external-llm-consulting（external LLM consulting 命令行）

External LLM Consulting

Audience: Developers seeking alternative AI perspectives on architecture and design decisions. Goal: Invoke external LLM CLIs in read-only/sandbox mode to get structured second opinions, then present results for comparison with Claude's perspective.

When to Use

• Validating architectural choices with diverse AI perspectives
• Exploring alternative approaches Claude might not consider
• High-stakes design decisions benefiting from model diversity
• Surfacing blind spots through cross-model analysis

General Workflow

1. Gather relevant context (CLAUDE.md, AGENTS.md, existing code patterns, constraints)
2. Formulate a focused prompt with context, specific questions, and structured output request
3. Execute the external CLI in sandbox/read-only mode
4. Parse and structure the response for comparison

Codex Consulting

Prerequisites

• Codex CLI installed (codex --version to verify)
• OpenAI authentication configured (codex login)

Available Models

Parse model from prompt if specified (e.g., "using codex-mini, analyze..." or "model: gpt-5.1-codex-max").

CLI Execution

Run Codex in read-only sandbox mode:

codex exec --sandbox read-only --model <model> "<prompt>"

Important flags:

• --sandbox read-only - Prevents any code modifications
• --model <model> - Model to use (default: gpt-5.1-codex)

For complex prompts, use heredoc:

codex exec --sandbox read-only --model gpt-5.1-codex "$(cat <<'EOF'
Context: [codebase context]

Question: [specific architectural question]

Please provide:
1. 2-3 alternative approaches
2. Trade-offs for each approach
3. Your recommendation with reasoning
EOF
)"

Codex Error Handling

Gemini Consulting

Prerequisites

• Gemini CLI installed (gemini --version to verify)
• Google authentication configured

Available Models

Parse model from prompt if specified (e.g., "using flash, analyze..." or "model: gemini-3.0-pro-preview").

CLI Execution

Run Gemini in sandbox mode:

gemini --sandbox --output-format text --model <model> "<prompt>"

Important flags:

• --sandbox - Prevents any code modifications
• --output-format text - Returns plain text (vs json/stream-json)
• --model <model> - Model to use (default: gemini-3.0-pro-preview)

For complex prompts, use heredoc:

gemini --sandbox --output-format text --model gemini-3.0-pro-preview "$(cat <<'EOF'
Context: [codebase context]

Question: [specific architectural question]

Please provide:
1. 2-3 alternative approaches
2. Trade-offs for each approach
3. Your recommendation with reasoning
EOF
)"

Gemini Error Handling

Prompt Formulation

Create a focused prompt that:

• Provides necessary context about the codebase
• Asks specific questions about the decision
• Requests structured output (options with trade-offs)

Include relevant sections from CLAUDE.md/AGENTS.md and any project constraints mentioned in the user's request.

Output Format

## [Codex/Gemini] Consulting Results

**Query:** [Original question/topic]

**Model:** [model used] (via [Codex/Gemini] CLI)

### Alternative Perspectives

#### Option 1: [Name]
- **Approach:** [Description]
- **Pros:** [Benefits]
- **Cons:** [Drawbacks]

#### Option 2: [Name]
- **Approach:** [Description]
- **Pros:** [Benefits]
- **Cons:** [Drawbacks]

#### Option 3: [Name]
- **Approach:** [Description]
- **Pros:** [Benefits]
- **Cons:** [Drawbacks]

### Recommendation
[LLM's preferred approach and reasoning]

### Key Insights
- [Insight 1 - something Claude might not have considered]
- [Insight 2]
- [Insight 3]

### Raw Output
<details>
<summary>Full Response</summary>

[Complete unedited response]

</details>

Safety Constraints

• Always use sandbox/read-only mode - Never allow code modifications
• No secrets in prompts - Don't include API keys, credentials, or sensitive data
• Verify responses - External LLM suggestions are opinions, not authoritative answers