agentleFS
felipeJJfelipeJJ/fullStack-pipeline/skills/pipeline/SKILL.md

pipeline

Skill1 starsChanged 42 days agoMITView on GitHub

Use when starting any development task (feature, bugfix, refactor, hotfix) in a project that has a pipeline profile (.claude/pipeline/profile.json), before writing or editing any production code.

29 other projects use a version of this.

Adopted by

Most recently taken up by

Used with

Projects with it also use follow-up, status and 7 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.
  • n24q02mSkill

    follow-up

    Find emails needing follow-up -- unanswered sent emails, flagged unreplied, draft contextual responses

    Also in 3 projects that use this

  • AnthropicSkill

    status

    Case status summary by audience — client-facing (plain language), internal (for the professor), or court-ready (formal caption format per local rules). Same facts, different framing and depth. Use when a student needs to update the client, brief the professor, or prepare a court status report.

    Also in 5 projects that use this

  • genomewalkerSkill

    explore

    RLM-style recursive memory exploration - dynamically navigate the memory graph

    Also in 3 projects that use this

  • genomewalkerSkill

    resume

    Resume a thread by loading its ~800-token context capsule

    Also in 3 projects that use this

  • AnthropicSkill

    analyze

    Answer data questions -- from quick lookups to full analyses. Use when looking up a single metric, investigating what's driving a trend or drop, comparing segments over time, or preparing a formal data report for stakeholders.

    Also in 3 projects that use this

  • dubwoahSkill

    ask

    Answer questions about the codebase without generating or modifying any code. Use when the user wants to understand how something works, find where something is defined, or explore the project structure. Invoked with /ask.

    Also in 3 projects that use this

The file

We split this skill into the rules it gives an agent

5 rules. 0 are common, 3 are distinctive, and 2 are very specific to this project.

  • 0

    Common

    Most of the 22 edited versions we read keep these.

  • 3

    Distinctive

    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
  1. Very specific:

    If .claude/pipeline/profile.json does not exist, stop and run the init-profile skill first.

    line 10
  2. Distinctive:

    Read the profile now — it is the project's contract and several of its fields drive this skill directly:

    line 12
  3. Distinctive:

    Create one todo per step when the task begins.

    line 34
  4. Very specific:

    Never prefix a commit with PIPELINE_SKIP_GATE=1 on your own initiative — the bypass exists for the user to authorize explicitly, and every …

    line 54
  5. Distinctive:

    Verify deliverables directly (read the files, run the tests) — an agent reporting "done" is a claim, not evidence.

    line 63

“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

Development Pipeline

You are the Orchestrator of one task, from request to committed code. You coordinate agents, decide architecture with the user, and on level 3+ tasks you never write production code yourself.

If .claude/pipeline/profile.json does not exist, stop and run the init-profile skill first. If the project has no CLAUDE.md, stop as well and run init-profile's CLAUDE.md step — planning without project context (domain glossary, permanent decisions, "never" rules) produces generic plans that fail review.

Read the profile now — it is the project's contract and several of its fields drive this skill directly:

  • areas — every task-cli call below means node <plugin-root>/scripts/task-cli.js <command>. The plugin root is announced at session start by the plugin's SessionStart hook (and is two directories above this skill's base directory). Never store it in the project — the install path is versioned and changes on every plugin update; an old profile's cliPath field is obsolete and ignored.
  • areas.<area>.skills — project/user skills the area's agents must follow. In every tester/dev/reviewer prompt, name the area's skills and state they are binding; agents load them via the Skill tool. If a listed skill fails to resolve at spawn time, read its SKILL.md yourself and inline the content in the agent's prompt instead — never silently drop a declared skill.
  • models — when spawning an agent, apply the matrix cell for that role and the task's level band. A cell is either a model name ("sonnet") or {"model": "sonnet", "effort": "low"} — pass model at spawn, and effort too where the spawn surface supports it (e.g. Workflow agents). inherit/absent = session model. task-cli start snapshots the applied cells into the registry for cost reporting.
  • branch — when perTask is true and the task level ≥ minLevel, create the branch from pattern (e.g. feat/task-{id}) right after task-cli start, before any test or code is written.
  • rework.maxAttempts — replaces the "3 attempts" default in step 5; rework.incrementalReview toggles per-phase review on level 5.
  • language — talk to the user in language.user; code identifiers/comments follow language.code; strings shown by the product follow language.productStrings.
  • lean — off | lite | full | ultra (absent = off): the lean mode adapted from Ponytail. The Lean notes in the steps below apply when it isn't off; the agents' part reaches them through the spawn brief.

Agent prompts stay short. The plugin's SubagentStart hook hands every tester, dev and reviewer a brief at spawn — plugin root, active task, plan path, TDD state, all areas (globs, test command, notes, skills), languages, CLAUDE.md and the lean rules for its role. A prompt needs only what the brief can't know: the area, the plan path, the tests to pass and adjust (dev), the changed files (reviewer). Don't paste profile data, CLAUDE.md or lean rules into prompts.

Complexity level (decide first, with the user)

LevelMeaningMechanics
1-2Small fix, localized changePlain subagents (tester → dev), fire-and-collect
3-4Medium feature, refactor with impactNamed agents per area + SendMessage coordination
5Large feature, architectural changeNamed agents + phased execution + incremental review per phase

Steps

Create one todo per step when the task begins.

  1. Register + discovery — task-cli new "<title>" --type=<improvement|bugfix> --level=N creates TASK-<id> and its plan file. Ask the open questions with AskUserQuestion (batch up to 4 per call, never one-by-one). Write the answers into the plan file's Discovery section immediately after receiving them — before planning, not later.

    Lean: before asking, climb rungs 2–5 yourself — grep the repo for an existing helper, check the standard library, a native platform feature, an already-installed dependency — and turn what you find into options. Rung 1 ("does this need to exist?") and adding a new dependency are always the user's call: they go in the same AskUserQuestion batch, never decided by you or an agent. lite: offer the lazier option, but recommend what was asked. full: recommend the laziest option that works (listed first, "(Recommended)"). ultra: also ask rung 1 for every requirement beyond the core.

  2. Plan — fill the plan file: mini-tasks with checkboxes, architecture options (present to the user, they decide), impact analysis (which existing tests cover the code being touched — list them, or state none: <reason>), and, when the task spans areas, the cross-area contract in its own section: exact field names, value formats, endpoints, shared params. The contract lives in the file, not in prompts — it must survive context compression. Run task-cli start <id>. Tell the user the plan file path so they can review it before execution.

    Lean (full, ultra): every mini-task names its rung (— rung 2: reuse src/lib/slug.ts, — rung 4: native <input type="date">, — rung 7: minimal, with validation); a bugfix's impact analysis lists every caller of the function touched; whatever the ladder or the user deferred goes under "Out of scope" as plain bullets with the trigger to revisit — testers don't test it and devs don't build it.

  3. Tests first (red) — spawn one tester agent per affected area, all in parallel. Each writes failing tests for the new behavior, runs them, and reports: which tests failed (expected) and which passed (suspicious — a test that passes before implementation proves nothing). Review the tests against the plan, then run task-cli tdd-red <id> --area=<area> for each area: it runs the area's testCommand (or --cmd="..." narrowed to the new tests) and records red only if the run fails — a green run is refused. Read the output: the failures must be assertions on the new behavior, not syntax/import errors. The TDD hook tracks red per area and blocks production edits in an area — through Write/Edit and through Bash (>, sed -i, cp, mv, rm, ...) — until that area is red. An area with genuinely no testable code path (docs, config): task-cli tdd-skip <id> --area=<area> --reason="..." (without --area it applies to every area) — the reason is recorded and visible in the registry.
  4. Implement — spawn dev agents per area; parallel when they touch different files (the default), sequenced when one depends on the other's output. Each dev prompt states: the plan file path, which tests must pass, and which existing tests to adjust (from the impact analysis). Devs report back via message; they never tick checkboxes or edit the registry. A dev reporting a higher rung than the plan's (a helper nobody listed, a part a native feature covers): inside the plan's scope → let it proceed; a scope change → AskUserQuestion first, and record the decision.
  5. Review — spawn the reviewer agent with the task id and the changed-file list. It runs the affected areas' test commands, the complexity ratchet (scripts/complexity-check.js — it compares against the commit recorded by task-cli start, so untouched legacy code is never blamed) on changed files, and checks the plan was fully delivered. Verdict: approve, or return to the dev with a precise defect description — and on every rejection run task-cli rework <id> --area=<area> --reason="<short>" so the attempt is recorded (this feeds the cost/rework report and is how the attempt limit is counted). Max rework.maxAttempts per dev per mini-task (default 3); one attempt before the last, check whether the test is wrong before blaming the implementation; at the limit, stop and ask the user. Level 5 with incrementalReview: review after each phase, not only at the end. Lean findings (gate 6) are handled as references/reviewer-protocol.md says: advisory in full, reuse:/delete: reject in ultra.
  6. Conclude — tick every checkbox in the plan file, present the changed-file summary to the user, and wait for commit approval. One commit per task, never grouped; the message follows commit in the profile (format: "conventional" → type(scope): subject, requireTaskRef → include TASK-<id>, written in commit.language). task-cli ready <id> before committing — it refuses while the plan has unticked checkboxes, and the commit gate blocks git commit until it passes and validates the message convention (heredoc messages included). A deliberately partial delivery (rollback option a, decided by the user) uses ready <id> --partial --reason="...". The post-commit hook records the hash, closes the task and attaches its token usage from the session transcripts automatically.

    Lean: if the reviewer reported lean findings, ask before the commit, in one AskUserQuestion: apply now (one rework cycle), defer (a bullet under "Out of scope"), or ignore — and record the choice in Decisions. task-cli debt lists every deferral and lean: marker with the task it came from.

    Never prefix a commit with PIPELINE_SKIP_GATE=1 on your own initiative — the bypass exists for the user to authorize explicitly, and every bypassed commit is registered as a bypass task that shows up in list and report.

Tests are the spec

If a change makes an existing test fail, never silently edit the test to make it pass. Stop and ask the user: is this an intentional behavior change (update test and code together) or a regression (fix the code)? There is no third option, and "the test was probably outdated" is not a decision you get to make alone.

Agent hygiene

  • Release each named agent as soon as its deliverable is validated — never leave agents idle after their work is accepted.
  • Verify deliverables directly (read the files, run the tests) — an agent reporting "done" is a claim, not evidence.

References

  • references/orchestration.md — read only for level 3+: agent naming, SendMessage, parallelism rules, when a Workflow script beats a team
  • references/reviewer-protocol.md — what to do with a verdict: rework loop, lean findings, rollback options on partial failure
  • references/lean.md — the agents' lean rules (injected by the SubagentStart hook); read it only to change them

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.