agentleFS
nickthelegendnickthelegend/nick-skills/skills/orchestrate/SKILL.md

orchestrate

Skill0 starsChanged 4 days agoMITView on GitHub

Run a whole hackathon entry end to end, on autopilot, without stopping. A task graph (orch.py) covers every phase (event rules, credential sweep, /grok-ideation, /idea-judge, spec, Codex build workers, /dribbble-frontend UI, /chatgpt-assets art, /gemini-video clip, live proof, /idea-judge on the build, /judge-skill gallery and walkthrough, the film, the submission pack), and a Stop hook refuses to let the session end its turn while any task is runnable. It stops only for steps only the human can do, and keeps working on everything else meanwhile. Use it when someone says "do it all", "end to end", "run the pipeline", names a hackathon and wants entries, or types /orchestrate.

  • Reads credentials

169 other projects use a version of this.

Adopted by

Most recently taken up by

Used with

Projects with it also use define-requirements, orchestrate-plan 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.
  • sson0-erSkill

    define-requirements

    Work with the user, one question at a time, to complete the requirements document (docs/claudex/<slug>/01-requirements.md). Run by the orchestrator (parent). Stays out of design and implementation.

    Also in 3 projects that use this

  • EternallLightSkill

    orchestrate-plan

    Use when executing a locked, reviewed PLAN.md by delegating the implementation to visible worker agents in Orca, herdr, or cmux ONE PHASE at a time — splitting the plan into context-window-sized phases, giving every phase a pristine worker context, and (when the plan touches INDEPENDENT repos) running one worker per repo IN PARALLEL. Triggers on "run the plan with a worker", "execute the plan in the other pane", "split the plan into phases and build it with workers", "build the backend and frontend in parallel panes". Not for edits you can finish inline; outside Orca, herdr, or cmux use tgc-skills:orchestrate.

    Also in 3 projects that use this

  • sson0-erSkill

    record-evidence

    How to record external facts (docs, command output, codebase findings) as evidence and design choices as decisions. Used by the researcher, designer and reviewer agents.

    Also in 3 projects that use this

  • PinkneyJiangSkill

    improve-architecture

    扫描代码库中的架构摩擦点,识别浅模块并挖掘深模块演进机会。生成可视化交互式 HTML 诊断报告,并在用户确认后闭环实施深度重构。

    Also in 4 projects that use this

  • kjnezSkill

    ticket

    Work on a JIRA/Linear ticket end-to-end: read ticket details, explore codebase, create a branch, implement with TDD, run quality checks, update ticket status, and create a PR. Use when the user provides a ticket ID to implement. Triggers on requests like "/ticket PROJ-123", "work on this ticket", "implement PROJ-123".

    Also in 4 projects that use this

  • xfstudioSkill

    reverse-engineer

    Expert reverse engineer specializing in binary analysis,

    Also in 3 projects that use this

The file

We split this skill into the rules it gives an agent

8 rules. 0 are common, 5 are distinctive, and 3 are very specific to this project.

  • 0

    Common

    Most of the 151 edited versions we read keep these.

  • 5

    Distinctive

    This author's own choices. Worth a look.

  • 3

    Very specific

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

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    You: do design, film, copy and review.

    line 10
  2. Very specific:

    Run scripts/install-hooks.sh once per machine.

    line 28
  3. Very specific:

    DONE:: run orch.py finish and send the deliverables.

    line 52
  4. Very specific:

    Add any extra work with orch.py add --product <p> --phase <n> --title "..." --after <id>.

    line 88
  5. Distinctive:

    No re-skins of past builds.

    line 93
  6. Distinctive:

    No mocks of external APIs.

    line 97
  7. Distinctive:

    Never print or open secrets.

    line 99
  8. Distinctive:

    Never delete a submitted video.

    line 101

“Common” is measured: kept by at least 6 in 10 of the 151 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

/orchestrate

You are the conductor of a run that does not stop. The work is split up:

  • Codex writes code.
  • You do design, film, copy and review.
  • Grok scouts ideas.
  • ChatGPT and Gemini make art.

The run is a task graph in orch.py. A Stop hook holds you to it.

Why it's built this way: the old version ended its turn at every milestone ("Say go and I'll start…", a status report, "want me to…?"). It also launched Codex detached, so nothing woke it when the workers finished. The human came back hours later to "why u stopped". Both holes are closed now:

  • The Stop guard blocks the end of the turn while a task is runnable, and tells you the next task.
  • Workers run under Bash run_in_background: true, so the harness wakes you when each one exits. The guard then makes you review the result.
O="python3 $HOME/.claude/skills/orchestrate/scripts/orch.py"   # (zsh: use a function, not "$O")

Step 0: arm it (first time on a machine, then every run)

  1. Run scripts/install-hooks.sh once per machine. It is idempotent, and it's a no-op for sessions without a run.
  2. New run:
    orch.py init --work <grokbot>/runs/<event-slug> --event "<name>" --url <event page> --deadline <UTC> [--products a,b,c]
    • Name the products if the human already chose them; ideation is then skipped.
    • Otherwise ideation runs first and you add products after the pick with orch.py add-product a,b.
    • Continuing a run in a new session: orch.py resume --work <dir>, then read <dir>/STATE.md.
  3. init and resume bind this session to the run. From then on, ending the turn while tasks are runnable gets blocked, with the next task in the message. That's intended. Keep going.

The loop (this is the whole skill)

while true:
  orch.py next --all                 # NEXT task, its skill, its done-criteria; plus what's parallel
  orch.py start <id>
  do it (with the named skill; parallel independent tasks via subagents or Codex workers)
  orch.py done <id> --evidence "<commit / URL / path / test line>"
     or orch.py block <id> --needs "<exact credential or decision>" --who human   -> PushNotification it
     or orch.py skip <id> --why "<reason>"
  • Never end the turn while next prints NEXT: or REVIEW:.
  • WAIT: (only background workers are running): you may end the turn. Their exit wakes you.
  • HUMAN: (only human steps are left): make sure every ask was sent, report in five lines, and end.
  • DONE:: run orch.py finish and send the deliverables.
  • Evidence is required. "done" without a commit, URL, path or command output isn't done.
  • orch.py status re-renders STATE.md. Show it to the human at phase changes, not every task.

Autopilot rules (default mode)

  1. No permission questions. No "should I continue", "say go", "want me to", or "which one?" when you have a recommendation. Pick it, orch.py decide "<what and why>", tell the human in one line, keep going. They redirect you if they disagree.
  2. A missing key blocks one task, never the run.
    • block it, PushNotification the exact ask (what, where to get it, which .env variable), and move on.
    • Soft dependencies (~id) route around a blocked task. Film on a local fork or testnet when mainnet keys are missing, label it honestly, and re-shoot when the key lands.
  3. The only real gates are steps only the human can do:
    • real money or a mainnet spend (ask once, with an amount);
    • making a repo public, if it wasn't pre-approved;
    • submitting the form, signing in, or posting. Everything else you do yourself.
  4. Parallel by default. next --all lists every runnable task.
    • Independent tasks go to subagents (research, QA) or Codex workers (code).
    • At most 2–3 heavy workers on a 16 GB laptop.
    • Design, art direction, film and review stay with you.
  5. The human said stop → orch.py pause (the guard goes off). "Carry on" → orch.py unpause.
  6. Stalls: the guard lets you stop after 3 turn-ends in a row that didn't change the task graph. If you are truly stuck, block the task with what's needed rather than looping.

What each phase means (the graph already holds these tasks)

PhaseTasksUse
0 intakerules, credsRead the event page live → RULES.md (deadline UTC + IST, rubric weights, must-haves, entries/team, prizes copied, bans). Credential sweep → CREDENTIALS.md + one PushNotification checklist in the first hour. Losing streaks came from "BLOCKED: needs a key" found on deadline day.
1–2 ideasideate, pick/grok-ideation, then /idea-judge (you judge, never Grok). Autopilot takes the top pick. If the human rejects the batch, change the lens, not the wording.
3 buildspec, repo, brief, code, reviewWrite specs/<slug>.md; set up a monorepo with MIT (Copyright (c) 2026 Nivesh Gajengi) and a private remote; write briefs from templates/codex-brief.md; run ORCH_TASK=<id> scripts/codex-run.sh <repo> <brief> with run_in_background: true. Review before every commit: deleted or weakened tests, gutted docs, staged secrets, fake data. Commit each diff and push straight away.
4 lookui, art, clip, qa/dribbble-frontend (you, never Codex): layout from the real category leader's screens, finish from the Dribbble library, never an invented layout or a clone of another of our apps. Then /chatgpt-assets, /gemini-video. QA at 375 and 1440 px. Live data, empty states, or "Setup required" only.
5 livelive, testingReal transactions with explorer links in the README; TESTING.md from real runs. Mainnet spend is a block for the human's yes.
6 judge + filmjudge, gallery, walk, film/idea-judge on the running build, then fix the weakest criterion. /judge-skill gallery and raw walkthrough. Cut a ~3-minute film with /hyperframes: problem first, then live demos only, full-width app screens, captions, speed badges. Never edit real output; redact with a visible box.
7 submitpack, public, submitHACKATHON_SUBMISSION_FORM.md, README, post text. Run a secret scan, then make the repo public. submit is the human's: send a 3-line checklist with links and the deadline in IST.

Add any extra work with orch.py add --product <p> --phase <n> --title "..." --after <id>. That covers judge fixes, a second chain check, or a re-shoot when a key lands.

Standing constraints (every product)

  • One chain per project.
  • No re-skins of past builds.
  • A product that won on one chain stays there.
  • Several products in one event must not look like one team made them.
  • Brandable names (Japanese, Greek or Korean).
  • No mocks of external APIs.
  • Fail loudly ("Setup required").
  • Never print or open secrets.
  • Never force-push.
  • Never delete a submitted video; upload a new version of the same file.

Delivery

  • Send each result the moment it lands with SendUserFile: a gallery link, a film, a live URL.
  • Reports are plain words: what's done, what's live (with links), what only the human can do, and by when.
  • Push back honestly. If an entry won't be live by the deadline, say so early and add the cut that makes it live.

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.