Claude Code 命令指南：内置命令速查与团队技能模板

---
description: Review a branch against the team's business-risk checklist
argument-hint: "[base-ref]"
disable-model-invocation: true
---

Review the current branch against this base ref: $ARGUMENTS
If no base is supplied, ask for it. Confirm it exists locally.
Read CLAUDE.md and any repository review guidance first.
Inspect the diff from the merge base to HEAD. Report uncommitted
changes separately; do not silently mix them into the branch review.

Check:
- Authorization and tenant boundaries on every changed data path.
- Validation, failure handling, retries and duplicate side effects.
- Schema compatibility, rollout assumptions and rollback options.
- Regression coverage for the changed behavior.
- Secrets, sensitive logs and changes to public contracts.

For each finding: severity, file and line, triggering scenario,
user impact and a concrete repair or test. Separate verified findings
from questions. Mark checks you could not assess as unknown.
Return findings first, then the checklist with pass/fail/unknown
and supporting evidence. If none are found, say so and name gaps.
Review only. Do not edit files, post comments, approve or merge.

---
description: Reproduce a bug with a failing test before fixing it
argument-hint: "[bug description or local issue file]"
disable-model-invocation: true
---

Bug: $ARGUMENTS
If the expected behavior is unclear, ask before changing code.
Read CLAUDE.md, inspect the working tree and preserve unrelated edits.
Find the relevant code and the repository's existing test runner.
Use local fixtures or mocks; do not call production services.

Add the smallest regression test that demonstrates the reported bug.
Run it before changing implementation. Confirm that it fails for
this bug, not because the environment or test setup is broken.
If you cannot demonstrate that failure, explain the blocker and stop.

Make the smallest implementation change that fixes the behavior.
Rerun the regression test and relevant nearby tests. Do not weaken
assertions to obtain a pass or expand into unrelated refactoring.

Report the cause, changed files, failing-before evidence,
passing-after evidence and any tests you could not run.
Do not commit, push or deploy.

---
description: Plan a migration with compatibility checks and rollback gates
argument-hint: "[scope] [target-state]"
arguments: [scope, target]
disable-model-invocation: true
---

Plan a migration of $scope to $target.
If either input is missing or ambiguous, ask for clarification.
Read CLAUDE.md and inspect the current implementation and consumers.
Do not edit files, execute migrations or connect to production.

Describe the current state with file references. List affected data,
APIs, background jobs, integrations and deployment dependencies.
Separate verified facts from assumptions that need an owner.

Propose small phases with dependencies, acceptance checks and a
stop/go decision for each. Include compatibility between old and new
versions, backfill validation, retries and observability.
For each phase, describe rollback or recovery. Flag irreversible
steps and data that a code rollback cannot restore.

Return a reviewable plan, open questions and the smallest safe pilot.
Do not invent a duration, data volume or owner.

---
description: Draft product release notes from a verified Git range
argument-hint: "[base-ref] [head-ref]"
arguments: [base, head]
disable-model-invocation: true
---

Draft release notes for $base..$head.
If either ref is missing, ask. Confirm both resolve locally.
Read CLAUDE.md and the project's existing release-note format.
Inspect commits in the range and compare the endpoint trees.
Use the changed code and tests to verify what each change does.

Group the draft into Added, Changed, Fixed and Breaking Changes.
Omit empty sections. Describe user impact without exposing internal
secrets or vulnerability details that need coordinated disclosure.
Attach a commit or file reference to each item for the reviewer.
Do not claim a change is deployed, enabled for every customer or
faster by a measured amount without evidence of that fact.

Finish with upgrade actions, unresolved questions and any changes
excluded because their impact could not be established.
Return a draft in chat. Do not edit the changelog, tag a release,
publish, send messages or deploy.

---
description: Draft a weekly engineering changelog from branch history
argument-hint: "[since] [until] [branch]"
arguments: [since, until, branch]
disable-model-invocation: true
---

Prepare a weekly changelog for $branch from $since inclusive to
$until exclusive. If an input is missing, ask. Confirm the timezone
and that the branch exists locally. State whether its checkout may
be stale; do not fetch or assume it matches the remote.

Read CLAUDE.md and the team's update format. Inspect commits reachable
from the branch, filtering by committer timestamp in this window.
Explain that commit timestamps do not prove merge or release dates.
Read diffs for claimed outcomes. Group related changes into themes
and avoid counting a merge and its constituent commits as separate
outcomes. Separate product changes, fixes and maintenance.

For each item, give its impact and a commit or file reference.
Label release status unknown unless deployment evidence is available.
Do not infer blockers, revenue, adoption or completion percentages
from commit messages. List evidence gaps at the end.
Return a concise draft in chat. Do not write files or send the update.