Second Opinion Skill

Proactively suggest getting another LLM's perspective when the situation warrants it.

Trigger Conditions

Suggest a second opinion when you detect:

1. Architectural Decisions

• Choosing between design patterns (e.g., repository vs service layer)
• Database schema design decisions
• API design choices (REST vs GraphQL, versioning strategy)
• Service decomposition (monolith vs microservices)
• State management approaches

2. Complex Trade-offs

• Performance vs. readability
• Flexibility vs. simplicity
• DRY vs. explicit code
• Build vs. buy decisions
• Consistency vs. availability trade-offs

3. Critical Code Reviews

• Security-sensitive code (authentication, authorization, crypto)
• Performance-critical paths
• Complex algorithms or data structures
• Code handling financial transactions or PII
• Concurrency and threading logic

4. Explicit Requests (trigger words)

• "another perspective"
• "second opinion"
• "sanity check"
• "what do you think"
• "am I on the right track"
• "does this make sense"
• "is this a good approach"

How to Suggest

A second opinion is most valuable from a different model family than the one that wrote the code. Bind the active assistant surface before suggesting an invocation.

Gemini CLI

Offer only commands packaged by the active gopher-ai-llm-tools extension:

• /ollama <specific question> asks an installed Ollama model. Treat it as

local only after verifying that OLLAMA_HOST resolves to a loopback host.

• /llm-compare <specific question> compares the available providers.

If a bare command conflicts with a user or project command, use the extension-prefixed form shown by /commands list, such as /gopher-ai-llm-tools.ollama. Never suggest command syntax from another assistant surface on Gemini CLI.

Codex

Offer only installed Codex skills:

• $llm-tools:gemini <specific question> asks Google Gemini. Explain that the

prompt goes to a cloud provider and that adding code or repository context requires explicit confirmation.

• $llm-tools:ollama <specific question> asks an installed Ollama model. Its

privacy boundary depends on the configured Ollama endpoint; the provider skill verifies loopback destinations or confirms a non-loopback transfer.

For security-sensitive or proprietary code, present Ollama only with that endpoint check. For architectural decisions, offer either provider and let the user choose; do not invoke one implicitly.

Claude Code

Offer the command that fits the request:

• /codex:review gets OpenAI analysis through the official Codex Claude Code

plugin when installed.

• /llm-tools:codex review <scope> uses the Codex CLI fallback when the

official plugin is missing or declined.

• /gemini <specific question> asks Google Gemini.
• /ollama <specific question> asks a local model.
• /llm-tools:review-loop --llm fable requests a fresh-context Claude review.
• /llm-compare <specific question> compares multiple providers.

In Claude Code, prefer /codex:review, /codex:adversarial-review, or /codex:rescue from codex@openai-codex when that official plugin is installed. Keep scripted pipelines on the existing CLI flow.

For security-sensitive code, explicitly mention the local /ollama option. For a focused challenge review, use /codex:adversarial-review when available. For complex reasoning or rescue work, use /codex:rescue when available.

When NOT to Suggest

Do not suggest second opinions when:

• User is actively implementing (don't interrupt flow)
• Task is simple/straightforward (typos, formatting, simple fixes)
• User has already made a firm decision
• User said "just do it" or similar
• It's routine code changes with clear requirements
• User previously declined suggestions in this session

Privacy Consideration

Before sending code or repository context to a non-local provider, identify the provider and the material that would leave the machine, then obtain explicit confirmation. Never include secrets or unrelated files. Ollama is local only when its effective OLLAMA_HOST is a verified loopback endpoint; model downloads remain a separate user decision.