Make Goal

Make Goal

Use this skill to turn engineering intent into a runnable task contract for an agent. The output is a /goal contract, a review of an existing goal, or a catalog-entry draft when explicitly requested. Do not implement the underlying software task unless the user separately asks for execution.

A good goal is not a wish. It is a contract with one measurable objective, real context to inspect, hard constraints, auditable completion, fresh verification, and stop rules for uncertainty or risk.

Operating Modes

• create: The user gives a task, issue, failure, plan, or vague request and wants a new /goal.
• tighten: The user gives an existing draft and wants it made safer or more executable.
• review: The user asks whether a goal is good; lead with defects, missing evidence, and risk.
• catalog: The user wants to add a goal pattern to an awesome-goal-prompts style catalog; read references/catalog-contribution.md.

If the user asks for a goal and also asks you to execute it, first produce the goal contract. Execute only after the user confirms or separately asks for execution.

Read Context First

Before drafting, inspect available context instead of guessing:

1. Read the user's latest request and any pasted error, issue, plan, or draft.
2. If working in a repository, read the nearest AGENTS.md or CLAUDE.md, relevant issue/PLAN/docs/logs, and the files or commands named by the user.
3. If the task references CI, tests, browser behavior, screenshots, migrations, production, data, auth, or security, include the concrete evidence source in CONTEXT.
4. Separate confirmed facts from assumptions. Do not present assumptions as facts.

Ask at most three concise questions only when missing information changes the contract. The highest-value questions are:

• What is the one finish line for this goal?
• What real files, issue, logs, screenshots, failing command, or baseline must be read first?
• What must not change, and what command/report/screenshot proves completion?

If the request is only "make this project better", "fix everything", or similarly broad, use clarify_first: ask for the primary goal, context, and done-when condition instead of drafting a fake comprehensive goal.

Choose The Shape

Use the full shape by default. Use compact only for routine, low-risk work with clear context and one or two simple verification commands.

Use full when any of these are true:

• The task is high-risk, multi-step, long-running, cross-module, migration, security, auth, data, deployment, production, or user-visible frontend work.
• The user mentions plan mode, AGENTS/CLAUDE rules, screenshots, browser smoke tests, CI repair, database changes, or PR readiness.
• The constraints are unclear or the task could drift into unrelated cleanup.

Do not over-process tiny tasks. If the requested change is routine and the diff could be described in one sentence, use compact or answer directly when the user did not ask for a goal.

Read references/contract-shape.md for the full and compact templates.

Artifact-Backed Long Goals

For long-horizon work where a single prompt would become a backlog, keep the /goal itself short and make the goal require durable local artifacts:

• a roadmap or plan file with phases, dependencies, deliverables, acceptance criteria, and mandatory commands;
• a state or progress file updated at phase boundaries;
• one phase spec per independently verifiable unit of work;
• a final audit that re-reads the original roadmap, re-runs aggregated verification, and checks the current working tree for declared deliverables.

Use this only when the task genuinely needs multiple independently verifiable phases. Do not turn a small one-file task into an artifact system.

Pattern Selection

Use source-backed patterns first when a close match exists. They are trusted shapes backed by public examples, not text to copy verbatim.

Use seed patterns only as raw reusable shapes when no source-backed pattern fits. Never describe a seed as collected from X, GitHub, docs, or a named author.

Read:

• references/pattern-selection.md when choosing between source-backed and seed patterns.
• references/source-backed-index.md for common rescue contracts and goal-workflow patterns.
• references/seed-patterns-index.md only when the source-backed index lacks a close shape.

Optional Lint

When reviewing generated output, run the bundled linter if a local shell is available:

python3 scripts/lint_goal.py <goal-or-response.md>

Use --profile compact, --profile data-migration, --profile security-xss, --profile read-only, --profile launch-readiness, or --profile clarify when the task type is known. Combine profiles with commas, such as --profile read-only,security-xss, when a goal has multiple risk modes. Use --json when another script will consume the result.

Run the deterministic fixture evals after changing this skill:

python3 scripts/run_evals.py

Required Contract

Every full goal must contain exactly these sections:

/goal
GOAL:
<One measurable goal.>

CONTEXT:
- <Files, docs, issues, logs, screenshots, commands, baseline, and conventions to inspect first.>

CONSTRAINTS:
- <What must not change; what must be preserved; safety, data, security, test, and scope boundaries.>

DONE WHEN:
- <Mechanically auditable completion condition.>

VERIFY:
- <Fresh command, report, screenshot, artifact, manual check, or explicit blocker.>

OUTPUT:
- <Changed files if execution occurs, key decisions, verification output, remaining risk, follow-up.>

STOP RULES:
- <When to pause instead of guessing.>

Compact form:

/goal <single measurable goal>

Read first: <files/issues/logs>.
Constraints: <what must not change; what must be preserved>.
Done when: <mechanically checkable end state>.
Verify with: <commands/reports/screenshots/evidence>.
Stop if: <missing input, destructive operation, repeated failed fixes, production access, or unclear product decision>.
Final output: <changed files, verification, risks, next action>.

Safety Defaults

Include these constraints unless they conflict with a more specific user requirement:

• Keep scope limited to this goal; do not add unrelated cleanup or "easy improvements".
• Include a Verification integrity: constraint when edits, validation, generated outputs, tests, lint, typecheck, screenshots, or external checks are involved. Use this shape: Verification integrity: do not weaken or bypass tests, assertions, lint, typecheck, validation, generated-output checks, screenshots, or external blockers to make the goal pass; fix the root cause or report the blocker.
• Respect repository instructions and existing patterns.
• Stop if secrets, production access, destructive data operations, missing credentials, or product decisions are required.
• Stop after three failed attempts on the same symptom and revisit the root-cause hypothesis.
• Do not mark complete until the current state has been checked against DONE WHEN.

For full goals that may lead to file edits, include the scope, verification-integrity, and repository-instruction defaults as actual CONSTRAINTS text, not only as internal guidance.

Keep context precise. Name the minimum useful files, issues, logs, screenshots, URLs, commands, or baselines instead of asking the agent to read the whole repository. Treat AGENTS.md and CLAUDE.md as context sources with different tool semantics; do not imply every coding agent loads the same instruction files automatically.

Add category-specific constraints when relevant:

• Security: preserve auth/authz, validate inputs, avoid unsafe HTML sinks, string-built SQL, shell string execution, and secret exposure.
• Data and migrations: write the literal term dry-run into the goal. Require a dry-run when the stack supports it; if true dry-run is unavailable, say so explicitly and require a disposable-environment rehearsal instead. Also require rollback or forward-only recovery notes, row-count/checksum or equivalent integrity evidence, and no silent data loss.
• Frontend: preserve accessibility, keyboard behavior, existing design system, and include browser or screenshot evidence for user-visible changes.
• CI/testing: reproduce or locate the current failure, fix production or fixture causes before changing assertions, and require fresh command output.
• Investigation: keep changes read-only unless the goal explicitly asks for a fix; separate evidence from hypotheses.
• Long-horizon/autonomous execution: if the goal spans multiple phases, require durable roadmap/state/phase artifacts, phase-boundary evidence, a final audit against the original roadmap, and complete working-tree deliverable checks that include committed, staged, unstaged, and untracked work.
• Launch/readiness/release-prep: add a scope fuse. If the work expands into multiple independent PR-sized changes, keep the current goal to the smallest launch-readiness pass and list follow-up PRs instead of expanding scope.
• External platform guidance: distinguish required local acceptance criteria from advisory platform recommendations. Record deviations and decide whether to keep or update; do not imply every recommendation must be implemented or that external outcomes such as trending, ranking, traffic, or approval are guaranteed.
• Growth/launch plans: treat stars, traffic, virality, Trending placement, rankings, external review approval, and awesome-list acceptance as external outcome targets. Keep the deliverable to a source-backed report, first PR scope, launch sequence, and measurement plan unless the user explicitly asks for execution.
• Platform settings: if a goal references repository settings, topics, social preview, Pages, GitHub metadata, traffic APIs, or external submissions, document owner-access work as manual steps or exact blockers instead of treating it as completed.

For non-trivial goals, include risk, trade-off, or blast-radius reporting in OUTPUT rather than adding a new top-level section.

Read references/quality-bar.md before finalizing high-risk goals.

Output Rules

When creating or tightening a goal:

1. Output the final contract in one fenced text block.
2. After the block, add a short note only if needed: assumptions made, missing context, or why full/compact was chosen.
3. Ask one concise handoff question after the note: 是否现在启动这个 goal？如果你回复“启动”或“开始执行”，我会先注册 active goal（如果当前运行环境支持），再按这个 contract 执行；否则只保留为可复制的 goal。
4. Do not start executing the goal in the same response. If the user confirms in the next message, treat that as an explicit request to execute the most recently generated goal contract.
5. In Codex or any runtime with an active-goal mechanism, confirmation means: check whether an active goal already exists, call the runtime's goal-registration tool such as create_goal with the contract objective before executing, and call the completion tool such as update_goal only after DONE WHEN and VERIFY are satisfied. If no active-goal mechanism is available, say that explicitly before executing manually.
6. Do not include lengthy internal analysis or catalog trivia.

When reviewing a goal:

1. Lead with findings ordered by severity.
2. Cite missing sections, vague done conditions, unverifiable VERIFY, unsafe constraints, or weak stop rules.
3. Provide a revised contract if the user asked for one.

When writing a catalog entry:

1. Read references/catalog-contribution.md.
2. Distinguish source-backed from seed.
3. Do not invent source metadata.

Final Self-Check

Before responding, verify:

• There is one primary goal, not a backlog.
• CONTEXT names real files, issues, logs, docs, commands, screenshots, or baselines when available.
• CONSTRAINTS say what must not change.
• DONE WHEN is mechanically checkable or evidence-based.
• VERIFY requires fresh evidence from the actual run or an exact blocker, not only a generic phrase such as "run tests".
• STOP RULES cover secrets, production access, destructive operations, repository settings or owner approval when relevant, unclear decisions, and repeated failed fixes.
• Launch, release-prep, growth, roadmap, Trending, and awesome-list goals include a scope fuse and manual platform blocker language.
• Seed patterns are not presented as source-backed evidence.