agentleFS
arxdsilvaarxdsilva/ai-loops/skills/parallel-work/SKILL.md

parallel-work

Skill0 starsChanged 7 days agoMITView on GitHub

Use when the lead plans or runs a loops wave of role agents, or when asked to work in parallel - contract first, disjoint ownership, spawn and label lanes, relay, gates, lifecycle and model tiers.

8 other projects use a version of this.

Adopted by

Most recently taken up by

The file

We split this skill into the rules it gives an agent

1 rule. 0 are common, 0 are distinctive, and 1 is very specific to this project.

  • 0

    Common

    Most of the 7 edited versions we read keep these.

  • 0

    Distinctive

    This author's own choices. Worth a look.

  • 1

    Very specific

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

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

    Integrate and verify: After each wave run loops gate, which checks every detected pack (specs/WORKFLOW.md §6).

    line 29

“Common” is measured: kept by at least 6 in 10 of the 7 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

Parallel work (lead checklist)

Rules: specs/PARALLEL_WORK.md. Run phases: specs/WORKFLOW.md §2-4. Roles and tiers: ROLES.md. This skill is the checklist; a one-file fix is still a wave with one agent.

  1. Plan has its waves. The plan (templates/PLAN.md) carries the contract, a waves table per wave and a block per agent (specs/PARALLEL_WORK.md §1). plan-reviewer rejects a plan without them.
  2. Map areas to roles. One role agent per layer or package area (ROLES.md). One role may run as several agents when their globs are disjoint (W1-domain-orders, W1-domain-payments), each with its own qa-verifier. Size waves to the effective cap = min(parallel.max_agents, default 100; loops hw recommendation; opt-in limits.max_agents); split larger waves into sub-waves (W1a, W1b) (specs/WORKFLOW.md §2.2). Every lane carries a language (a detected pack); test parallelism per agent comes from the pack (env_hints, test_jobs; Go: GOFLAGS=-p=2). Check the plan with loops plan check.
  3. Pick each model (ROLES.md tiers): haiku for mechanical work, sonnet for implementation against a contract, opus for contracts, consistency, concurrency and migrations. State the reason in the table. Escalate one tier after two failed QA rounds.
  4. Contract first (wave 0). The architect writes ports, typed commands, DTOs, stubs and shared config; one <id>-engineer per pack a lane uses sets up the toolchain and test prerequisites (loops:lang-<id>). Manifests and lockfiles have one owner. Wave 0 is done when each pack's build and lint checks pass (loops gate --only <build,lint checks>; Go: build,vet).
  5. Launch a wave in one message, in the background: producers (domain, application) first; consumers (adapters) may start against stubs. Use subagent_type: loops:<role>, with the agent's plan block and these prompt lines (specs/WORKFLOW.md §3):
Run dir: <abs>   Specs: <abs: loops path specs>   Role: <role>   Agent id: <W1-domain>   Wave: <n>
Language: <id> (root <dir>)
  1. Label each lane right after spawn:
loops run label --agent <agentId from the Agent result> --name W1-domain --role domain-engineer --wave 1
  1. Pipeline QA. Launch a qa-verifier as soon as its implementer reports. Record verdicts and loop-backs with loops run verdict and loops run loopback.
  2. Relay cross-agent facts through the lead with SendMessage; broadcast additive contract changes to every affected agent. Do not enter plan mode while agents run.
  3. Integrate and verify. After each wave run loops gate, which checks every detected pack (specs/WORKFLOW.md §6); route failures to the owning agent, resumed with SendMessage.
  4. Lifecycle (specs/PARALLEL_WORK.md §7). Reuse the owning agent for fixes and the next wave. Do one-fact lookups and edits under ~30 lines yourself. Commit with git commit -m ... -- <paths> (only product-reviewer commits, after you confirm). TaskStop agents you will not resume. Split a worker before about 200 API calls.

Agent prompt skeleton

Model: <haiku|sonnet|opus> - <one-line reason>
You are agent <id> (<role>, language <id>) on <repo>. Load loops:lang-<id>. Read first: <plan file §>, AGENTS.md, nearest package AGENTS.md, <spec sections>.
Contract: <the plan's contract section>
Your job: <the agent's block from the plan>
Rules: owns ONLY <globs>; must NOT touch <globs>, manifests, lockfiles, shared config; no git writes; strict TDD;
  targeted checks only (loops gate --only <check>,... --pkgs <owned>; the pack's race policy).
Final report: templates/CLAIMS_REPORT.md (<=400 words).

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.