LLM Advisor

Use Simon Willison's llm CLI to consult other LLMs for second opinions, alternative perspectives, and expert advice on complex problems.

When to Use This Skill

Proactive Use (autonomous)

Use this skill proactively without being asked when:

• Stuck >15 minutes on a bug or problem
• Complex debugging with unclear root cause
• Architecture decisions with significant trade-offs
• Planning complex features that need validation
• Unfamiliar codebase/language where you need guidance
• Security-sensitive code (auth, crypto, input validation)

On-Demand Use (user requests)

Use when the user says:

• "Ask Gemini/GPT/Claude about X"
• "Get a second opinion on this"
• "What would GPT think about this approach?"
• "Check with another model"

Prerequisites

Installation

# Install llm CLI
brew install llm
# or
pip install llm

# Set up OpenAI API key
llm keys set openai

# Install Gemini plugin (optional)
llm install llm-gemini
llm keys set gemini

# Install Anthropic plugin (optional)
llm install llm-anthropic
llm keys set anthropic

Verify Setup

# Check available models
llm models

# Test a simple prompt
llm "Hello, what model are you?"

Model Selection (Current as of 2026)

OpenAI Models

Note: Check llm models to see exactly which model IDs are installed. OpenAI model names change frequently. As of 2026, gpt-4.1 is the flagship GPT and o3/o4-mini are the reasoning models.

Google Gemini Models

Anthropic Claude Models

Model Selection Guidelines

1. Quick questions: Use gpt-4o-mini or gemini-2.0-flash
2. General advice: Use gpt-4.1 or claude-sonnet-4-5
3. Complex debugging: Use o4-mini -o reasoning_effort=high or gemini-2.5-pro -o thinking_budget=32000
4. Code review: Use gpt-4.1 or claude-sonnet-4-5 for thorough analysis
5. Architecture / deep reasoning: Use o3 or claude-opus-4-5
6. Long context (>100K tokens): Use gemini-2.5-pro (1M+ context window)

Command Reference

Basic Prompts

# Simple question
llm "What's the best way to handle rate limiting in a REST API?"

# With specific model
llm -m gpt-4.1 "Explain this error: <error message>"

# With system prompt
llm -s "You are a senior software architect" "Review this design: <design>"

Piped Input

# Analyze code from file
cat src/auth.ts | llm -m gpt-4.1 "Review this authentication code for security issues"

# Analyze git diff
git diff | llm -m gpt-4.1 "Review these changes for bugs"

# Analyze error logs
cat error.log | llm -m gpt-4o-mini "What's causing these errors?"

Options

# Control reasoning effort (OpenAI o-series)
llm -m o4-mini -o reasoning_effort=high "Complex question"
llm -m o4-mini -o reasoning_effort=low "Simple question"

# Enable thinking mode (Gemini)
llm -m gemini-2.5-pro -o thinking_budget=32000 "Complex reasoning task"

# Extract code blocks from response
llm -x "Write a function to parse JSON safely"

Conversations

# Continue previous conversation
llm -c "What about error handling?"

# Start new conversation with context
llm -m gpt-4.1 "I'm debugging a memory leak in a Node.js app..."
llm -c "Here's the heap snapshot: <data>"

Workflow Examples

Hard Bug Debugging

When stuck on a difficult bug:

# Describe the problem with context
llm -m o4-mini -o reasoning_effort=high "I'm debugging this issue:

Error: Connection refused on port 5432
Environment: Docker container, Node.js 22
Stack trace: <paste stack trace>

What I've tried:
1. Verified PostgreSQL is running
2. Checked network settings
3. Tested connection from host

What else should I check?"

Architecture Decision

When evaluating design choices:

llm -m claude-opus-4-5 "I need to decide between these approaches:

Option A: Event-driven with Redis pub/sub
Option B: Direct API calls with circuit breaker

Context:
- Microservices architecture
- ~1000 requests/second
- Eventual consistency is acceptable

What are the trade-offs? Which would you recommend?"

Code Review

Get a second opinion on code changes:

git diff HEAD~1 | llm -m gpt-4.1 "Review this code change for:
- Bugs or logic errors
- Security vulnerabilities
- Performance issues
- Missing edge cases

Be specific about line numbers and issues."

Plan Validation

Before implementing a complex feature:

llm -m o3 "Review this implementation plan:

Feature: User authentication with OAuth2
Steps:
1. Add OAuth2 middleware
2. Create /auth/callback endpoint
3. Store tokens in Redis
4. Add refresh token rotation

What am I missing? What could go wrong?"

Long Document Analysis

For large codebases or long documents, use Gemini's large context window:

# Analyze large codebase
find . -name "*.ts" | head -20 | xargs cat | llm -m gemini-2.5-pro "Summarize the architecture and identify any security concerns"

# Analyze long log files
cat large-app.log | llm -m gemini-2.5-pro "Find patterns indicating memory leaks or performance regressions"

Multi-Model Comparison

For critical decisions, consult multiple models:

# Ask the same question to different models
QUESTION="Should I use JWT or session cookies for authentication in a Next.js app?"
llm -m gpt-4.1 "$QUESTION"
llm -m claude-sonnet-4-5 "$QUESTION"
llm -m gemini-2.5-pro "$QUESTION"

Alternative Approaches

When you want fresh ideas:

llm -m gpt-4.1 "Here's my current approach to <problem>:
<describe approach>

What are alternative ways to solve this? What am I missing?"

Best Practices

Providing Context

1. Include relevant code - Pipe files or paste snippets
2. Describe the environment - Framework, language version, dependencies
3. Explain what you've tried - Helps avoid repeated suggestions
4. State your constraints - Performance requirements, compatibility needs

Formatting Prompts

# Good: Structured with clear sections
llm -m gpt-4.1 "Problem: <description>

Context:
- Language: TypeScript
- Framework: Express
- Environment: Docker

Error: <error message>

Question: What's causing this and how do I fix it?"

# Bad: Vague without context
llm "Why doesn't this work?"

Interpreting Results

1. Verify suggestions - Always test recommendations against your codebase
2. Consider multiple perspectives - Run the same question through different models
3. Cross-reference - Check suggestions against documentation
4. Adapt to your context - Generic advice may need adjustment
5. Models can be wrong - Especially about recent APIs; verify against official docs

When NOT to Use

• Simple, obvious changes (typos, formatting)
• Well-documented operations with clear solutions
• When you need to move fast on low-risk changes
• Repetitive tasks where the pattern is established

Troubleshooting

Model Not Found

# List available models
llm models

# Install missing plugin
llm install llm-gemini
llm install llm-anthropic

# Update to get latest model names
pip install --upgrade llm llm-gemini llm-anthropic

API Key Issues

# Re-set API key
llm keys set openai
llm keys set gemini
llm keys set anthropic

# Verify key is set
llm keys

Rate Limiting

If you hit rate limits:

1. Switch to a different model temporarily (e.g., gpt-4o-mini instead of gpt-4.1)
2. Use reasoning_effort=low for less token usage
3. Wait and retry after a minute

Long Responses Truncated

For complex questions that need detailed answers:

# Use a model with longer output limits
llm -m gemini-2.5-pro "detailed question requiring long answer"

Model Names Drift

OpenAI, Google, and Anthropic update model names frequently. If a model ID doesn't work:

llm models list  # See all installed models with current IDs