claude（Claude 搜索）

Claude Code CLI

Use this skill when you need the local claude CLI from a Codex-style harness. Bias toward non-interactive claude -p runs. Treat the CLI as an automation tool, not an interactive brainstorming partner, unless the user explicitly asks to stay inside Claude.

First Check

Confirm the CLI is healthy before you build a workflow around it:

claude --version
claude auth status --text

If either command fails, stop and report the setup or auth problem. Do not retry task prompts until preflight works.

Default Flows

Use exactly one prompt source per claude -p run:

• short prompt: pass it as the trailing argument
• long or multi-line prompt: pass it over stdin

Do not mix stdin with a trailing prompt argument. The CLI can merge both inputs, which makes harness behavior unreliable.

One-shot analysis

Default choice for summarization, critique, and read-only analysis:

claude -p \
  --model sonnet \
  --effort medium \
  --permission-mode plan \
  --no-session-persistence \
  "Summarize the uncommitted changes"

One-shot edits

Use this when Claude should change files in the current repo:

claude -p \
  --model sonnet \
  --effort high \
  --permission-mode acceptEdits \
  --no-session-persistence \
  "Fix the failing tests in the current repo"

Narrow the working set first if the change is broad or the repo is noisy.

Continue or resume

Use persisted sessions only when you expect to continue later.

• If you want a disposable one-shot run, keep --no-session-persistence.
• If you want to continue later, omit that flag.
• Prefer -r <session-id> -p when the exact session matters.
• Use -c -p only when reusing the latest local conversation is clearly safe.

printf '%s\n' "Continue and focus on the failing tests only" | claude -c -p

printf '%s\n' "Finish the refactor and summarize the remaining risks" | \
  claude -r <session-id> -p

Structured output

Use JSON only when another tool will parse the result:

claude -p \
  --model sonnet \
  --effort medium \
  --permission-mode plan \
  --no-session-persistence \
  --output-format json \
  "Summarize the current diff"

--output-format json returns an envelope, not raw model text. The main body is usually in result. Check subtype and errors, not only is_error.

Review And Plan Critique

Use the same workflow for code review and implementation-plan critique. Start broad once, then narrow scope if needed. Do not keep retrying with slightly different wording.

Broad attempt

Use repo-native review for ordinary local-change review:

claude -p \
  --model sonnet \
  --effort medium \
  --permission-mode plan \
  --no-session-persistence \
  "Review the current uncommitted changes in this repo. Focus on concrete bugs, regressions, misleading docs, packaging issues, and risky assumptions. Output sections in this exact order: Findings, Open questions, Residual risks. Findings should come first with file references. If there are no findings, say 'No findings'."

Use stdin-only prompts for longer plan critique requests:

claude -p \
  --model sonnet \
  --effort medium \
  --permission-mode plan \
  --no-session-persistence \
  < /tmp/claude_prompt.txt

Narrowing ladder

If scope needs tightening, narrow in this order:

1. git diff --staged when the user clearly means staged work
2. git diff --unified=3 -- <paths...> when the worktree is noisy or large
3. one file or subsystem at a time if the narrowed diff is still too broad

Example narrowed diff review:

git diff --unified=3 -- path/to/file1 path/to/file2 | \
  claude -p \
    --model sonnet \
    --effort medium \
    --permission-mode plan \
    --no-session-persistence \
    "Review this diff only. Focus on concrete bugs, regressions, misleading docs, packaging issues, and risky assumptions. Output sections in this exact order: Findings, Open questions, Residual risks. Findings should come first with file references. If there are no findings, say 'No findings'."

Repo-native review can inspect unrelated worktree changes. Use staged or file-scoped diffs when scope matters.

Hard Rules

Treat every failed or suspicious run as one of these cases, then take the single next action listed here.

Preflight failure

claude --version or claude auth status --text fails.

Next action: stop and report the setup or auth problem.

Non-zero exit

claude -p exits non-zero.

Next action: stop and report the exit reason or stderr. Do not rerun the same shape.

Slow but still running

The command is still running and has not failed.

Next action: wait once. If it remains too slow for the task, narrow scope.

Empty stdout

The command exits cleanly but prints nothing useful.

Next action: inspect stderr, then rerun once in plain text with a smaller prompt. If that also fails, stop and report it.

JSON envelope issue

JSON mode returns an envelope but the useful payload is missing or unusable.

Next action: check result, subtype, and errors. If the envelope still does not give a usable answer, rerun once with --output-format text.

Scope too broad

Claude is slow, vague, or obviously reviewing the wrong surface.

Next action: switch to the narrowing ladder. Change scope, not wording.

Tool or configuration limit

Claude reports a tool limit, configuration limit, or similar execution problem.

Next action: simplify the scope or prompt. Do not retry the same broad command with slightly different phrasing.

Retry budget

For review and plan critique:

• allow at most one broad attempt
• allow at most one narrowed attempt
• after that, stop and summarize what failed, what was tried, and the next safest narrower attempt

Advanced Knobs

Keep the main workflow simple. Advanced flags drift more often and should be verified against the local CLI help before first use in a new environment.

Read references/claude-cli.md only when you need:

• tool allow or deny lists
• system prompt control
• streaming output
• worktree mode
• settings overrides
• more detailed JSON behavior notes

Update Check

On first use in a session, silently check:

1. https://raw.githubusercontent.com/abpai/skills/main/versions.json
2. Compare the remote claude version to this file's metadata.version
3. If remote is newer, pause before the task and ask whether to run:

npx skills update claude

If the fetch fails, skip the check silently.