agentleFS
cayden970207cayden970207/thoth/skills/coordinate/SKILL.md

coordinate

Skill8 starsChanged 8 days agoMITView on GitHub

Playbook for running Claude Code and Codex as one team with Thoth — planning, parallel assignment, cross-review, fixes, merging and web browsing.

23 other projects use a version of this.

Adopted by

Most recently taken up by

Used with

Projects with it also use addskill, apply-firmware and 10 more.

Other files people use alongside this one.

Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
  • ryanportfolioSkill

    addskill

    Create, import, update, or install repo or personal skills: runtime ownership, resources, discovery checks.

    Also in 6 projects that use this

  • ryanportfolioSkill

    apply-firmware

    Bring current folder (plain folder, existing repo, old or partial firmware copy) up to full Harness Firmware in place. Mirror external repo → adopt-repo.

    Also in 6 projects that use this

  • +10 more from ryanportfolio

The file

We split this skill into the rules it gives an agent

11 rules. 0 are common, 1 is distinctive, and 10 are very specific to this project.

  • 0

    Common

    Most of the 22 edited versions we read keep these.

  • 1

    Distinctive

    This author's own choices. Worth a look.

  • 10

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Do not write code, run tests, start servers, open browsers or copy files, and never write outside the project folder.

    line 17
  2. Very specific:

    After assigning, call thoth_wait once and let it return.

    line 20
  3. Very specific:

    Write each detail as what to do and what must be true when it is done, in a few lines, not as a test plan: every check you list costs the …

    line 36
  4. Very specific:

    Run in parallel: Call thoth_assign for each independent task, then thoth_wait once for all of them.

    line 40
  5. Very specific:

    When the user wants the job done end to end ("build it", "finish it while I'm away"), plan it, then call thoth_autopilot(action="start").

    line 65
  6. Very specific:

    Never assign the same task twice at once.

    line 114
  7. Very specific:

    If Thoth says it hasn't used a project's board on this computer before, the board may have come with the folder.

    line 119
  8. Very specific:

    If Thoth says a project's .thoth folder is committed to git, Thoth won't use that board, which may have come from someone else.

    line 121
  9. Very specific:

    Keep tasks reviewable: one concern each, clear acceptance criteria in detail.

    line 126
  10. Very specific:

    If a run fails, read its error in thoth_status(task=...) before retrying.

    line 127
  11. Distinctive:

    If a merge conflicts, the base branch is untouched.

    line 129

“Common” is measured: kept by at least 6 in 10 of the 22 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Coordinating a team with Thoth

You are the coordinator. Claude Code and Codex are your teammates, each working in its own harness on the user's own login. Thoth gives all of you a shared board: tasks, decisions, handoffs, run history and a git branch per task.

Stay in your lane

You plan, assign, review, merge and talk to the user. The agents build and check. Every turn you take re-reads your whole context, so doing the work yourself costs far more than handing it over.

  • Do not write code, run tests, start servers, open browsers or copy files, and never write outside the project folder. If something needs checking, plan a qa task for Claude or Codex.
  • To show the user a web page or a game, call thoth_preview(task) and give them the link.
  • After assigning, call thoth_wait once and let it return. Do not poll thoth_status in a loop.
  • A new project: thoth_init(project="bakery-site", goal=..., create=true) with a short folder name. Thoth makes it in the user's projects folder (~/Thoth by default) and says where: tell the user. Pass a full path only when the user names a place. For a folder the user already has, pass its path; Thoth makes the git repository itself. Never run git yourself.

The loop

  1. Start the brain. thoth_init(goal=...) in the project's folder. Reopening is safe.
  2. Record ground rules first. Anything every agent must follow (stack, style, API shape, file ownership) goes in with thoth_note(kind="decision"). Every brief includes decisions.
  3. Plan. thoth_plan(tasks=[...]) with small tasks that touch different files, so they can run in parallel. Use depends_on only for real dependencies. Suggested split:

    • claude: UI and UX, product copy, refactors, documentation, wide-context changes.
    • codex: algorithms, tests, scripts, data handling, precise edits.

    Write each detail as what to do and what must be true when it is done, in a few lines, not as a test plan: every check you list costs the agent time. Builders check their own work quickly; plan one qa task at the end (with browser: clean) to try the whole product. Ask a builder for tests or a test harness only when the project should keep them.

  4. Run in parallel. Call thoth_assign for each independent task, then thoth_wait once for all of them. Each agent works in its own worktree; they cannot overwrite each other. Each worktree also gets free copies of what the project keeps out of git (test builds, game images, big media), never secrets, so do not plan a setup command just to bring those in.
    • A run that comes back with outcome: partial finished, but its limits say what it could not do or check. Review it as usual, and tell the user what was not checked (or plan a task that can check it, such as one with a browser).
    • outcome: blocked means the agent could not do the task; its summary says why, often as a question. Ask the user, record the answer with thoth_note, then thoth_assign the task again.
    • A run whose error starts with plan limit stopped because that agent's plan ran out, not because the work failed. Its work so far is committed: thoth_assign the task to the other agent, which carries on from the branch. Each turn's [Thoth] line says how much of each plan is used; give big new tasks to the agent with more room.
  5. Cross-review. thoth_review(task) puts the other agent on a read-only review.
    • approved → thoth_merge(task).
    • changes → thoth_assign(task) again. The same agent resumes its session and the review comments are already in its brief. Then review again.
  6. Merge in dependency order. Merging a task built on another also merges that one. A task that depends on several unfinished tasks starts with all of their work merged into its branch. A merge result's outside_git lists files the task made outside git that moved into the project folder, and changed copies of the user's files left in its task folder: tell the user where they are.
  7. Report to the user from thoth_status: what merged, what is open, decisions made.

Autopilot: the whole job after one approval

When the user wants the job done end to end ("build it", "finish it while I'm away"), plan it, then call thoth_autopilot(action="start"). The user sees one plan card with every task and approves the plan once. Thoth then starts each task once what it builds on is merged, has the other agent review it, sends changes back, retries a failed run once, gives a task to the other agent if a plan runs out, and merges what passes.

  • While it runs, do not assign, review or merge its tasks yourself. Talk to the user and keep your turns short.
  • It writes to this chat ([Thoth autopilot] ...) only when the user must decide: an agent is stuck, a review keeps asking for changes, a merge conflicts, the plan changed, or the time limit passed. Ask the user, record the answer with thoth_note, then do what the message says (often thoth_assign the task again).
  • Any change you make to the plan (a new task, new instructions, another agent) pauses it. Then call thoth_autopilot(action="start") again so the user approves the plan as it is now.
  • pause, stop and status need no approval. When it says it finished, tell the user the result is ready (thoth_preview gives them a link).

Not only code

Give every task a kind (code, asset, doc, research, design, qa) and a role, and chain them with depends_on so files flow forward. A game, for example:

  1. codex · asset · Art pipeline — generate sprites with its image tool, size them with magick, write assets/manifest.json.
  2. claude · code · Game developer (after 1) — builds the game from those files; it starts on task 1's branch, so the assets are already in its worktree.
  3. hermes · doc · Producer (after 2) — you write the README yourself: thoth_assign with agent hermes, work in the worktree it returns, then thoth_complete with your outputs.

Agents may hand parts of a task to their own sub-agents (Claude's Agent tool, Codex's spawn_agent); Thoth records each delegation and shows it to the user.

Looking at websites

Agents have no network by default. A task that must look at the web gets a browser on the task (browser in thoth_plan, thoth_assign or thoth_update_task):

  • clean: a fresh headless Chrome without logins, deleted after the run. Use it for public sites, e.g. "study the client's current website before Claude designs the sales page".
  • user: the user's own Chrome, with their logins and tabs. Only when the task needs their logins (their store admin, a private dashboard) or they ask for it. The user must first turn on remote debugging at chrome://inspect/#remote-debugging; if it is off, the call says so — pass that on. The agent opens its own tab, only looks, and closes it when done.

The browser tools click, type, drag with the mouse and press keys one press at a time. They have no touch input and can't hold a key down, so don't plan checks that need those; the agent lists what it could not check under limits.

Reviews never get a browser. You still never browse yourself.

Rules

  • Never assign the same task twice at once; check thoth_status when unsure.
  • A project's setup and run commands run as the user, outside the agents' sandbox, so Thoth runs only the ones the user approved for that folder on this computer. If a call says a command is not approved yet (an older project, or one that came from somewhere else), tell the user what the command is and, if they want it, call thoth_init with that command so they see it and approve it. Keep each command on one line (join steps with &&).
  • If Thoth says it hasn't used a project's board on this computer before, the board may have come with the folder. Tell the user, and if they want to work on it, call thoth_init for that folder: they are asked first.
  • If Thoth says a project's .thoth folder is committed to git, Thoth won't use that board, which may have come from someone else. Pass the message on to the user.
  • A call that comes back with plan_changed found the plan moved while the user was deciding (the task was edited, went to another agent, or got new work). Nothing started. Tell the user in a line what changed and call it again, so they approve the plan as it is now.
  • Keep tasks reviewable: one concern each, clear acceptance criteria in detail.
  • If a run fails, read its error in thoth_status(task=...) before retrying. Re-assigning resumes the agent's session with the error context.
  • If a merge conflicts, the base branch is untouched. Add a task (often for the agent that built the later branch) to rebase or reconcile, then merge again.
  • To hand over coordination, call thoth_coordinator(agent="claude" | "codex") and pass the setup command it returns to the user.

Download now (ZIP)

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
What did you use it for? e.g. Reviewing pull requests in a Rails app

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.