second-opinions

name	second-opinions
description	Get validation from a different AI model before committing major changes — detects available LLM CLIs and routes to the best one.
display_name	Second Opinions
brand_color	#4F46E5
local_only	false
group	For Anyone
usage	/second-opinions:run
summary	About to make a big call? Get a gut-check from a second AI with a different perspective before you commit.
favorite	true
default_prompt	Get a second opinion on this implementation or design decision and summarize the strongest agreement, disagreement, and actionable feedback.

Second Opinions

Get validation from a different AI before committing. Any single model — regardless of which one is running — has blind spots shaped by its training, context, and the conversation so far. A different architecture, temperature, or framing catches different things.

When to Use

Mandatory:

• Complex multi-file changes before merge
• Design decisions with multiple valid approaches
• After 2+ hours on a single approach (tunnel vision risk)
• Security-sensitive or performance-critical code

Skip for: trivial fixes, style questions, crystal-clear requirements

Agent Detection

Use the bundled detection script, with an inline fallback if SKILL_DIR isn't set:

bash "${SKILL_DIR}/scripts/detect-llms.sh" --quiet 2>/dev/null || \
  for t in agent claude codex llm; do command -v "$t" >/dev/null 2>&1 && echo "$t"; done

Use the first one found. If none are available, tell the user and skip this step. Treat agent as a router, not a requirement: direct claude -p, codex exec, or any subscribed local-agent CLI is valid. Do not call Anthropic/Claude models through OpenRouter; use local claude -p directly or via agent <opus|sonnet|haiku> --no-fallback, or skip Claude.

Codex caveat: codex loads ~/.codex/config.toml → your CLAUDE.md, so by default it delegates the review back to the cheap-model toolchain (quick-check) — silently defeating the point of a different model's opinion. The detect script already adds --ignore-user-config; also end the prompt with: "Do this review YOURSELF — do not delegate to any other tool or model." Heavy reviews can exceed 10 min (use -c model_reasoning_effort="xhigh", run in background).

Model Selection

Second opinions are about deep analysis, not speed. Use the smartest model available:

Tool	For deep analysis	For quick checks
agent	agent --frontier (codex subscription / GPT-5.5; no Anthropic OpenRouter)	agent --fast (deepseek-v4-flash)
agent when Claude is wanted	agent opus --no-fallback (local claude -p; fails closed)	agent sonnet --no-fallback
codex	GPT-5.5 — see Codex caveat above	—

When this skill is invoked for pre-merge review, design validation, or architecture decisions, prefer agent --frontier, or codex for a genuinely different architecture. If the desired second opinion is specifically Claude, use agent opus --no-fallback; the agent command must route Claude through local claude -p, never OpenRouter. For quick sanity checks, agent --fast is fine.

How to Ask

The prompt is the same regardless of agent — adapt the invocation to whatever's available. The detect-llms.sh script outputs NAME|INVOKE_PATTERN|MODEL_FAMILY|NOTES — use the INVOKE_PATTERN field, substituting {prompt} with your actual prompt.

Pre-Merge Review (Most Common)

Show the diff and ask for a production-readiness check:

Review my git changes for production readiness.

Show the diff from main and check for:
- Correctness and edge cases
- Architecture and design
- Performance implications
- Security concerns

Design Decision Validation

Describe the options and constraints, then ask: What trade-offs am I not seeing?

Targeted Question

Ask one specific question about the implementation — don't fish for general feedback.

Interpreting Results

The other agent is a collaborator, not an authority. Classify each piece of feedback:

Category	Action
Must-fix	Bug, security issue, correctness problem → implement immediately
Should-fix	Genuine simplification, better error handling → implement if clean
Nice-to-have	Alternative approach, style preference → mention to user
Reject	Over-engineering, conflicts with project conventions → skip with reason

If the other agent and your analysis disagree, explain the disagreement to the user and let them decide.

The Red Flags

These thoughts mean STOP and get a second opinion:

• "It works in my tests" — tests only prove known scenarios
• "I've spent 3 hours on this" — sunk cost isn't validation
• "I'm confident this is right" — confidence correlates with blind spots
• "It's obviously the best approach" — obvious to you ≠ optimal

5 minutes of external validation prevents hours of debugging.