plan-execution
Installation
SKILL.md
Plan Execution
Purpose
Turn a confirmed plan into finished work. The plan already did the thinking: execute it exactly, review each task before moving on, and leave a record that survives your own forgetting. Verification evidence is required only when the user explicitly authorizes the relevant command; otherwise report implementation separately from unrun verification.
Workflow
- Set up: establish an isolated workspace via
git-worktrees(never start implementation on the default branch without explicit user consent). Read the plan and its spec; the spec is the binding authority when they conflict. Keep a progress ledger file — context and todo lists are a live view, the ledger is the record; after any context loss, trust the ledger andgit logover recollection, and never re-implement tasks whose commits already exist. - Pre-flight: scan the plan for cross-task interface conflicts — for every task consuming what an earlier task produces, check that producer and consumer agree. Rule on each conflict against the spec before Task 1.
- Choose the execution mode:
- Subagent mode (preferred when available): dispatch one fresh implementer per task with a review after each.
- Inline mode: execute tasks yourself when the user chose it or no subagent tool exists — never fabricate a dispatch.
- Parallel dispatch: only for independent problem domains with no shared state and no overlapping write scopes (e.g. unrelated failing test files); related failures are investigated together, not split.
- Execute continuously. Do not pause to ask "should I continue?" between tasks. Only four things stop you: an irreversible or destructive operation; a security-sensitive action; a side effect outside the workspace that norms require asking about (merge, push to a shared branch, publish); or a plan so broken every path forward is a guess.
- Verify each performed operation independently. Before choosing commands, check every operation the command may perform: test execution, dependency restore, compilation/build, packaging, migration, or deployment. Run only operations explicitly authorized and only when prerequisites already exist. If a requested test command would implicitly restore or build, report that conflict and select a no-restore/no-build form when available; otherwise do not run it. Record each skipped operation explicitly in the ledger and output (for example, “tests run; restore/build not run”).
- When output contradicts the plan: if the code is wrong, use
debugging— never patch a symptom to make output match. If the plan is wrong, rule on the smallest spec-satisfying change and ledger it asRuling: <finding> — <decision and why> — <cost if wrong>; deviating without a ledgered ruling is a decision made in secret. - After all tasks: one fresh whole-branch review (a reviewer subagent if available). Critical and important findings get one fix pass; minor findings go to the ledger.
- Hand off the finished branch to
branch-integration.