second-opinion Skill
在把答案交给用户之前,由独立评审员交叉验证智能体自己的答案。当用户说 'second opinion'、'sanity check'、'cross-check'、'am I missing something'、'stress-test'、"devil's advocate"、'run a full review'、'this is important'、'high-stakes'、'help me choose between'、'critique this' 等时使用。三种评审风格:魔鬼代言人(单份批评)、董事会辩论(多评审员并行验证)、圆桌讨论(多轮头脑风暴,用于在多条路径间抉择)。不适用于事实查询——请用 WebSearch。当已有用户级……时跳过此技能。
安装方式:把技能目录放入 ~/.claude/skills/(Claude Code)或在 claude.ai 设置中启用;也可复制右侧安装命令一键添加。
技能指令原文(SKILL.md)
Second Opinion — cross-check before you commit
Stress-test your own answer with reviewers who don't share your framing. The reviewer set
depends on what you have available:
- Always available: the isolated
idea-validatoragent (same model family, but NO parent
context — it can't anchor on your framing, and it READS the actual files).
- If you have access to a second model family (Gemini, GPT, …, via a CLI wrapper or API):
add it as an external reviewer. A different training distribution catches different blind
spots. It reviews from the brief only (no repo access) — lean on it for concept and
tech-currency, not file-level facts.
Why the description says to skip it when you have your own board: a user-level review board or
second-opinion skill (your own reviewer agent, a model list, a wrapper for another model family) and this
skill are two procedures for one decision, and they drift apart. This skill is the default for people who
have none.
When to invoke
- Two or more viable paths to choose between → Round-Table.
- High-stakes decision (architecture, launch copy, anything with real rollback cost) →
Boardroom Debate.
- Stuck on the same problem after 2+ attempts, or a non-trivial proposal awaiting approval →
Devil's Advocate.
If none fires, don't invoke — a single-model answer is enough for routine work.
Out of scope: factual lookups ("what's the latest X") — that's WebSearch, not review.
Own thinking first (load-bearing)
Before ANY review style: do your own research, form your own proposal with rationale and
trade-offs, show it to the user. Only then invoke reviewers to critique it. Asking a reviewer
before forming your own position turns it into a seed for the decision instead of a validator —
and a reviewer without your codebase context can confidently seed something wrong.
The three styles
1. Devil's Advocate (single critique)
Spawn idea-validator with a self-contained artifact (the design/decision/plan pasted inline,
plus the file paths it should actually read). Or send the same artifact to your external model
if the claim is about tech or the outside world.
2. Boardroom Debate (parallel validation — the headline pattern)
- Write ONE self-contained artifact. Paste content inline — never rely on file references a
brief-only reviewer might silently fail to load.
- Launch ALL reviewers in the SAME message (parallel calls). Sequential calls destroy
independence — a later reviewer sees the earlier one's framing.
- Build an acceptance ledger: | Concern | Reviewer A | Reviewer B | My evaluation | Action |
- Adjudicate, never count votes. All agreeing can share a blind spot; one dissenter with a
file:line beats abstract agreement. A code-reading reviewer outranks a brief-only one on
facts about the code. The procedure in full — the three acceptance layers and the
claim→cheapest-decisive-check table — is reference/orchestrator-fact-check.md.
- Present the ledger critically: where you accept, where you push back, and why. You make the
final call — reviewer output is INPUT, not the decision.
3. Round-Table (choosing between paths)
Diverge → deepen → converge, with a mandatory user check-in before convergence:
ground the facts (web-check) → generate/challenge options broadly → kill weak ones down to 2-3
with concrete arguments → present survivors to the user, ask for preference → converge on
one primary + a kill list → fact-check the final claims → synthesize.
Critical evaluation rule
After every reviewer call: challenge each recommendation (fact or speculation?), check for
missing context the reviewer couldn't see, verify numbers (predictions are not measurements),
and present a table — reviewer said / my evaluation / action.
Anti-patterns: forwarding reviewer output as-is · accepting everything · treating reviewer
confidence as correctness · running reviews for trivial/stylistic choices · sequential
"parallel" calls.