coordinator-method
Skill0 starsChanged 36 days agoMITView on GitHub
The person's coordinator method: delegating to cloud and local workers, merging PRs, checks every round, usage stops, overnight loops, deciding instead of asking, UI changes. Use when acting as coordinator in any project: launching or briefing a worker, merging or closing a PR, starting a round, the person going to sleep, or changing an interface.
In this repository17 other files it gives its agents.
- clean-codeSkill
- cloud-workerSkill
- context-hygieneSkill
- git-conventionsSkill
- matt-bridgeSkill
- merge-gateSkill
- model-mixSkill
- mod-uiSkill
- overnightSkill
- release-watchSkill
- roadmap-trackerSkill
- screen-mapSkill
- skill-auditSkill
- smart-modsSkill
- smart-ultracodeSkill
- verified-researchSkill
- windows-opsSkill
The file
We split this skill into the rules it gives an agent
10 rules. About 0 are common, 8 are distinctive, and 2 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈8
Distinctive · estimate
This author's own choices. Worth a look.
2
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 16
When a feature maps to a named upstream skill or method, load its text verbatim with a thin binding to local tools.
- Distinctive: line 24
Before merging: merge the latest main into the branch, rerun the required checks on that result, and merge only if main has not moved since.
- Distinctive: line 26
After a merge: close the tickets it completes with a comment naming the PR, update the roadmap, launch what it unblocks, close obsolete or …
- Distinctive: line 27
Ask before merging anything unfinished or failing.
- Distinctive: line 32
After each merge, read main's CI and the failing test itself.
- Distinctive: line 39
When the person goes to sleep:
- Very specific: line 43
Keep the machine awake: ask the person to turn on Keep computer awake in Desktop (Settings → This computer → System) and leave the lid …
- Distinctive: line 56
Every visual choice states its meaning from the product's core metaphor in the project's visual-language doc, and one signal keeps one …
- Very specific: line 62
Start from the latest main and merge it again before every push: git fetch origin && git merge -m "chore: merge origin/main into <branch>" …
- Distinctive: line 63
Before every push run the full type check, test suite, build and UI check on the merged result.
≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill 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
How the person wants delegated work run, in every project. The project's own AGENTS.md, CONTRIBUTING.md, branch protection and .claude/CLAUDE.md win over this skill. Project facts (scripts, paths, roadmap issue, accounts, test app) live in the project's memory and its optional .claude/CLAUDE.md. Model and effort for every worker come from model-mix; whether and how to run a workflow comes from smart-ultracode.
Delegation
- The coordinator keeps the queue, briefs, reviews, merges and roadmap. Hands-on work of any size (debugging, conflicts, tests, docs, settings) becomes a delegated thread: gather only the facts a precise brief needs, then hand off.
- Cloud sessions take everything that runs in a checkout: code, tests, conflicts, docs, investigations. From the repo root:
expect ~/.claude/skills/coordinator-method/launch.exp <task> <rules|-> <log> <model> <effort>. The prompt is the task, the Worker brief at the end of this file (keep it the last section), then the project's rules file. Confirm the session URL before telling the person work runs in the cloud; an agent tool's remote option can fall back to local silently. - A local background agent in its own worktree takes only work that needs this machine: installed CLIs, the person's signed-in accounts, the local packaged build. One at a time.
- The local machine stays light: full suites, builds and UI checks run in cloud sessions and CI.
- One implementation thread at a time by default; run parallel threads, on disjoint areas, when the person asks or when the budget is green and today's profile in
model-mixallows more than one cloud session. Cloud sessions and routines draw from the same plan limits as everything else. Check rarely: one long wait beats many status reads. Workers commit often. - Real proofs against an AI provider pin the cheapest suitable model and the lowest effort in every request and ticket prompt. The person's own tool configs stay as they are unless they ask.
- When a feature maps to a named upstream skill or method, load its text verbatim with a thin binding to local tools; flag anything without an upstream source as a local addition.
- QA mode (the person asks for a critic, tester, evaluator): write user stories on the spot and fix flows through at most three local subagents in worktrees with one shared written brief. Each fix starts from a failing test; run only targeted tests, one worker, under
nice. Worktrees stay unpushed: review each branch and bring it into the working PR yourself. - The roadmap lives in the tracker (pinned issue plus milestones). After a restart read it and the open PRs, then continue the queue. Every new issue gets a milestone.
- A declared reliability cycle: no new features, one phase at a time in order. A phase closes when its PR is merged green with proofs run by the project's own tooling, never an agent's claim; then recap (changes, proofs, quota) and wait for the person's ok. A new security problem goes first. Refactors: plan, wait for the ok, one PR per responsibility, no behavior change.
Merging and git safety
- Merge finished, mergeable work yourself, without asking. The person never presses merge or turns on auto-merge, and neither do you: platform auto-merge squashes or writes a non-conforming subject.
- Before merging: merge the latest main into the branch, rerun the required checks on that result, and merge only if main has not moved since; otherwise realign again. Merge only on green required checks, never by override.
- For an approved PR, arm a monitor on its CI and run
gh pr merge N --merge --match-head-commit <sha> --subject "<type>(scope): ... (#N)"when it turns green. - After a merge: close the tickets it completes with a comment naming the PR, update the roadmap, launch what it unblocks, close obsolete or superseded PRs, rebuild and restart the local test build on the new main (local threads driving it reconnect).
- Ask before merging anything unfinished or failing. Dependabot minor and patch merge on green CI; majors wait for a dedicated thread.
- Naming of commits, branches, PR titles and merge subjects, and the history and tag rules (append-only pushed history, no force push, main only through PRs, immutable
v*tags):git-conventions. The sync subject here ischore: merge origin/main into <branch>, a valid Conventional Commit.
Every round
- After each merge, read main's CI and the failing test itself. Tell a flaky timeout from a deterministic failure and fix or hand off the second at once.
- List open PRs and new issues. For each idle worker, look in the tracker for its PR, branch or comments so finished work gets picked up.
- Compare each new PR with what just landed on main to catch duplicates. Review the diff (final review row of
model-mix) and the screenshots of every UI PR, including model-facing text leaking into the interface.
Usage and overnight
- Run the budget check in
model-mixat the moments it names, and follow its color: yellow is one profile down, red starts nothing new but lets open PRs finish. Usage is per account, so other sessions, cloud workers and the person's chats spend the same budget. - When the person goes to sleep:
- Compact first, or hand off to a fresh coordinator, because every wakeup sends the whole context again.
- Wake on events: a background watcher script or a CI monitor. Add a self-paced
/loopwith a 30 minute fallback wakeup, which keeps the 1-hour prompt cache warm. - Each wakeup runs the budget check before launching. With the 5-hour window at 90% or more (Pro and Max), sleep until its reset. Ultra has no such window; only the weekly and Fable weekly limits apply.
- Keep the machine awake: ask the person to turn on Keep computer awake in Desktop (Settings → This computer → System) and leave the lid open, or use
caffeinate -is -t 36000on macOS. - Push notifications only for decisions and milestones. Leave a morning summary, with the weekly points spent overnight, in the project's scratch state file.
Talking with the person
- Decide anything derivable from project rules, agreed decisions or earlier messages. Arbitrate disagreements between agents or reviewers, record the choice with its doubt, and break loops. Finish mechanical steps yourself, such as committing a resolved merge. Ask only product or design choices nothing covers.
- In grilling sessions map the whole frontier, then ask one question per message with options and a recommendation, and say how many remain.
- Public content for the person's accounts: show each draft (text, translation, visual) and publish only after their ok on that item, within their daily cap. Humanizer on every text, only merged facts, visuals from the real product with no private data, the person's own browser session. Videos reuse real components and tokens, few soft and motivated camera and mouse moves, short caption cards on why the product is useful, synthesized audio. Scheduled routines only prepare drafts.
UI changes
- Show first: edit the working tree, run the app from it, send a screenshot, wait for the ok, then commit and push.
- Mockups and prototypes decide layout only; every real component and feature stays. Each UI PR carries a before/after table of touched components, a feature map is the contract each change ticks, and CI blocks removed tests or screenshots.
- Every visual choice states its meaning from the product's core metaphor in the project's visual-language doc, and one signal keeps one meaning everywhere. Motion: about 150 ms for state changes, slow and continuous only for ongoing work, off under reduced motion.
- Calls to action sit on the right of an action row, primary or destructive last, through one shared row component.
Worker brief
Rules for every worker. The project's rules that follow win on conflict.
- Start from the latest main and merge it again before every push:
git fetch origin && git merge -m "chore: merge origin/main into <branch>" origin/main. - Before every push run the full type check, test suite, build and UI check on the merged result. If the environment cannot run one, say so in the PR and the report; report only checks that ran and passed as passed.
- Commits, merge commits and PR titles follow Conventional Commits; branches follow Conventional Branch with long types, lowercase, hyphens and the issue number, never an agent prefix. Each commit ends with the Co-Authored-By line your session's attribution gives for your model.
- Pushed history is append-only: a fix after a push is a new commit, no amend plus push, no
--forceor--force-with-lease. - Open or update the PR, then stop: the coordinator merges. Start no other ticket.
- Conflicts keep both sides' behavior. Existing tests and checks stay as strong as they are; list each resolved conflict in the PR.
- Code you add or change follows
clean-codewhen that skill is available; at least: clear names, no hidden side effects, no logic copied from elsewhere in the repo. Name in the PR anyclean-coderule left unapplied and any problem seen in existing code you did not change. - Tests use one clock: fake Date and move the system time with the test clock, or pass
nowexplicitly. - Mockups decide layout only: keep every real component and list the touched ones in the PR.
- Text the person sees carries no model-facing framing, tool or API names, or prompt instructions.
- Code reused from another repository is credited only in the third-party notices file.
- Finish with a short report: what changed, PR URL, check results, and the origin/main SHA tested.
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.

