execute

Implement a task and verify it. Use when the scope is understood.

SKILL.md

GitHub

Implement the requested task and verify the result.

Prepare

Load execution-workflow only when isolating the work in a jj workspace or dispatching several agents. Use Serena when available and read memories only when prior knowledge could change the work (for example a canonical verification command); if unavailable, use repository evidence. Run jj status before editing only if you need to tell your changes from existing ones at completion.

Name the requested outcome, affected paths, and acceptance checks. For a local, understood change, this is enough planning. For work spanning interfaces or uncertain behavior, identify atomic units, dependencies, and owners before implementation. State the assumption when interpretations differ; ask only under CLAUDE.md's work_selection rule.

Implement

  1. Read the edit targets and nearest relevant pattern. Before adding a family member, trace its sibling through registration/discovery and name required touch points. Before removing or migrating a definition, search the identifier itself, including declarations, comments, and test doubles. Use references to establish caller impact for interface changes.
  2. Implement directly or delegate substantial independent units under the delegation contract. Do not require an agent for each file. Each writer follows output_discipline and returns evidence, not self-approval.
  3. Add or update tests only when the task asks for tests or the repository has tests for that kind of change. A user instruction not to write tests takes precedence over any other guidance. Document changed public behavior. Do not create unrelated tests or docs to fill a role. Never substitute an always-passing test for behavior unavailable here; inspect existing helpers for a usable fixture or adapter first.

Verify and review

  1. Name the applicable test, lint, build, type, or artifact-validation commands from the project. Establish their input scope, configuration, generated outputs, and fixture/service requirements. Ensure new and untracked work is included; keep scratch outputs isolated from other sessions.
  2. Run the checks and inspect assertions as well as exit status. Use diagnostics when available, otherwise repository commands. For unavailable behavior, search relevant scripts/configuration for a fixture, in-memory adapter, or substitute backend; report the search and the remaining gap.
  3. For failing tests, allow one targeted fix attempt across the task, then rerun. Remaining failures are blockers, not permission to weaken the gate. Preserve fail-closed behavior when following a reference implementation.
  4. Review the settled change for correctness, caller impact, and missing acceptance coverage. Add a review or verification agent only when the user asks, the change involves security or data, or an interface change lacks an existing test for the changed behavior. Use a security agent for security or data changes and a verification agent for an untested interface change; use a design or performance agent when the user requests that role. Do not add an agent for uncertain behavior alone, and do not ask an agent to repeat a check already completed. Independent reviews may run together, read-only. Do not assign multiple reviewers the same check. Give claim-verification agents the claim and supporting commands, not merely a diff. If agents are unavailable, perform the checks here and disclose that they were not independent.
  5. Completion requires passing checks over the requested behavior. Static validation of prose/configuration does not establish runtime behavior, and a missing check remains a reported gap.

Finish

Apply memory_policy only when durable learning or unresolved findings warrant a write. Refresh only memories actually checked against the result; do not launch a memory agent or a freshness audit by default.

Return output_contract with changed paths, verification scope (including new files), review concerns checked, and any uncovered acceptance criteria. State the weakest consequential claim and what would confirm it when one remains.