Advisor Skill

This skill enables Claude to consult other AI models when facing challenging problems that benefit from an outside perspective.

When to Use This Skill

Use the advisor skill when:

• Stuck on a problem - Exhausted obvious approaches and need fresh ideas
• Complex architectural decisions - Need validation or alternative perspectives on design choices
• Technical uncertainty - Unclear about the best approach among multiple options
• Debugging challenging issues - Standard debugging hasn't revealed the root cause
• Validating assumptions - Want to verify an approach before implementing
• Need specialized expertise - Problem involves domains where another perspective helps

Available Advisors

Codex

Best for: Deep reasoning problems, complex logical challenges, multi-step analysis

Command:

codex e "your query here" 2>/dev/null

Use when:

• Complex reasoning or planning problems requiring deep thought
• Multi-faceted analysis with many interconnected factors
• Evaluating trade-offs between approaches
• Difficult algorithmic or logical challenges
• Need step-by-step reasoning through complex scenarios

Gemini

Best for: Online research, current information, real-world data, web-based queries

Command:

gemini "your query here" 2>/dev/null

Use when:

• Need current information or recent developments
• Looking up documentation, APIs, or libraries
• Research on tools, frameworks, or best practices
• Finding examples or solutions from the web
• Checking current status of technologies or standards

Usage Pattern

1. Identify the Need

Before consulting an advisor, clearly identify:

• What problem you're facing
• What you've already tried
• What specific insight you need

2. Formulate the Query

Craft a clear, focused query that:

• Provides necessary context
• States the specific question or problem
• Mentions constraints or requirements
• Avoids unnecessary details

3. Execute the Consultation

Use the Bash tool with appropriate timeout (10 minutes = 600000ms):

codex e "How can I efficiently find duplicate entries in a large dataset with 100M+ records while minimizing memory usage?" 2>/dev/null

or

gemini "I'm deciding between microservices and monolith architecture for a medium-sized team. What factors should guide this decision?" 2>/dev/null

4. Integrate the Response

After receiving advice:

• Evaluate its applicability to your specific context
• Combine it with your own knowledge
• Adapt recommendations to fit the actual requirements
• Make the final decision based on all available information

Best Practices

DO:

• Use advisors when genuinely stuck or uncertain
• Provide clear, focused queries with relevant context
• Allow the full 10-minute timeout for complex queries
• Consider the advisor's perspective alongside your own analysis
• Choose the right advisor for the problem type (Codex for deep reasoning, Gemini for online research)

DON'T:

• Consult advisors for straightforward problems you can solve directly
• Use advisors as a substitute for your own analysis
• Include sensitive or proprietary information in queries
• Accept advice blindly without evaluating its fit for your context
• Make excessive queries - reserve for genuinely challenging situations

Command Template

Always use the Bash tool with these parameters:

Bash tool parameters:
- command: codex e "query" 2>/dev/null  OR  gemini "query" 2>/dev/null
- description: "Consulting [Codex/Gemini] about [brief problem description]"
- timeout: 600000  (10 minutes in milliseconds)
- dangerouslyDisableSandbox: true  (REQUIRED - see note below)

Important: Sandbox Requirement

You MUST set dangerouslyDisableSandbox: true when calling codex or gemini.

These tools need to write session files outside the default sandbox-allowed directories. Without this parameter, they'll fail with:

Read-only file system (os error 30)

This is safe because these are user-installed, whitelisted tools that require filesystem access for session management.

Example Scenarios

Example 1: Deep Reasoning - Architecture Trade-offs

Situation: Need to deeply analyze trade-offs between multiple architectural approaches.

Query to Codex:

codex e "I need to design a notification system that handles 100K notifications/day with delivery guarantees. Compare event-driven (Kafka), queue-based (RabbitMQ), and polling approaches. Consider: failure handling, scaling, operational complexity, latency requirements, and team expertise (3 backend devs, no DevOps). What would you recommend and why?" 2>/dev/null

Example 2: Online Research - Current Best Practices

Situation: Need to know current best practices for a technology or framework.

Query to Gemini:

gemini "What are the current best practices for implementing authentication in Next.js 14 apps in 2024? I need to understand popular libraries, session vs JWT trade-offs, and security considerations." 2>/dev/null

Example 3: Deep Reasoning - Complex Algorithm Design

Situation: Complex algorithmic problem requiring multi-step reasoning.

Query to Codex:

codex e "I need to implement a job scheduler that: 1) respects dependencies between jobs, 2) handles priority levels, 3) supports job cancellation mid-execution, 4) distributes across multiple workers. Walk through the data structures and algorithm design. What are the edge cases and how would you handle them?" 2>/dev/null

Example 4: Online Research - Library/Tool Investigation

Situation: Need current information about tools, libraries, or frameworks.

Query to Gemini:

gemini "I need a Python library for parsing PDFs with table extraction. Research current options (2024), compare their accuracy, performance, and ease of use. What do developers recommend?" 2>/dev/null

Example 5: Deep Reasoning - Debugging Complex Logic

Situation: Complex bug requiring careful logical reasoning through multiple possibilities.

Query to Codex:

codex e "I have a distributed transaction that occasionally commits on service A but fails on service B, leaving inconsistent state. The retry logic sometimes creates duplicates. Given: idempotency keys, 2PC is not an option, compensation logic exists but races with retries. Reason through the failure scenarios and design a robust solution." 2>/dev/null

Example 6: Online Research - Documentation Lookup

Situation: Need to find specific API documentation or examples.

Query to Gemini:

gemini "Find documentation and examples for using PostgreSQL's LISTEN/NOTIFY feature with connection pooling (pgbouncer). What are the gotchas and how do production systems handle this?" 2>/dev/null

Integration with Regular Workflow

The advisor skill complements Claude's normal problem-solving:

1. Attempt direct solution - Use your knowledge and available tools first
2. Identify complexity - Recognize when a problem is particularly challenging
3. Consult advisor - Get outside perspective on the challenging aspects
4. Synthesize solution - Combine advisor input with your analysis
5. Implement - Execute the solution using all available information

Remember: Advisors provide valuable perspectives, but you remain responsible for the final implementation and ensuring it fits the specific context and requirements.