agentleFS
emiliodominguezemiliodominguez/flow-skills/skills/flow-work/SKILL.md

flow-work

Skill1 starsChanged 47 days agoMITView on GitHub

Implementation of a bounded change in verified slices. Use for most coding requests: features, changes, bug fixes or a plan task; multi-worker gated runs go to `flow-orchestrate`. Hands off to `flow-review`, or `flow-verify` via the coordinator.

The file

We split this skill into the rules it gives an agent

6 rules. About 0 are common, 6 are distinctive, and 0 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈6

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 0

    Very specific

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

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Read applicable instructions, nearby code and tests, configuration, and pinned dependency versions.

    line 23
  2. Distinctive:

    When unsure about an API, check official docs or the installed library source.

    line 26
  3. Distinctive:

    Verify in the runtime that carries the risk: browser interaction and visual inspection for UI, real CLI invocations for commands, API and …

    line 33
  4. Distinctive:

    Run required repository gates.

    line 35
  5. Distinctive:

    Before any "done", "fixed" or "passing" claim: name the check that proves it, run it fresh at this revision, read the full output, then …

    line 39
  6. Distinctive:

    Stop and gather evidence when you catch yourself thinking:

    line 49

≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “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

Work

Deliver one verified slice at a time, following repository conventions, with evidence proportionate to the change.

Phase 1: Select the task

  • An explicit plan path selects that plan. Read it and its referenced source. If the path is missing, stop; do not pick another plan.
  • A description selects direct work, even when old plans exist.
  • A bare invocation may select the only unfinished ready or in-progress direct plan. Ask if several are plausible. Orchestrated plans resume through flow-orchestrate.
  • A coordinator's task brief selects delegated mode. Read references/delegated-mode.md before editing.

Current user instructions override an older plan. Reconcile source drift before editing. A ready task has observable acceptance and satisfied dependencies. Report or verify already completed work instead of redoing it.

Phase 2: Ground the implementation

  • Read applicable instructions, nearby code and tests, configuration, and pinned dependency versions.
  • Documented repository rules (JSDoc, comments, formatting, logging, lint) beat generic preferences.
  • Resolve ordinary details from evidence; ask only about consequential missing choices.
  • When unsure about an API, check official docs or the installed library source. Do not add a dependency or build chain because a generic example uses one.

Phase 3: Implement and verify each slice

  • New behavior or a bug fix with a stable contract: work tests-first where useful. Run the failing case, confirm the failure is relevant, implement, rerun.
  • Risky rewrite of existing behavior: characterize the contract first.
  • Reversible prose or low-impact style edits need no test that mirrors implementation.
  • Verify in the runtime that carries the risk: browser interaction and visual inspection for UI, real CLI invocations for commands, API and data checks for integrations. Cover relevant failure, loading, empty, retry, and persistence states. Keep fixture data isolated.
  • Record command, working directory, observed result, and artifact revision. A passing typecheck alone does not demonstrate behavior.
  • Run required repository gates. Broaden checks only for affected dependents or a concrete unresolved risk.

Phase 4: Report completion

  • Before any "done", "fixed" or "passing" claim: name the check that proves it, run it fresh at this revision, read the full output, then state the result with that evidence.
  • Direct plan mode: check off a task only after its acceptance is evidenced. Append a short evidence note and set Status to in-progress or done. If the plan requires independent acceptance, route through flow-verify first.
  • Stay within owned scope. Inspect and preserve unrelated work in a required file; stop only when ownership or intent is unresolvable. Never stage a shared index another worker owns.

When blocked

Refine a falsifiable hypothesis from the failure, shrink the reproduction, and use flow-diagnose for a hard cause. Honor the task's correction budget; without one, stop after three unsuccessful attempts without progress and report the evidence. Missing access or contradictory requirements need a named prerequisite, not retries. Never weaken tests or acceptance to pass.

Red flags

Stop and gather evidence when you catch yourself thinking:

  • "It should work now" or "that was a trivial change": run the check anyway.
  • "The test is wrong": prove it against the contract before touching it.
  • "I'll verify at the end": each slice gets its own evidence.
  • "One more quick fix": after repeated failed attempts, step back to flow-diagnose.

Anti-patterns

  • Broadening a slice with opportunistic refactors or cleanup.
  • Treating an unrun command or unavailable service as passing evidence.
  • Asking to continue work the user already requested, or ignoring new steering.
  • Completing a coordinator's plan task on the worker's own authority.

Done when

  • The slice works at the inspected revision and required checks have real results.
  • Changes follow repository conventions and stay in scope.
  • Direct work has a clear outcome; delegated work has a reviewable artifact, execution-profile evidence, and known gaps.

Continue to flow-review when the task requires it, or return to the coordinator for flow-verify.

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.