second-opinion

name

second-opinion

description

Spawn 1-3 independent reviewers to provide critical perspectives on the agent's thinking before presenting to the user. Use when you need a fresh take, alternative approaches, or independent validation on a design, plan, code, or any non-trivial decision. Trigger on "second opinion", "another perspective", "ask opus", "ask codex", "independent review", or /second-opinion. Also trigger proactively during design and plan review phases.

Second Opinion

Spawn 1-3 independent reviewers in parallel to provide critical assessments of your thinking before you present to the user. A second opinion is what a second doctor does: look at the situation fresh, form an independent view, tell you where they agree and where they'd do differently. The output is for you, not the user — they see the result of your improved reasoning, not the raw reviewer responses.

Backends

• opus: Claude Opus subagent via the Agent tool (model: "opus").
• codex: OpenAI via the Codex MCP tool (mcp__codex__codex). Load via ToolSearch first.

Invocation

/second-opinion [backend...] — specify 1-3 backends in any combination.

When choosing how many to spawn:

• 1 reviewer: quick sanity check, low-stakes decisions.
• 2 reviewers: important design decisions, plans with significant tradeoffs.
• 3 reviewers: high-stakes architecture, irreversible decisions.

When spawning multiple reviewers, always launch them in parallel (multiple tool calls in the same turn).

What to send

Distill the situation — don't dump raw conversation. Include:

• What's being decided or evaluated, and why it matters.
• The approach under consideration, or the options on the table.
• Constraints, prior decisions, rejected alternatives.
• Supporting material if applicable (code snippets, data, drafts).

Each reviewer gets the same context. Their independence comes from separate execution, not different inputs.

Philosophy injection

Every reviewer prompt MUST include this block verbatim. Subagents inherit your CLAUDE.md, so this block focuses on the reviewer-specific guidance that isn't already there.

Review principles — follow these strictly:

• You are an independent reviewer, not the decision-maker. Form your own independent assessment, as if coming to it fresh. Agree where you agree, disagree where you disagree — let the disagreement emerge from your analysis, not from a prescribed adversarial stance.
• Apply your CLAUDE.md principles to this review — especially: clean, correct, elegant (no quick fixes, workarounds, or hacks); ideal-first (what's the right answer with no constraints?); depth, not broad; every alternative strong (range from a step toward the ideal to the ideal itself — no weak options, no band-aids).
• Be direct when you disagree. Challenge the approach if it's flawed — don't hedge or soften genuine pushback. But don't manufacture disagreement either.
• Ground in evidence, don't just reason from the summary. Use your tools — open referenced files, fetch primary sources, web-search for current practice, precedents, or known failure modes. Verify load-bearing claims rather than taking them at face value. The value of a second opinion is bringing in what the author didn't — not re-processing what they gave you.

Reviewer prompt template

Same structure for both backends. Inject the principles block verbatim.

You are an independent reviewer providing a second opinion.

[philosophy injection block here, verbatim]

Context:
{distilled summary}

Provide:
1. Your independent assessment of the situation.
2. Where your view aligns with the author's recommendation — and why.
3. Where it diverges — the specific mechanism, not vague hand-waving.
4. Alternative approaches worth considering, with tradeoffs.
5. Your recommendation.

Integrating the responses

Look for the signal across reviewers:

• Convergence — all reviewers reach the same conclusion: strong signal, absorb the confidence.
• Divergence on specifics — agree on direction but differ on details: pick the strongest elements from each.
• Fundamental disagreement — disagree on core approach: most valuable outcome. Decision has genuine tradeoffs you may not have mapped. Dig into the WHY before deciding.
• Unanimous concern — every reviewer flags the same issue: take it seriously even if you initially dismissed it.

Rules for integration

1. Absorb, don't relay. The user sees your improved reasoning, not the raw reviewer output.
2. Don't flip on style preference alone. A reviewer optimizing on a different tradeoff axis (e.g., pragmatic over clean) isn't new information — it's a different axis. Update only on new facts, logical errors, or genuine blind spots. The worst outcome is abandoning a well-reasoned position because someone else sounded confident.
3. Do update when warranted. Genuine blind spots, cleaner abstractions, missed failure modes — absorb them.
4. Own the changed mind. If reviewers shifted your assessment, present updated thinking as your own. The user cares about quality, not process.

Before finalizing: steel-man

After integrating reviewer input, apply the Hold the Line principle: re-articulate your prior position and the goal it served, then check whether the reviewer's input refutes the core on the same goal — or merely operates on a different axis. Reframes (where the reviewer redefines the question rather than critiquing your answer) deserve a harder pass, not a softer one.

The specific second-opinion failure mode: a reviewer's "fresh take" silently replaces the problem you were solving rather than critiquing your answer, and you absorb the reframe as wisdom because it came wrapped in confidence.

If the full procedural check would help, run /steel-man-own-position.