goal-prompting
Installation
SKILL.md
Goal Prompting
Turn intent into a goal prompt that the named harness can pursue and judge honestly. Support advice, drafting, review, cross-harness translation, and explicit activation without forcing one workflow on every request.
Quick Path
- Identify the requested mode: explain, draft, review, translate, or activate. Do not turn a request for guidance or copy into a running goal.
- Identify the target harness. Read references/codex-goals.md for Codex and references/claude-goals.md for Claude Code. When translating, read both.
- Inspect safely available context before asking for facts: repository instructions, plans, tests, scripts, issues, logs, or the supplied draft. Ask one concise question only when the missing answer changes the outcome or completion evidence materially. Continue drafting the settled parts while that answer is pending; do not invent a verification command or activate an unresolved goal to avoid asking.
- Decide whether goal mode fits. Use it for substantial work with a closed, verifiable end state. Recommend a normal prompt, plan, automation, or human review gate when that is the better control surface.
- Write the smallest contract that preserves the outcome, evidence, material boundaries, and real stop conditions. Keep a direct goal payload below the target's documented limit. For domain-specific proof or a file-backed contract, read references/goal-patterns.md.
- Audit the draft for false completion, unverified assumptions, impossible checks, hidden side effects, and target-specific evaluator blind spots.
- Return the requested artifact. Activate it only when the user explicitly asked to set or start the goal and the current session exposes an agent-callable goal-state capability.
For the dated official sources and reviewed community patterns behind these rules, read references/source-notes.md.
Goal Fit
A goal fits substantial multi-turn work with a closed, observable finish line: a bounded migration, failing-to-passing bug fix, benchmark target, finite review queue, or decision-ready research artifact.