agentleFS
seankoji-comseankoji-com/.agents/skills/implement/SKILL.md

implement

Skill2 starsChanged 9 days agoMITView on GitHub

Use when the user invokes implement or requests implementation through T3 delegates, returning reviewed commits for the main thread to integrate and ship.

191 other projects use a version of this.

Adopted by

Most recently taken up by

Used with

Projects with it also use setup-ts-deep-modules, ask-matt 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.
  • liufeng1976Skill

    setup-ts-deep-modules

    Wire dependency-cruiser into a TypeScript repo so each package is a deep module — implementation hidden in subfolders, reachable only through its entry-point files. User-invoked.

    Also in 43 projects that use this

  • GoogleSkill

    ask-matt

    Ask which skill or flow fits your situation. A router over the skills in this repo.

    Also in 91 projects that use this

  • RoentekSkill

    loop-me

    Grill me about specs for the workflows I want to build, within this workspace.

    Also in 50 projects that use this

  • AWSSkill

    to-questionnaire

    Turn a decision you can't fully answer into a questionnaire for someone else to fill in.

    Also in 88 projects that use this

  • ryotaroh180105Skill

    claude-handoff

    Hand the current conversation off to a fresh background agent that picks up the work immediately.

    Also in 42 projects that use this

  • GoogleSkill

    to-spec

    Turn the current conversation into a spec and publish it to the project issue tracker — no interview, just synthesis of what you've already discussed.

    Also in 168 projects that use this

  • +3 more from google

The file

We split this skill into the rules it gives an agent

12 rules. 0 are common, 12 are distinctive, and 0 are very specific to this project.

  • 0

    Common

    Most of the 171 edited versions we read keep these.

  • 12

    Distinctive

    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 the repository rules and working tree.

    line 15
  2. Distinctive:

    If the required T3 route is unavailable, report it without substituting a shell launcher, native agent or top-level thread.

    line 30
  3. Distinctive:

    When cost is comparable, prefer eligibility against the main thread too, so shipping can reuse the review.

    line 60
  4. Distinctive:

    Use T3-isolated worktrees for coding workers.

    line 72
  5. Distinctive:

    Run review-loop as its coordinator, using this skill's model-selection policy and review's shared evidence, severity and eligibility rules.

    line 81
  6. Distinctive:

    Before returning for a decision, checkpoint scoped work locally.

    line 90
  7. Distinctive:

    Do not launch a new isolated lead from the main thread's original base as a resume route.

    line 98
  8. Distinctive:

    If nested reviewer dispatch is unavailable, the lead returns checkpointed work marked review-pending.

    line 102
  9. Distinctive:

    If the upgraded attempt is still stuck, or safe escalation is unsupported, return the blocker and preserved work.

    line 130
  10. Distinctive:

    Create scoped local commits and complete review-loop against the returned revision.

    line 136
  11. Distinctive:

    Stop owned workers, servers and monitoring when finished.

    line 159
  12. Distinctive:

    Keep source/evidence required for pending issue filing accessible to the main thread and issue writer.

    line 163

“Common” is measured: kept by at least 6 in 10 of the 171 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

Implement off the main thread

The main thread dispatches one working lead in an isolated T3 worktree. The lead codes directly, delegates substantial independent parts when useful, and runs review-loop before returning reviewed commits. The main thread owns acceptance, integration and shipping.

Main-thread responsibilities

  1. Read the repository rules and working tree. Give the lead a compact brief: request and authority, scope and exclusions, settled product/design decisions, acceptance criteria, verification, repository identity, exact source HEAD and the main thread's model identity/lineage. Include the paths to this skill, review-loop and review. Delegates receive their task, not the parent conversation.
  2. Isolated delegates start from the parent's committed HEAD, without its dirty changes. Identify required uncommitted prerequisites before dispatch. Commit them only with existing or newly obtained authority; otherwise report the blocker. Do not silently omit them, copy unrelated work or invent a handoff route. Verify the prepared worktree's actual base before editing.
  3. Select the working lead using the model-selection rules below. Discover exact T3 capabilities and use delegate_task with isolation: 'worktree'. Explicitly authorise the lead in its task to act as the implementation coordinator and dispatch scoped worker and reviewer children. Follow the returned next. If the required T3 route is unavailable, report it without substituting a shell launcher, native agent or top-level thread.
  4. On return, verify the worktree, exact HEAD, checks, review and remaining gates against the brief before acceptance or integration. Record every deferred-bug packet under tracking, dispatch it through raise-issue and retain ownership until its issue URL is verified. Code readiness and tracking completion are separate states. Continue independent work under the caller's authority.

Model and effort selection

Optimise expected total cost across implementation, context, review, retries and repairs while meeting the quality bar. Apply this policy to the lead, workers and every review round or additional opinion inside implement. Other workflows retain review's default preferences.

  • Discover provider/model identities, options and tool support; apply T3 preflight before dispatch. Preserve explicit model choices subject to review's eligibility rules.
  • Match capability to the work. A bounded mechanical edit with deterministic checks can use a smaller model and lower supported effort. Ambiguous behavior, shared interfaces, security or state/concurrency work needs stronger reasoning from the start. A low price does not establish suitability.
  • Compare feasible choices using available cost and capability evidence, including likely extra attempts and coordination overhead. If reliable prices or task performance are unavailable, state the heuristic and uncertainty. Do not invent prices, claim a measured optimum or build a model registry.
  • Reserve an eligible reviewer route before assigning coding workers. The reviewer must differ in lineage from the lead/coordinator and every material model author for primary review; additional judges may use review's higher-reasoning exception. When cost is comparable, prefer eligibility against the main thread too, so shipping can reuse the review. Report when that later eligibility is absent. Reassess eligibility after repairs or model changes. Record each model, effort, role, selection reason and exposed usage; unknown usage stays unknown.

Working lead

  1. Work directly by default. Add a worker only for a substantial independent part whose expected benefit justifies startup, context and integration cost. Give each worker non-overlapping ownership, acceptance criteria, exact base, local commit authority and instructions to read review as an implementation participant. Workers and assigned reviewers do not delegate further.
  2. Use T3-isolated worktrees for coding workers. Commit required lead checkpoints before dispatch so workers inherit the intended state. Integrate their scoped commits into the lead's worktree through the repository's permitted method, preserving other work. Evaluate the integrated result, not just each worker's report. Keep one writer per owned scope.
  3. Return any material product, design or scope decision to the main thread with the affected work, recommendation and evidence. Continue only independent work; do not ask the user directly from a delegate or settle the decision by expanding scope. A pending decision makes the handoff incomplete.
  4. Run review-loop as its coordinator, using this skill's model-selection policy and review's shared evidence, severity and eligibility rules. Supply that selection context in every review brief. Reviewers work read-only against the lead's frozen current state; they do not need a separate coding worktree. Include all integrated changes and acceptance criteria. Preserve findings, dispositions, scores, best candidate and stagnation counter across repairs and resumes.

Pause and resume

Before returning for a decision, checkpoint scoped work locally. Report task handles, worktree, branch, HEAD, dirty paths, ledger, scores, counters and explicit limits. Mark work incomplete; exclude secrets and unrelated changes.

The main thread returns the settled answer to the same lead child thread through t3_thread_send, restating the coordinator role, dispatch authority, scores, counters and explicit limits, and recording the returned task handle. Before resuming, the lead verifies its existing worktree, branch, HEAD and ledger against the handoff. Do not launch a new isolated lead from the main thread's original base as a resume route. If the bound checkout cannot be recovered, report the blocker; do not set up a replacement checkout in the shell or a child's prompt.

If nested reviewer dispatch is unavailable, the lead returns checkpointed work marked review-pending. The main thread obtains review against that worktree's exact committed HEAD using the same preserved review-loop progress and selection policy and rechecks eligibility against its own lineage. Give the reviewer the lead worktree's absolute path and exact HEAD; require it to verify that HEAD before reading the scoped source there. Do not review the main checkout as a substitute. The main thread sends any required repairs back to the preserved lead through the resume route above. Only a completed eligible review can establish readiness.

One automatic upgrade

Allow one model or effort upgrade for the whole task, shared by the lead, workers and reviewers. It requires evidence of a capability failure or diagnosed lack of progress, with a reason the stronger choice should help. An expected failing regression test alone is not an escalation reason. Start high-risk work at a suitable level rather than relying on a cheap failed attempt.

Record the consumed upgrade before dispatching it. Preserve the worktree state, author identities, findings, review-loop scores and stagnation counter; changing models does not reset any budget. Stop replaced tasks through T3 before handing their ownership to another writer. Cancelling a lead also stops its descendants. For stronger implementation, the lead may use the upgrade on a worker launched from its current committed checkpoint and remain the coordinator. Do not pretend the running lead's model changed. If the coordinator itself must be replaced and T3 cannot preserve its checkout through a supported route, return a blocker with the checkpoint handoff rather than starting again from the parent's base.

If the upgraded attempt is still stuck, or safe escalation is unsupported, return the blocker and preserved work. Do not launch a second upgrade or label budget exhaustion as success.

Ready work and return

Create scoped local commits and complete review-loop against the returned revision. Ready means the agreed criteria, required checks and applicable runtime/UI evidence are met, with an eligible completed cross-lineage review and finding dispositions allowed by review. Missing material evidence or unresolved gates must be reported as incomplete. Apply review's deferred-bug tracking rule. Code-ready work may return with non-blocking filing requests for the main thread to dispatch through raise-issue. Include the evidence, finding IDs, source revision and posting authority, and mark tracking pending until URLs are verified.

Return a compact plain-text handoff containing:

  • Repository, worktree, branch, exact base and reviewed HEAD, and commit IDs.
  • Implemented scope, checks and evidence, findings and dispositions, remaining limitations and the next integration action.
  • Models/lineages, selection reasons, upgrade use, scores, stagnation counter, explicit limits and exposed usage/cost.
  • Review eligibility for main-thread reuse and any integration/base changes that would require affected behavior to be reviewed again.
  • Task/thread manifest including descendants, failures and cancellations; worker worktree paths, evidence locations and cleanup status.
  • Verified deferred-bug issue URLs or pending filing requests and their owner.

Stop owned workers, servers and monitoring when finished. Keep the lead worktree and commits available for parent integration. Remove integrated worker worktrees only through T3 tooling when supported; otherwise report them for T3 cleanup. Preserve evidence needed by the handoff before removing task artifacts. Keep source/evidence required for pending issue filing accessible to the main thread and issue writer; a code-ready handoff does not complete that tracking.

Invoking implement authorises scoped local commits by the lead and its workers in their T3 worktrees, and filing of confirmed deferred bugs arranged by the main thread in the task's repository. Explicit no-post limits apply. It does not authorise committing prerequisites in the parent checkout, parent integration, pushing, PR creation, merging or deployment. Explicit narrower limits still apply; if they prevent the requested commit handoff, report that conflict.

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.