goal

Execute Goal

Execute an approved package at ai-artifacts/goals/<slug>/goal.md. Invoking this skill with a package is approval to begin implementation; it is not permission to rediscover scope or rewrite the contract.

Load the contract

1. Read repository instructions first.
2. Require an explicit goal.md path from the user. Do not guess the latest or nearest package.
3. Read goal.md and resolve its links relative to the package directory.
4. Require and read facts.md, facts.meta.json, plan.md, and dev-log.md.
5. Read intent.md when the package contains or links it. Older approved packages without intent.md remain valid.
6. Stop if a required file is missing, a link is broken, fact metadata does not match the accepted facts, or a product, scope, architecture, security, or destructive decision remains unresolved.

Accepted facts are the implementation contract. Do not edit intent.md when present, facts.md, or facts.meta.json during execution. Do not revise plan.md unless the user explicitly sends the package back to setup or replanning.

Execute

1. Briefly state the goal, approved scope, and execution approach.
2. Follow the approved plan in order. Use the narrowest implementation that satisfies every accepted fact.
3. Preserve unrelated work. Keep one writer per worktree unless parallel work is deliberately isolated.
4. Append concise entries to dev-log.md: completed work, changed files, validation, blockers, and remaining work. Never rewrite prior entries or claim work not performed.
5. Run every verification required by plan.md.
6. For each fact whose automatedVerification is true, run or add a concrete automated check. Record a blocker instead of pretending a check passed.
7. Record manual evidence for facts that cannot be verified automatically.

If implementation exposes a new material decision or contradicts an accepted fact, stop and ask the user. Do not silently expand scope or mutate the contract.

Delegation