Magi

MAGI

Multi-model brainstorming panel. Three teammates explore a question in parallel, each backed by a different model family, then the lead consolidates their proposals for the user.

• Scientist: reasons directly as Claude Opus (no external dispatch)
• Mother: delegates to OpenAI Codex via mcp__codex__codex MCP tool
• Woman: delegates to Google Gemini via gemini CLI

Process

1. Clarify

If the question is underspecified, use AskUserQuestion to nail down purpose, constraints, and success criteria. Skip if already clear and actionable.

• Ask questions one at a time to refine the idea
• Prefer multiple choice questions when possible, but open-ended is fine too
• Only one question per message

2. Setup

Read the personality and reference files, then spawn all teammates in parallel.

Files to read:

• Personalities: [MAGI-1.md](personalities/MAGI-1.md), [MAGI-2.md](personalities/MAGI-2.md), [MAGI-3.md](personalities/MAGI-3.md)
• References: [codex.md](references/codex.md), [gemini.md](references/gemini.md)

Create team with TeamCreate using name magi-{topic} (e.g., magi-auth-strategy).

Spawn all 3 teammates in a single message (3 parallel Agent calls with team_name set):

| Teammate | name | subagent_type | Prompt includes | | --------- | ----------- | ----------------- | ----------------------------------------------------------------------- | | Scientist | scientist | general-purpose | MAGI-1.md personality + question (reasons directly as Opus) | | Mother | mother | general-purpose | MAGI-2.md personality + codex.md (dispatches to Codex MCP) + question | | Woman | woman | general-purpose | MAGI-3.md personality + gemini.md (dispatches to Gemini CLI) + question |

Include all clarified context in each spawn prompt: teammates have no conversation history.

3. Parallel Exploration

The lead's role is coordination only:

• Wait for teammates to send proposals via SendMessage
• Forward any teammate clarifying questions to the user via AskUserQuestion, noting which teammate (and model) asked. Never answer on the user's behalf: only the user answers.

4. Consolidate + Present

Collect all proposals, then:

1. Deduplicate similar proposals (attribute to all teammates/models that proposed it)
2. Group by theme if many proposals
3. Present each option with:

• Which teammate(s) and model(s) proposed it (e.g., "Scientist [Opus]", "Mother [Codex]")
• Trade-off analysis from each perspective
• Who tagged it as their top pick and why

1. Ask the user to select an option via AskUserQuestion
2. Ask via AskUserQuestion what to do next:

• Write a plan: teardown, then handoff to writing-plans
• Debate: another round of critique (see below)
• Done: teardown, no further action

5. Debate (optional, user-triggered)

Only runs if the user requests it. Can be repeated.

1. Broadcast the consolidated option list to all 3 teammates via SendMessage
2. Each teammate critiques the proposals through their model (Scientist reasons directly; Mother and Woman follow Debate Mode in their reference files)
3. Collect updated stances and re-present to the user (back to step 4)

6. Teardown

Tear down only when the user selects Write a plan or Done.

1. shutdown_request to each teammate
2. Wait for all shutdown approvals
3. TeamDelete

7. Handoff (write a plan path only)

After teardown, invoke writing-plans skill with the chosen option(s) as context.

Gotchas

• Teammates have no conversation history. Everything they need must be in the spawn prompt — the user's question, clarified context, CLAUDE.md, and their personality/reference files. If you forget context from the Clarify step, the teammate works blind.
• TeamDelete fails if teammates are still active. Always send shutdown_request to all three and wait for approvals before calling TeamDelete.
• Save the Codex threadId. Mother's first mcp__codex__codex call returns a threadId needed for debate follow-ups via mcp__codex__codex-reply. If lost, the debate round must re-send full context.
• Gemini has no persistent thread. Unlike Codex, each Gemini call is stateless. For debate rounds, the full proposal list and persona must be re-sent every time.
• Teammates may not report back. If a teammate goes silent, send a SendMessage nudge. After 2 minutes of silence, collect what you have and present partial results.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: vinta
• Source: vinta/hal-9000
• License: MIT

Install and usage instructions live in the source repository linked above.