agentleFS
Sagargupta16Sagargupta16/agent-brain/skills/parallel/SKILL.md

parallel

Skill2 starsChanged 11 days agoMITView on GitHub

Fan-out recipe for when the user wants speed through agents. Partitions work by unit (page, route group, file lane, repo), gives each agent sole write ownership of disjoint files and a bounded prompt with a verify step, launches all lanes in one Agent tool call, then runs the gates once and reports per lane. Use when the user says "run parallel agents to do things fast", "one agent for each page, fix all pages in parallel", "run more agents, I don't see much progress", or "strictly use parallel agents". Never spawns agents for a single-file task, lets two agents own one file, re-runs gates per agent, or commits before integration.

9 other projects use a version of this.

Adopted by

Most recently taken up by

Used with

Projects with it also use commit.

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.
  • SentrySkill

    commit

    Create commit messages following Sentry conventions. Use when committing code changes, writing commit messages, or formatting git history. Follows conventional commits with Sentry-specific issue references.

    Also in 3 projects that use this

The file

We split this skill into the rules it gives an agent

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

  • 0

    Common

    Most of the 7 edited versions we read keep these.

  • 8

    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:

    When progress looks slow, the answer is more lanes, never smaller scope: "run parallel agents to do things fast", and later "run more …

    line 17
  2. Very specific:

    Check disjointness mechanically: git ls-files <glob> per lane, then sort | uniq -d across all lanes must print nothing.

    line 45
  3. Very specific:

    Commit only after step 5 passes and only if the same message asked.

    line 69
  4. Distinctive:

    Never spawn agents for a single-file or single-unit task.

    line 87
  5. Distinctive:

    Never let two lanes own the same file.

    line 89
  6. Distinctive:

    Never run type, lint, build, or tests per agent or per lane.

    line 91
  7. Distinctive:

    Never accept "done" from an agent without reading its diff.

    line 93
  8. Distinctive:

    Never commit, push, or open a PR before the integration pass, and never one PR per lane.

    line 95
  9. Very specific:

    Never let an agent commit, format unrelated files, or edit outside owns.

    line 97
  10. Distinctive:

    Never report a lane as verified when the check was the agent's own claim.

    line 98
  11. Distinctive:

    Never chain commands with a double ampersand, in the shell or in a lane prompt.

    line 99

“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

When progress looks slow, the answer is more lanes, never smaller scope: "run parallel agents to do things fast", and later "run more agents, I don't see much progress". This is the scouts -> one synthesizer -> per-file fixers -> verify shape as a recipe on the Agent tool.

Read first

  • Prior lessons in the brain: brain recall parallel, brain recall workflow (and read the diff yourself; a subagent summary is not the diff)
  • The repo's CLAUDE.md for the gate command (pnpm run check, uv run pytest) and any worktree rule (for example: worktrees at the repo root, never under a nested package such as frontend/, or typescript-eslint sees two tsconfig roots)

Procedure

  1. Qualify. Count the independent units. One file, or units that all edit the same file, means no agents: do it inline and say why in one line. Two or more disjoint units means fan out with as many lanes as units; the user asked for speed, not a cap.
  2. Partition by the natural unit and print the lane table BEFORE launching:
| lane | owns (write) | reads | verify | agent |
| pages/budget | frontend/src/pages/budget/** | components/ui/** | render /budget at 375px, no console errors | 1 |
| pages/goals | frontend/src/pages/goals/** | same | render /goals | 2 |

Check disjointness mechanically: git ls-files <glob> per lane, then sort | uniq -d across all lanes must print nothing. A shared file (util, type, barrel) gets its own lane that runs FIRST, or one lane owns it and the rest treat it read-only. 3. One prompt per lane, same skeleton for all, the user's verbatim ask included: the goal in their words; owns: list; read-only: list; rules: no formatting or cleanup outside owns, no commit, no typecheck, lint, build or test run (those run once, at integration), Edit tool only (CRLF repos); the lane's own verify step (render the page, run the one test file, hit the one endpoint); return shape: lane | files changed | what changed | verified how | left open. 4. Launch every lane in ONE message with the Agent tool: subagent_type general-purpose for edits, Explore for read-only scouts, run_in_background: true. Do not poll and narrate; work the shared lane or the integration checklist while they run. isolation: "worktree" only when two lanes must touch the same file, and then integration is a cherry-pick per worktree, not a merge. 5. Integrate, once, in this order: git status --porcelain and git diff --name-only; every path maps to exactly one lane's owns, or it is reverted and that lane re-run with a tighter prompt. Read the full diff yourself. Then the repo's gate command one time, plus the test suite. A red gate is fixed here in the main loop, not by re-spawning every lane. 6. Report per lane with the agent's return line beside your own verification. A lane whose claim you did not check is marked (unverified). 7. Commit only after step 5 passes and only if the same message asked; then ship (one PR per work stream, never one per lane).

Output

lanes: 6 launched 10:02 | 6 returned by 10:19 | files touched 23 (0 overlaps)

| lane | files | agent said | I verified | open |
| pages/budget | 4 | fixed truncated amount cells, 375px ok | diff read; /budget rendered | none |
| pages/goals | 3 | ... | (unverified: page needs login) | login-gated check |

gate: pnpm run check pass (once) | tests 214/214 (once)
commit: not asked / <sha> via ship

Must not

  • Never spawn agents for a single-file or single-unit task; say "one unit, doing it inline" and do it.
  • Never let two lanes own the same file; a shared file is its own lane or read-only to all but one.
  • Never run type, lint, build, or tests per agent or per lane; once, at integration.
  • Never accept "done" from an agent without reading its diff; the summary describes intent.
  • Never commit, push, or open a PR before the integration pass, and never one PR per lane.
  • Never let an agent commit, format unrelated files, or edit outside owns.
  • Never report a lane as verified when the check was the agent's own claim.
  • Never chain commands with a double ampersand, in the shell or in a lane prompt; one per line.

Non-goals

Deciding what to fix (repo-audit and fix-all enumerate the work; this skill only runs it wide), merging (ship), workflow scripts (a workflow-authoring skill, if installed), per-lane domain rules (design or clean-code skills, if installed, load inside the lane prompt when relevant).

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.