mcp-adr-analysis-server
AGENTS.md35 starsChanged 20 days agoMITView on GitHub
Instructions for coding agents that build changes in this repository (OpenCode, Claude Code in a builder loop, or any harness that reads AGENTS.md). Project commands, architecture and conventions are in CLAUDE.md; read it first. This repo follows the Invoice Exception Desk PLAYBOOK: a human rules on every change in a GitHub issue, a builder implements exactly that ruling, checks the builder never saw grade it, and a human merges. Only when you are working an issue labelled agent-ready (an "Agent…
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 repository13 other files it gives its agents.
- .agents/skills/document-specialist/AGENTS.mdAGENTS.md
- .github/copilot-instructions.mdCopilot instructions
- documentation-specialistSkill
- speckit-analyzeSkill
- speckit-checklistSkill
- speckit-clarifySkill
- speckit-constitutionSkill
- speckit-convergeSkill
- speckit-implementSkill
- speckit-planSkill
- speckit-specifySkill
- speckit-tasksSkill
- speckit-taskstoissuesSkill
The file
We split this AGENTS.md into the rules it gives an agent
7 rules. About 1 is common, 1 is distinctive, and 5 are very specific to this project.
≈1
Common · estimate
Many files like this have a rule of this kind.
≈1
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
- could be automatic: a setting could enforce it
- Very specific: line 13
Only when you are working an issue labelled agent-ready (an "Agent ruling" issue).
- Very specific: line 31
Run the tests: npm run typecheck && npm test.
- Very specific: line 37
Never run /speckit.specify, /speckit.clarify, /speckit.plan or /speckit.tasks during a builder loop.
- Very specific: line 40
If a task is unclear, conflicts with the spec, plan or ruling, or needs something that isn't done: don't guess and don't skip ahead.
- Very specific: line 45
Don't report sizes or counts.
- Distinctive: line 49
Stop rather than edit any of these, in addition to anything the ruling lists:
- Common: line 59
Never push, open PRs, or use credentials.
≈ “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
AGENTS.md
Instructions for coding agents that build changes in this repository (OpenCode, Claude Code
in a builder loop, or any harness that reads AGENTS.md). Project commands, architecture and
conventions are in CLAUDE.md; read it first.
This repo follows the Invoice Exception Desk PLAYBOOK: a human rules on every change in a GitHub issue, a builder implements exactly that ruling, checks the builder never saw grade it, and a human merges.
When these rules apply
Only when you are working an issue labelled agent-ready (an "Agent ruling" issue). That
issue is the ruling: it answers every question you would otherwise guess. Without the label,
don't start.
Which row gets built next is the operator's pick from ROADMAP.md ("How the
next row is picked"), never the builder's.
Builder rules
Inputs: .specify/memory/constitution.md, the ruling issue, and specs/<NNN-feature>/spec.md,
plan.md and tasks.md. A human approved them. Read them, every ADR listed in the spec's
Governing Decisions section (in docs/adrs/), and ralph-memory.md if it exists, before you
change anything. If a task would contradict an Accepted ADR, that's a BLOCKED stop (rule 6).
- Work only in the feature folder you were given.
- Take the first user story (phase) in
tasks.mdthat still has unchecked tasks. Implement its tasks in order, dependencies first, non-[P]tasks before[P]ones. Work on that one story only. - Run the tests:
npm run typecheck && npm test. If they fail, fix them within this story, at most 3 tries. Tick only the tasks whose tests pass; if none pass, tick nothing, commit nothing, end your turn. - Tick the completed tasks
[x]intasks.md, updateralph-memory.mdandprogress.md, and make one commit with the change,tasks.md,ralph-memory.mdandprogress.md. Then end your turn. Don't start the next story. - Never run
/speckit.specify,/speckit.clarify,/speckit.planor/speckit.tasksduring a builder loop. Never editspec.md,plan.mdor the constitution. Don't add, reword or delete tasks. - If a task is unclear, conflicts with the spec, plan or ruling, or needs something that isn't
done: don't guess and don't skip ahead. Append
BLOCKED: <task ID>: <question>toprogress.md, commit nothing, and end your turn. - Print
<promise>COMPLETE</promise>on its own line only when every task is[x], the tests pass andgit status --short --untracked-files=allis empty. - Don't report sizes or counts. CI (
Acceptance Grader) does that.
Never touch
Stop rather than edit any of these, in addition to anything the ruling lists:
.github/,.specify/,.opencode/,.claude/,acceptance/AGENTS.md,CLAUDE.md,ROADMAP.md,.repo-governor*docs/adrs/: decisions are the owner's; a builder reads ADRs and never writes or edits thempackage.json,package-lock.json,server.json,CHANGELOG.md- the grader's own files:
vitest.config.ts,vitest.acceptance.config.ts,scripts/check-base-test-ids.mjs - existing test files, except the exact edits the ruling allows
Never push, open PRs, or use credentials. A human or the operator does that.
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.

