Docs / Reference / Working together (for AIs)

Working together: helpers, models and feedback

For any Claude conversation that works for the owner: how to use other AIs as helpers, choose the right model for each part of the work, keep the quality perfect, and report how each model did.

The owner has several AI models available on his Macs: Claude, Codex (GPT) and Antigravity (Gemini). You are not alone on a job. Split big work into parts and give each part to the AI that does that kind of work best. The work goes faster and each AI's usage lasts longer. The result must still be perfect: you stay responsible for everything a helper hands back.

Talking to other conversations is covered in Talking to other conversations. Reporting how models did is covered in Model feedback.

1. Your key

Every conversation on the owner's Macs can use its Mac's feedback key. It is already in the Keychain:

# on Mac Studio 1
MMF=$(security find-generic-password -s maximind-feedback-key -a studio1-claude -w)
# on Mac Studio 2
MMF=$(security find-generic-password -s maximind-feedback-key -a studio2-claude -w)
# on the MacBook Pro
MMF=$(security find-generic-password -s maximind-feedback-key -a macbook-claude -w)

Never print, log or copy the key. It can only read the model guide and send or read feedback. In each record you send, put your own conversation's name in source.project, so the owner can see who reported.

2. Before choosing a model: read the guide

curl -s -H "x-api-key: $MMF" https://maximind.dev/api/ai/model-guide
curl -s -H "x-api-key: $MMF" "https://maximind.dev/api/ai/model-advice?task_kind=code&language=en"

The guide is built from every AI's real results and from the owner's own verdicts. It is rebuilt every night and after new feedback. It says which model is best for which task kind, language and size, which defects to watch for, and how big one call's output can safely be. If the guide says a model is weak at a kind of task, do not give it that task. The owner sees the same knowledge at /admin/model-tips.

What we have learned so far. The live guide overrides this list:

Helper Good at Watch out
Codex (gpt-6-sol) Page and UI design, full features in a web app, code audits, security-sensitive code, big files Uses Codex quota: split sensibly
Gemini 3.1 Pro High (Antigravity) Self-contained modules with an exact interface, data mapping, text tasks Review for global side effects and error paths. Weak on app features (iOS). It may refuse when a file reads secrets, so keep those files out of its task. It has claimed tests that did not run: re-run them yourself.
Claude helpers (your Agent tool) Reviews, careful edits in your own repo, integration, research Same quota as you

3. Split the work

  1. Plan the parts by file area, so no two helpers edit the same file.
  2. Write one task file per part:
    • the goal in the owner's words;
    • the exact interface or contract;
    • which files it may change;
    • the tests it must run;
    • what "done" means.
  3. Give each helper its own git worktree: git worktree add ../<repo>-<part> -b <branch>.
  4. Run the parts in parallel and keep working on your own part meanwhile.
  5. Start helpers from your conversation. Then they appear under you on the owner's Cockpit page automatically.

Commands:

# Codex: use the vietlingo account first; codex-3 only when vietlingo is out of quota or signed out
CODEX_HOME=~/.ai-profiles/codex-vietlingo /opt/homebrew/bin/codex exec -C <worktree> -m gpt-6-sol \
  -c model_reasoning_effort=high -s danger-full-access - < task.md > job.log 2>&1 &

# Antigravity (Gemini): run it inside the worktree
cd <worktree> && ~/.local/bin/agy --dangerously-skip-permissions -p "$(cat task.md)" \
  --model gemini-3.1-pro-high --effort high > job.log 2>&1 &

Both use the owner's existing logins on this Mac. Never add API keys of your own.

4. Quality: you check everything

A helper's "all tests pass" is a claim, not a fact. Before you use any result:

  • Re-run the tests yourself, and the project's full test suite.
  • Read the diff:
    • changes outside the allowed files;
    • global side effects;
    • swallowed errors;
    • weakened or skipped tests;
    • leftover scratch files (git status).
  • For user-facing work, look at it: screenshots at phone and desktop sizes.
  • Send exact, numbered fix lists back to the helper, or fix it yourself when that is quicker.
  • Only merge when it is right. Speed never excuses a worse result.

5. Report every helper job

After each helper job, send one feedback record to MaxiMind (format: Model feedback). Report good results as well as bad ones:

  • the model and how it was run;
  • the task kind and language;
  • verdict and ratings;
  • first-check pass or fail;
  • defects;
  • evidence: test output, counts.
curl -s -X POST -H "x-api-key: $MMF" -H "Content-Type: application/json" \
     https://maximind.dev/api/ai/feedback --data @records.json

This is how every conversation, and the owner, learns which model to trust for what.