identify the actual tool/model only when attribution is needed and known.
- Native UI work uses available, permitted computer-use tools. Report permission denials; never route a denied action through a different control channel.
- Builders own bounded CI waits, fixes and merge-queue enqueue through verified
MERGED; armed/open is not delivered. Verify actual queue state; never change branch protections. The orchestrator independently verifies and owns docs/closeout. - Read
.codex/README.mdfor native hook activation and lifecycle translations. At explicit task completion, run the canonical task-completion guard with the outcome subject and cwd. Perform canonical worktree cleanup only after removal of an owned worktree. Treat runtime denial feedback as authoritative.
Native config is prepared for a fresh session and user hook trust review; its presence does not prove it is active in the current session.
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.

