ai-panel

AI Panel — Second Opinions from Other Models

When to Use

• Plan mode: Get alternative perspectives on architecture, technology choices, or trade-offs before committing to an approach.
• Implementation: Sanity-check a specific decision (e.g. library choice, data model, algorithm).
• Debugging: Ask for fresh eyes on a tricky problem.

When NOT to use: Trivial decisions, obvious fixes, or when you're already highly confident. Don't waste premium requests on things that don't benefit from diversity of thought.

Available Models

Model	Command	Strengths
Google Gemini	gemini -p "prompt"	Web-grounded knowledge, long context, research questions
GPT (via Copilot)	copilot -p "prompt" --model gpt-5.2	Strong reasoning, code generation, structured output

Important: Copilot defaults to Claude Sonnet — always pass --model gpt-5.2 to get a genuinely different perspective. The point of the panel is model diversity.

How to Call

Always call both models in parallel (two Bash calls in a single message) and wrap in timeout to prevent hangs.

# Gemini (30s timeout)
timeout 30 gemini -p "Context: ... Question: ... Be concise, 3-5 key points."

# GPT via Copilot (30s timeout)
timeout 30 copilot -p "Context: ... Question: ... Be concise, 3-5 key points." --model gpt-5.2 --allow-all-tools

Prompt Guidelines

• Always end with: "Be concise, 3-5 key points." — both models ramble without this.
• Provide enough context for a useful answer (tech stack, constraints, goals).
• Ask a specific question — not "what do you think?" but "which approach is better for X given Y?"
• Keep prompts under ~300 words. Longer prompts cause Gemini to hang.
• Include the most relevant constraint or trade-off to focus the response.

Model Routing — When to Use Which

Ask both (default for /second-opinion)

Architecture decisions, technology choices, approach comparison — anything where genuine disagreement is valuable.

Gemini only

• Questions about recent technologies, APIs, or documentation (web-grounded)
• Broad "what are the options" research questions
• When you need a non-code perspective

GPT only

• Code-specific questions (algorithm choice, data structure, implementation pattern)
• When you need structured/formatted output
• Quick sanity checks on a specific approach

Ask neither

• You're already confident and the decision is straightforward
• The question is project-specific (only you have the context)
• The question is about style/preference rather than correctness

Output Handling

Gemini

Strip the Loaded cached credentials. prefix line if present. The useful response starts after it.

Copilot (GPT)

Strip the trailing usage stats block. Everything from the blank line before Total usage est: onwards should be removed. Only keep the actual response content.

Timeout

If a model times out (30s), note it in the output and proceed with whichever model responded. Don't retry — the user can re-run if needed.

Presenting Results

Format the panel's output clearly:

### Gemini
<cleaned response>

### GPT
<cleaned response>

### Synthesis
- **Agreement**: Where both models align
- **Disagreement**: Where they differ and why
- **Recommendation**: Your own assessment, weighing all perspectives

Always give your own recommendation last. The panel informs your judgement — it doesn't replace it.