magi

Name: magi
Author: vinta

120stars21forks

How to install magi?

vinta/hal-9000/magi

$npx -y skills add vinta/hal-9000 --skill magi

Installs into the current project.

Summary

Use when brainstorming ideas, features, or directions for a project where independent perspectives from different model families (Claude/Codex/Gemini) would surface blind spots and spark creative options the user hasn't considered — especially "what cool things can I add", "what

Files · 1

View on GitHub

SKILL.md

1	# MAGI
2
3	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.
4
5	- Scientist: reasons directly as Claude Opus (no external dispatch)
6	- Mother: delegates to OpenAI Codex via `mcp__codex__codex` MCP tool
7	- Woman: delegates to Google Gemini via `gemini` CLI
8
9	## Process
10
11	### 1. Clarify
12
13	If the question is underspecified, use `AskUserQuestion` to nail down purpose, constraints, and success criteria. Skip if already clear and actionable.
14
15	- Ask questions one at a time to refine the idea
16	- Prefer multiple choice questions when possible, but open-ended is fine too
17	- Only one question per message
18
19	### 2. Setup
20
21	Read the personality and reference files, then spawn all teammates in parallel.
22
23	Files to read:
24
25	- Personalities: [MAGI-1.md](personalities/MAGI-1.md), [MAGI-2.md](personalities/MAGI-2.md), [MAGI-3.md](personalities/MAGI-3.md)
26	- References: [codex.md](references/codex.md), [gemini.md](references/gemini.md)
27
28	Create team with `TeamCreate` using name `magi-{topic}` (e.g., `magi-auth-strategy`).
29
30	Spawn all 3 teammates in a single message (3 parallel `Agent` calls with `team_name` set):
31
32	\| Teammate \| `name` \| `subagent_type` \| Prompt includes \|
33	\| --------- \| ----------- \| ----------------- \| ----------------------------------------------------------------------- \|
34	\| Scientist \| `scientist` \| `general-purpose` \| MAGI-1.md personality + question (reasons directly as Opus) \|
35	\| Mother \| `mother` \| `general-purpose` \| MAGI-2.md personality + codex.md (dispatches to Codex MCP) + question \|
36	\| Woman \| `woman` \| `general-purpose` \| MAGI-3.md personality + gemini.md (dispatches to Gemini CLI) + question \|
37
38	Include all clarified context in each spawn prompt: teammates have no conversation history.
39
40	### 3. Parallel Exploration
41
42	The lead's role is coordination only:
43
44	- Wait for teammates to send proposals via `SendMessage`
45	- 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.
46
47	### 4. Consolidate + Present
48
49	Collect all proposals, then:
50
51	1. Deduplicate similar proposals (attribute to all teammates/models that proposed it)
52	2. Group by theme if many proposals
53	3. Present each option with:
54	- Which teammate(s) and model(s) proposed it (e.g., "Scientist [Opus]", "Mother [Codex]")
55	- Trade-off analysis from each perspective
56	- Who tagged it as their top pick and why
57	4. Ask the user to select an option via `AskUserQuestion`
58	5. Ask via `AskUserQuestion` what to do next:
59	- Write a plan: teardown, then handoff to `writing-plans`
60	- Debate: another round of critique (see below)
61	- Done: teardown, no further action
62
63	### 5. Debate (optional, user-triggered)
64
65	Only runs if the user requests it. Can be repeated.
66
67	1. Broadcast the consolidated option list to all 3 teammates via `SendMessage`
68	2. Each teammate critiques the proposals through their model (Scientist reasons directly; Mother and Woman follow Debate Mode in their reference files)
69	3. Collect updated stances and re-present to the user (back to step 4)
70
71	### 6. Teardown
72
73	Tear down only when the user selects Write a plan or Done.
74
75	1. `shutdown_request` to each teammate
76	2. Wait for all shutdown approvals
77	3. `TeamDelete`
78
79	### 7. Handoff (write a plan path only)
80
81	After teardown, invoke `writing-plans` skill with the chosen option(s) as context.
82
83	## Gotchas
84
85	- 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.
86	- `TeamDelete` fails if teammates are still active. Always send `shutdown_request` to all three and wait for approvals before calling `TeamDelete`.
87	- 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.
88	- Gemini has no persistent thread. Unlike Codex, each Gemini call is stateless. For

People who liked this also viewed

6 picks

Type

magi

120stars21forks

View on GitHub

How to install magi?

vinta/hal-9000/magi

$npx -y skills add vinta/hal-9000 --skill magi

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	# MAGI
2
3	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.
4
5	- Scientist: reasons directly as Claude Opus (no external dispatch)
6	- Mother: delegates to OpenAI Codex via `mcp__codex__codex` MCP tool
7	- Woman: delegates to Google Gemini via `gemini` CLI
8
9	## Process
10
11	### 1. Clarify
12
13	If the question is underspecified, use `AskUserQuestion` to nail down purpose, constraints, and success criteria. Skip if already clear and actionable.
14
15	- Ask questions one at a time to refine the idea
16	- Prefer multiple choice questions when possible, but open-ended is fine too
17	- Only one question per message
18
19	### 2. Setup
20
21	Read the personality and reference files, then spawn all teammates in parallel.
22
23	Files to read:
24
25	- Personalities: [MAGI-1.md](personalities/MAGI-1.md), [MAGI-2.md](personalities/MAGI-2.md), [MAGI-3.md](personalities/MAGI-3.md)
26	- References: [codex.md](references/codex.md), [gemini.md](references/gemini.md)
27
28	Create team with `TeamCreate` using name `magi-{topic}` (e.g., `magi-auth-strategy`).
29
30	Spawn all 3 teammates in a single message (3 parallel `Agent` calls with `team_name` set):
31
32	\| Teammate \| `name` \| `subagent_type` \| Prompt includes \|
33	\| --------- \| ----------- \| ----------------- \| ----------------------------------------------------------------------- \|
34	\| Scientist \| `scientist` \| `general-purpose` \| MAGI-1.md personality + question (reasons directly as Opus) \|
35	\| Mother \| `mother` \| `general-purpose` \| MAGI-2.md personality + codex.md (dispatches to Codex MCP) + question \|
36	\| Woman \| `woman` \| `general-purpose` \| MAGI-3.md personality + gemini.md (dispatches to Gemini CLI) + question \|
37
38	Include all clarified context in each spawn prompt: teammates have no conversation history.
39
40	### 3. Parallel Exploration
41
42	The lead's role is coordination only:
43
44	- Wait for teammates to send proposals via `SendMessage`
45	- 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.
46
47	### 4. Consolidate + Present
48
49	Collect all proposals, then:
50
51	1. Deduplicate similar proposals (attribute to all teammates/models that proposed it)
52	2. Group by theme if many proposals
53	3. Present each option with:
54	- Which teammate(s) and model(s) proposed it (e.g., "Scientist [Opus]", "Mother [Codex]")
55	- Trade-off analysis from each perspective
56	- Who tagged it as their top pick and why
57	4. Ask the user to select an option via `AskUserQuestion`
58	5. Ask via `AskUserQuestion` what to do next:
59	- Write a plan: teardown, then handoff to `writing-plans`
60	- Debate: another round of critique (see below)
61	- Done: teardown, no further action
62
63	### 5. Debate (optional, user-triggered)
64
65	Only runs if the user requests it. Can be repeated.
66
67	1. Broadcast the consolidated option list to all 3 teammates via `SendMessage`
68	2. Each teammate critiques the proposals through their model (Scientist reasons directly; Mother and Woman follow Debate Mode in their reference files)
69	3. Collect updated stances and re-present to the user (back to step 4)
70
71	### 6. Teardown
72
73	Tear down only when the user selects Write a plan or Done.
74
75	1. `shutdown_request` to each teammate
76	2. Wait for all shutdown approvals
77	3. `TeamDelete`
78
79	### 7. Handoff (write a plan path only)
80
81	After teardown, invoke `writing-plans` skill with the chosen option(s) as context.
82
83	## Gotchas
84
85	- 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.
86	- `TeamDelete` fails if teammates are still active. Always send `shutdown_request` to all three and wait for approvals before calling `TeamDelete`.
87	- 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.
88	- Gemini has no persistent thread. Unlike Codex, each Gemini call is stateless. For

People who liked this also viewed

6 picks

Type