Ask a Co-Worker

Get a fresh perspective from a different AI model when you're stuck, looping, or want to validate an approach. Sometimes all it takes is a different way of looking at the problem.

When to Use This Skill

Auto-invoke (Claude should proactively use this when):

• •You've attempted the same approach 3+ times and it keeps failing
• •You're going in circles — reverting changes, re-trying the same fix
• •An error message is opaque and your debugging isn't converging
• •You need a fundamentally different architectural perspective
• •You've exhausted your ideas and are about to tell the user "I'm stuck"

Manual invoke (user triggers with /ask-coworker):

• •User explicitly wants a second opinion
• •User wants to compare approaches across models
• •Brainstorming — getting multiple angles on a design decision

Available Co-Workers

Always use the best coding models. The point is expert-level second opinions, not saving tokens.

Codex Models (Responses API) — Best for Code

These are OpenAI's dedicated coding models. Use ask-codex.py (they require the Responses API, not Chat Completions).

Chat Models (llm CLI)

General-purpose models available via llm. Use for architecture, design, and non-code questions.

Model Selection

• •Code questions → codex-mini-latest (default) or gpt-5.2-codex (complex)
• •Architecture/design → o3 for deep reasoning or gpt-4.1 for quick takes

How to Ask

Two tools available

ask-codex.py — for Codex models (Responses API):

python C:/Users/robert/project/ask-coworker/ask-codex.py "Your question here"
python C:/Users/robert/project/ask-coworker/ask-codex.py -m gpt-5.2-codex "Your question here"
python C:/Users/robert/project/ask-coworker/ask-codex.py -s "You are a TypeScript expert" "Your question here"

llm — for chat/reasoning models (on PATH):

llm -m o3 "Your question here"
llm -m gpt-4.1 "Your question here"
llm -m gpt-4.1 -s "You are a senior backend engineer" "Your question here"

Rules for Asking Good Questions

DO:

1. •Include the error message — paste the exact error, not a summary
2. •Include the relevant code snippet — 10-30 lines of context, not entire files
3. •State what you've tried — "I tried X and Y, both failed because Z"
4. •Ask a specific question — "Why might this deadlock?" not "Help"
5. •Include the tech stack — "Node 20, TypeScript, PostgreSQL 16, Express"

DON'T:

• •Paste entire files (token waste, dilutes the question)
• •Ask vague questions ("What's wrong with this code?")
• •Use this as a first resort — try solving it yourself first
• •Blindly trust the response — validate before applying

Question Templates

Debugging a persistent error (use Codex)

python C:/Users/robert/project/ask-coworker/ask-codex.py -s "You are a debugging expert. Be concise and specific." "
I'm getting this error repeatedly and can't figure out why:

ERROR: [paste exact error]

Relevant code:
[10-30 lines of relevant code]

What I've tried:
- [approach 1 and why it failed]
- [approach 2 and why it failed]

Tech stack: [languages, frameworks, versions]
What am I missing?
"

Architecture/design decision (use o3)

llm -m o3 "
I need to decide between these approaches for [problem]:

Option A: [description]
Option B: [description]

Context: [constraints, scale, team size, timeline]
Give pros/cons and a clear recommendation.
"

Alternative approach brainstorm

python C:/Users/robert/project/ask-coworker/ask-codex.py -m gpt-5.2-codex -s "Suggest 3 completely different approaches. Be creative." "
Current approach: [what you're doing]
Problem: [why it's not working]
Constraints: [what must be preserved]
"

Interpreting Responses

After receiving a co-worker's response:

1. •Don't blindly apply suggestions — evaluate them critically
2. •Look for the insight, not the exact code — the value is the different perspective
3. •If the suggestion doesn't work, that's fine — it may still have shifted your thinking
4. •Report back to the user — summarize what the co-worker suggested and your assessment:
   • •"I asked Codex about the deadlock. It suggested [X]. I think that's worth trying because [Y]."
   • •"o3 suggested [approach], but I don't think it applies here because [reason]. However, it did make me realize [insight]."

Checking Available Models

# List all installed models
llm models list

# Check which keys are configured
llm keys list

Troubleshooting