菜单
Claude Code automatic advisor for security, reviews, high-risk design, substantial multi-step plans, complex refactors, recurring failures, and risky completion checks. Skip trivial work, ordinary planning, other harnesses, or sessions without an advisor tool.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/jpcaparas/skills --skill claude-code-auto-advisor命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
基于 SOC 职业分类
正在显示 SKILL.md
| name | claude-code-auto-advisor |
| description | Claude Code automatic advisor for security, reviews, high-risk design, substantial multi-step plans, complex refactors, recurring failures, and risky completion checks. Skip trivial work, ordinary planning, other harnesses, or sessions without an advisor tool. |
| compatibility | Requires Claude Code v2.1.98+ with a configured advisor model. Optional helper scripts require python3. |
| metadata | {"version":"1.0.0","short-description":"Auto-consult Claude Code advisor for high-stakes work","openclaw":{"category":"development","requires":{"bins":"[Truncated]"}}} |
| references | ["advisor-policy","claude-code-mechanics"] |
Invoke Claude Code's configured advisor automatically for tasks where a second model materially reduces risk.
Before applying this skill, answer these gates in order:
Not running in Claude Code: Stop. Do not simulate the advisor with subagents, web search, another model, or a different harness feature.
No configured advisor model or no callable advisor tool in this Claude Code session: Do not call the advisor. Continue with normal verification and record the skip reason only when it matters to the final handoff.
Task is security-related, a code review, high-stakes design, multi-step high-level plan validation before substantial work, a risky refactor, production/data-affecting work, a recurring failure, or a complex completion check: Gather enough local evidence, then consult the advisor before committing to the approach or final conclusion. Read .
references/advisor-policy.mdTask is routine, reversible, single-step, ordinary planning, a small checklist, or already fully determined by local evidence: Do not consult the advisor.
Unsure whether the advisor is enabled:
Prefer the actual Claude Code tool list as source of truth. Optionally run python3 scripts/detect_advisor_config.py --format text for a conservative settings preflight, then read references/claude-code-mechanics.md.
| Situation | Action |
|---|---|
| Security, auth, secrets, permissions, sandboxing, supply chain, data exposure | Consult the advisor after initial context gathering and before recommendations or edits |
| Code review, PR review, or requested independent review | Consult the advisor before final findings |
| Big multi-step implementation plan, migration plan, multi-agent handoff, or high-level plan with meaningful blast radius | Draft the plan with evidence and constraints, then consult before work begins |
| Large refactor, migration, concurrency, data model, irreversible deploy, production impact | Consult before locking the plan; consult again before done if risk remains |
| Same error pattern repeats after two credible attempts | Consult before another speculative fix |
| No advisor model/tool, unsupported main model, disabled advisor, or non-Claude harness | Do not call; do not ask the user to configure it unless they requested advisor setup |
| Small formatting change, simple command output, obvious typo, short stable Q&A, one-step plan, or ordinary task planning | Do not consult |
| Need | Read |
|---|---|
| Decide whether a task requires advisor consultation | references/advisor-policy.md |
| Understand Claude Code advisor setup, requirements, model pairing, cost, caching, and failure modes | references/claude-code-mechanics.md |
| Check local Claude Code settings without invoking advisor | scripts/detect_advisor_config.py |
advisorModel is not the same as a callable advisor in the current request if the main model, provider, organization allowlist, or disable flag blocks it.