Research Brainstorm Skill

Description

Generates and cross-validates research ideas using Gemini and Codex in parallel, then synthesizes results with Claude.

Usage

/research-brainstorm "research topic" [--domain physics|ai_ml|statistics|mathematics|paper] [--weights '{"novelty":0.4,"feasibility":0.3,"impact":0.3}'] [--depth low|medium|high|max] [--personas N] [--claude-only]

Arguments

• $ARGUMENTS — The research topic and optional flags:
  • --domain — Research domain (physics, ai_ml, statistics, mathematics, paper). Auto-inferred if omitted.
  • --weights — JSON object of scoring weights for direction ranking. Keys: novelty, feasibility, impact, rigor, scalability. Values must sum to 1.0. If omitted, Claude analyzes the prompt and recommends adaptive weights for user confirmation (see Step 0a).
  • --depth — Controls review depth (default: medium):
    • low — Skip cross-review, go directly to synthesis
    • medium — Standard one-shot cross-review (current behavior)
    • high — Cross-review + adversarial debate round
    • max — Hierarchical MAGI-in-MAGI: N persona subagents run parallel mini-MAGI pipelines, then meta-review + adversarial debate across all perspectives
  • --personas N|auto — Number of domain-specialist subagents for --depth max (default: auto, range: 2-5). When auto, Claude analyzes the topic to determine the optimal persona count. Ignored for other depth levels.
  • --claude-only — Replace all Gemini/Codex MCP calls with Claude Agent subagents. Use when external model endpoints are unavailable or for a Claude-only workflow. Two subagents with distinct cognitive styles (Creative-Divergent and Analytical-Convergent) ensure perspective diversity.

Instructions

MCP Tool Rules

• Gemini: Use the following model fallback chain. Try each model in order; if a call fails (error, timeout, or model-not-found), retry with the next model:
  1. model: "gemini-3.1-pro-preview" (preferred)
  2. model: "gemini-2.5-pro" (fallback)
  3. Claude (last resort — skip Gemini MCP tool, use Claude directly)
• Codex: Use mcp__codex-cli__brainstorm for ideation, mcp__codex-cli__ask-codex for analysis/review. If Codex fails 2+ times, fall back to Claude directly.
• File References: Use @filepath in the prompt parameter to pass saved artifacts (e.g., @brainstorm/codex_ideas.md) instead of pasting file content inline. The CLI tools read files directly, preventing context truncation.
• Web Search: Use web search freely whenever factual verification, recent developments, or literature context would strengthen the discussion:
  • Claude: Use the WebSearch tool directly
  • Gemini: Add search: true to mcp__gemini-cli__ask-gemini or mcp__gemini-cli__brainstorm calls
  • Codex: Add search: true to mcp__codex-cli__ask-codex or mcp__codex-cli__brainstorm calls
  • When to search: prior work verification, methodological precedents, dataset/library availability, related approaches, fact-checking quantitative claims
  • Claude-only mode: Claude Agent subagents cannot use WebSearch. The main Claude agent should search beforehand and include findings in the subagent prompt.

Claude-Only Mode

When --claude-only is active, all Gemini/Codex MCP tool calls are replaced with Claude Agent subagents (subagent_type: general-purpose). The table below maps each original call to its replacement:

Original Call	Replacement	Cognitive Style
mcp__gemini-cli__brainstorm	Agent subagent A	Creative-Divergent: unconventional connections, "What if?" scenarios, wide exploration, questioning assumptions
mcp__codex-cli__brainstorm	Agent subagent B	Analytical-Convergent: step-by-step feasibility, established methodologies, deep evaluation, practical constraints, risk assessment
mcp__gemini-cli__ask-gemini	Agent subagent A	Same Creative-Divergent style
mcp__codex-cli__ask-codex	Agent subagent B	Same Analytical-Convergent style

Key rules for claude-only mode:

1. File access: Subagents use the Read tool to access files (no @filepath syntax).
2. Output filenames: Keep original names (gemini_ideas.md, codex_ideas.md, etc.) — add a header > Source: Claude Agent subagent (claude-only mode, {style}) to each output file.
3. Independence: Both subagents are spawned simultaneously so neither can see the other's output.
4. --depth max internal subagents: Within each persona's mini-MAGI pipeline, use Expansive Explorer (replaces Gemini) and Grounded Builder (replaces Codex) cognitive styles to maintain internal diversity.

LaTeX Formatting Rules

When writing mathematical expressions in any output document (brainstorm ideas, synthesis, etc.):

• Inline math: Use $...$ for short expressions within a sentence (e.g., $x^2 + y^2 = r^2$)
• Display equations: Use $$ on its own line, with the equation on a separate line:

  $$
  \mathcal{L} = -\frac{1}{4} F_{\mu\nu} F^{\mu\nu} + \bar{\psi}(i\gamma^\mu D_\mu - m)\psi
  $$
  

• Never write display equations inline as $$..equation..$$ on a single line — always use line breaks
• Include this formatting instruction in prompts to Gemini and Codex when the topic involves mathematical content

When this skill is invoked, follow these steps exactly:

Step 0: Setup

1. Parse the research topic from $ARGUMENTS. If a --domain flag is provided, note the domain (physics, ai_ml, statistics, mathematics, paper). Otherwise, infer the domain from the topic.
2. Create the output directory: outputs/{sanitized_topic}_{YYYYMMDD}_v{N}/brainstorm/
   • Sanitize the topic: lowercase, replace spaces with underscores, remove special characters, truncate to 50 chars.
   • Use today's date in YYYYMMDD format.
   • Version: Glob for outputs/{sanitized_topic}_{YYYYMMDD}_v*/ and set N = max existing + 1 (start at v1).
3. If a domain template exists at ${CLAUDE_PLUGIN_ROOT}/templates/domains/{domain}.md, read it for context.
4. Parse --weights:
   • If --weights is explicitly provided: Validate that keys are a subset of {novelty, feasibility, impact, rigor, scalability} and values sum to 1.0. Save immediately to brainstorm/weights.json with metadata:

     {
       "weights": { <user-provided weights> },
       "_meta": {
         "method": "explicit",
         "domain": "<detected domain>"
       }
     }
     

     Skip Step 0a entirely.
   • If --weights is not provided: Load domain defaults as a baseline reference only (do not save yet — Step 0a will handle saving after user confirmation):
   • physics: {"novelty": 0.30, "feasibility": 0.15, "impact": 0.25, "rigor": 0.20, "scalability": 0.10}
   • ai_ml: {"novelty": 0.25, "feasibility": 0.25, "impact": 0.25, "rigor": 0.10, "scalability": 0.15}
   • statistics: {"novelty": 0.25, "feasibility": 0.20, "impact": 0.20, "rigor": 0.25, "scalability": 0.10}
   • mathematics: {"novelty": 0.35, "feasibility": 0.10, "impact": 0.20, "rigor": 0.30, "scalability": 0.05}
   • paper: {"novelty": 0.20, "feasibility": 0.25, "impact": 0.30, "rigor": 0.15, "scalability": 0.10}
5. Parse --depth: Accept low, medium, high, or max (default: medium).
   • low — Skip Step 1b (cross-review), go directly to Step 1c (synthesis)
   • medium — Standard one-shot cross-review (current default behavior)
   • high — Cross-review + adversarial debate (Step 1b+)
   • max — Hierarchical MAGI-in-MAGI pipeline (Steps 1-max-a through 1-max-d replace Steps 1a/1b/1b+/1c)
6. Parse --personas N|auto: Accept integer 2-5 or the string auto (default: auto). Only used when --depth max; ignored otherwise.
   • If auto: Defer persona count determination to Step 0b, where Claude analyzes the topic's complexity, number of distinct sub-disciplines, and methodological diversity to select the optimal N (2-5).
   • If an explicit integer is given: Use that value directly.
7. Parse --claude-only: Boolean flag (default: false). When present, all Gemini/Codex MCP calls are replaced with Claude Agent subagents. See the Claude-Only Mode section above for the replacement table and cognitive style definitions.

Step 0a: Adaptive Weight Recommendation

If --weights was explicitly provided: Skip this step entirely (weights already saved in Step 0).

When --weights is omitted, Claude analyzes the research prompt to recommend topic-appropriate weights:

1. Analyze the research prompt for research-nature signals:

   Signal	Example Keywords	Effect
   Theoretical/Formal	proof, formal, theorem, analytical	rigor↑, feasibility↓
   Experimental/Empirical	experiment, data, empirical, measurement	feasibility↑, rigor↑
   Applied/Engineering	deploy, production, tool, system	feasibility↑, scalability↑
   Exploratory/Innovative	novel, first, unexplored, new paradigm	novelty↑, feasibility↓
   Practicality emphasis	practical, real-world, industry	feasibility↑, scalability↑
   Rigor emphasis	rigorous, convergence, guarantee	rigor↑
   Impact emphasis	breakthrough, paradigm, transformative	impact↑, novelty↑
   Narrow scope	specific algorithm, single method	rigor↑, feasibility↑
   Broad scope	interdisciplinary, broad, cross-domain	novelty↑, impact↑

2. Generate recommended weights: Starting from the domain baseline loaded in Step 0, apply adjustments based on detected signals:

   • Each signal adjusts relevant dimensions by ±0.05 to ±0.10
   • After all adjustments, normalize so values sum to 1.0
   • Clamp each dimension to the range [0.05, 0.45] before final normalization

3. Present a comparison table to the user:

   Detected signals: [list of signals found in the prompt]
   
   | Dimension     | Domain Default | Recommended | Adjustment Reason            |
   |---------------|---------------|-------------|------------------------------|
   | novelty       | 0.30          | 0.35        | +0.05 (exploratory research) |
   | feasibility   | 0.15          | 0.10        | -0.05 (theoretical focus)    |
   | impact        | 0.25          | 0.25        | (no change)                  |
   | rigor         | 0.20          | 0.25        | +0.05 (formal methods)       |
   | scalability   | 0.10          | 0.05        | -0.05 (narrow scope)         |
   

4. Ask the user for confirmation using AskUserQuestion:

   • Option A: "Accept recommended weights" (Recommended) — use the adaptive weights
   • Option B: "Use domain defaults" — use the unmodified domain baseline
   • Other: User provides custom weights as JSON

5. Save to brainstorm/weights.json with metadata based on the user's choice:

   {
     "weights": { <chosen weights> },
     "_meta": {
       "method": "<adaptive-recommended|domain-default|custom>",
       "domain": "<detected domain>",
       "domain_defaults": { <original domain baseline> },
       "signals_detected": ["theoretical work", "narrow scope", ...],
       "adjustments_applied": {
         "rigor": "+0.05 (theoretical focus)",
         "feasibility": "-0.05 (theoretical focus)"
       }
     }
   }
   

   • If "Accept recommended": method = "adaptive-recommended"
   • If "Use domain defaults": method = "domain-default", signals_detected and adjustments_applied are still recorded for traceability
   • If custom JSON: method = "custom"

Step 0b: Dynamic Persona Casting

After setup, Claude analyzes the topic and domain to assign specialist personas:

For --depth low|medium|high (2 personas):

1. Analyze the topic's sub-disciplines, methodologies, and key challenges.
2. Assign a Gemini persona — a domain expert profile best suited for creative/theoretical ideation on this topic (e.g., "Theoretical cosmologist specializing in dark energy models" or "Bayesian statistician with expertise in causal inference").
3. Assign a Codex persona — a practitioner/builder profile best suited for implementation-focused ideation (e.g., "Computational physicist with GPU simulation experience" or "ML engineer specializing in scalable training pipelines").
4. Each persona definition should include: name/title, expertise areas (3-5 bullet points), and a guiding question that shapes their perspective. Name the persona after a real historical figure (위인) whose work is closely related to the persona's domain (e.g., a statistical learning theorist → "R. A. Fisher", a theoretical physicist → "Emmy Noether", a computational pioneer → "John von Neumann"). This immediately signals the persona's intellectual lineage and analytical style.
5. Personas complement the domain template — they do not override it. The domain template provides general context; personas provide topic-specific focus.
6. If --claude-only: Relabel the personas in brainstorm/personas.md:
   • "Gemini persona" → "Subagent A (Creative-Divergent)"
   • "Codex persona" → "Subagent B (Analytical-Convergent)"
   • Include the cognitive style directive in each persona definition so subagents receive it directly.
7. Save to brainstorm/personas.md.

For --depth max (N personas):

1. Analyze the topic's sub-disciplines, methodologies, and key challenges.
2. Determine N (if --personas auto):
   • Evaluate the topic along these dimensions:
     • Sub-discipline count: How many distinct research sub-fields does this topic span?
     • Methodological diversity: Does it require both theoretical and empirical approaches? Simulation? Formal proofs?
     • Interdisciplinary breadth: Does it bridge multiple domains (e.g., physics + ML, statistics + biology)?
   • Selection heuristic:
     • N=2: Narrow, well-defined topic within a single sub-field (e.g., "optimizing a specific algorithm")
     • N=3: Standard research topic spanning theory and practice (e.g., "novel regularization methods for deep learning")
     • N=4: Broad topic crossing 2+ sub-fields or requiring distinct application perspectives (e.g., "physics-informed neural networks for fluid dynamics")
     • N=5: Highly interdisciplinary or contentious topic where a dedicated critic perspective adds value (e.g., "quantum advantage claims in machine learning")
   • Announce the chosen N and reasoning to the user before proceeding.
   • If --personas was given as an explicit integer (2-5), use that value directly and skip this analysis.
3. Cast N domain-specialist personas (model-independent — each persona runs both Gemini and Codex):
   • N=2: Theory/Concepts, Computation/Implementation
   • N=3: Theory/Concepts, Computation/Implementation, Empirical/Validation
   • N=4: + Application/Interdisciplinary
   • N=5: + Critique/Skepticism
4. Each persona definition must include:
   • Name/title — Use a real historical figure (위인) whose work aligns with this persona's domain (e.g., "Alan Turing — Computation Theory Specialist", "Marie Curie — Experimental Methodology Pioneer", "Claude Shannon — Information-Theoretic Analyst"). The figure's intellectual legacy should resonate with the persona's analytical lens.
   • Expertise areas (3-5 bullet points)
   • Guiding question that shapes their perspective
   • Primary lens (one sentence summarizing their analytical viewpoint)
5. Personas are complementary — they should cover distinct dimensions of the research space with minimal overlap.
6. If --claude-only: Within each persona's mini-MAGI pipeline, relabel the internal model roles:
   • "Gemini" → "Expansive Explorer" (pushes boundaries, explores emerging approaches, proposes frontier ideas)
   • "Codex" → "Grounded Builder" (focuses on proven approaches, clear implementation paths, established tools)
   • Include these cognitive style directives in the persona file so subagents receive them directly.
7. Save to brainstorm/personas.md.

Step 1a: Parallel Independent Brainstorming

Execute these two calls simultaneously (in the same message). Prepend the assigned persona from brainstorm/personas.md to each prompt:

Gemini Brainstorming:

mcp__gemini-cli__brainstorm(
  prompt: "[Persona: {gemini_persona_name} — {gemini_persona_expertise}]\nGuiding question: {gemini_guiding_question}\n\nDomain context: @{domain_template_path}\n\n{topic} — Generate diverse, creative research ideas. Consider theoretical foundations, practical applications, novel approaches, and potential breakthroughs.",
  model: "gemini-3.1-pro-preview",  // fallback: "gemini-2.5-pro" → Claude
  domain: "{domain}",
  methodology: "auto",
  ideaCount: 12,
  includeAnalysis: true
)

Note: Omit the Domain context: @{domain_template_path} line from the prompt when no domain template exists.

Codex Brainstorming:

mcp__codex-cli__brainstorm(
  prompt: "[Persona: {codex_persona_name} — {codex_persona_expertise}]\nGuiding question: {codex_guiding_question}\n\nDomain context: @{domain_template_path}\n\n{topic} — Generate implementation-focused research ideas. Consider feasibility, existing tools/libraries, computational requirements, and step-by-step approaches.",
  domain: "{domain}",
  methodology: "auto",
  ideaCount: 12,
  includeAnalysis: true
)

Note: Omit the Domain context: @{domain_template_path} line from the prompt when no domain template exists.

Note: If Codex MCP is unavailable, fall back to mcp__gemini-cli__brainstorm with the Gemini fallback chain and implementation-focused framing.

If --claude-only: Replace both calls above with two Agent subagents, executed simultaneously:

Subagent A (Creative-Divergent, replaces Gemini):

Agent(
  subagent_type: "general-purpose",
  prompt: "You are a Creative-Divergent thinker. Prioritize unconventional connections across adjacent fields, 'What if?' scenarios, wide exploratory breadth, and questioning fundamental assumptions.

[Persona: {gemini_persona_name} — {gemini_persona_expertise}]
Guiding question: {gemini_guiding_question}

Use the Read tool to read the domain template at {domain_template_path} (skip if no template exists).

Research topic: {topic}

Generate 12 diverse, creative research ideas. Consider theoretical foundations, practical applications, novel approaches, and potential breakthroughs. Include feasibility and impact analysis for each idea.

Save your complete output to {output_dir}/brainstorm/gemini_ideas.md. Start the file with:
> Source: Claude Agent subagent (claude-only mode, Creative-Divergent)
> Persona: {gemini_persona_name}
> Timestamp: {ISO timestamp}"
)

Subagent B (Analytical-Convergent, replaces Codex):

Agent(
  subagent_type: "general-purpose",
  prompt: "You are an Analytical-Convergent thinker. Prioritize step-by-step feasibility analysis, established methodologies, deep evaluation of each option, practical constraints, and risk assessment.

[Persona: {codex_persona_name} — {codex_persona_expertise}]
Guiding question: {codex_guiding_question}

Use the Read tool to read the domain template at {domain_template_path} (skip if no template exists).

Research topic: {topic}

Generate 12 implementation-focused research ideas. Consider feasibility, existing tools/libraries, computational requirements, and step-by-step approaches. Include feasibility and impact analysis for each idea.

Save your complete output to {output_dir}/brainstorm/codex_ideas.md. Start the file with:
> Source: Claude Agent subagent (claude-only mode, Analytical-Convergent)
> Persona: {codex_persona_name}
> Timestamp: {ISO timestamp}"
)

Save results to:

• brainstorm/gemini_ideas.md — Gemini's (or Subagent A's) raw output with header noting source, persona, and timestamp
• brainstorm/codex_ideas.md — Codex's (or Subagent B's) raw output with header noting source, persona, and timestamp

Step 1b: Cross-Check (--depth medium or --depth high only)

If --depth low: Skip this step entirely and proceed to Step 1c.

After both brainstorming results are saved, execute these two calls simultaneously. Prepend the assigned persona to each review prompt:

Gemini reviews Codex ideas (Round 1):

mcp__gemini-cli__ask-gemini(
  prompt: "[Persona: {gemini_persona_name} — {gemini_persona_expertise}]\n\nReview the following research ideas for technical feasibility, scientific rigor, novelty, and potential impact. Identify strengths, weaknesses, and suggest improvements for each idea.\n\n@{output_dir}/brainstorm/codex_ideas.md",
  model: "gemini-3.1-pro-preview"  // fallback: "gemini-2.5-pro" → Claude
)

Codex reviews Gemini ideas (Round 1):

mcp__codex-cli__ask-codex(
  prompt: "[Persona: {codex_persona_name} — {codex_persona_expertise}]\n\nReview the following research ideas for implementation feasibility, computational practicality, available tools/datasets, and timeline realism. Identify strengths, weaknesses, and suggest improvements for each idea.\n\n@{output_dir}/brainstorm/gemini_ideas.md"
)

Note: If Codex MCP is unavailable, fall back to mcp__gemini-cli__ask-gemini with the Gemini fallback chain.

If --claude-only: Replace both calls above with two Agent subagents, executed simultaneously:

Subagent A (Creative-Divergent, replaces Gemini reviewing Codex):

Agent(
  subagent_type: "general-purpose",
  prompt: "You are a Creative-Divergent reviewer. Prioritize unconventional connections, wide exploratory breadth, and questioning assumptions.

[Persona: {gemini_persona_name} — {gemini_persona_expertise}]

Use the Read tool to read: {output_dir}/brainstorm/codex_ideas.md

Review these research ideas for technical feasibility, scientific rigor, novelty, and potential impact. Identify strengths, weaknesses, and suggest improvements for each idea.

Save your review to {output_dir}/brainstorm/gemini_review_of_codex.md. Start with:
> Source: Claude Agent subagent (claude-only mode, Creative-Divergent)"
)

Subagent B (Analytical-Convergent, replaces Codex reviewing Gemini):

Agent(
  subagent_type: "general-purpose",
  prompt: "You are an Analytical-Convergent reviewer. Prioritize step-by-step feasibility, practical constraints, and risk assessment.

[Persona: {codex_persona_name} — {codex_persona_expertise}]

Use the Read tool to read: {output_dir}/brainstorm/gemini_ideas.md

Review these research ideas for implementation feasibility, computational practicality, available tools/datasets, and timeline realism. Identify strengths, weaknesses, and suggest improvements for each idea.

Save your review to {output_dir}/brainstorm/codex_review_of_gemini.md. Start with:
> Source: Claude Agent subagent (claude-only mode, Analytical-Convergent)"
)

Save results to:

• brainstorm/gemini_review_of_codex.md
• brainstorm/codex_review_of_gemini.md

Step 1b+: Adversarial Debate (--depth high only)

If --depth low or --depth medium: Skip this step entirely.

After Round 1 cross-review, Claude identifies the top 3 points of disagreement between Gemini and Codex (e.g., conflicting feasibility assessments, divergent novelty ratings, opposing recommendations). Save the disagreement summary to brainstorm/disagreements.md before the debate calls.

Execute Round 2 simultaneously:

Gemini Round 2 — Defend/Concede/Revise:

mcp__gemini-cli__ask-gemini(
  prompt: "[Persona: {gemini_persona_name}]\n\nYou reviewed Codex's ideas and Codex reviewed yours. Here are the top 3 points of disagreement:\n\n@{output_dir}/brainstorm/disagreements.md\n\nFor each disagreement:\n1. **Defend** your position if you believe it is correct, providing additional evidence or reasoning\n2. **Concede** if the opposing argument is stronger, explaining why\n3. **Revise** your assessment to a new position if appropriate\n\nYour original review:\n@{output_dir}/brainstorm/gemini_review_of_codex.md\n\nCodex's review of your ideas:\n@{output_dir}/brainstorm/codex_review_of_gemini.md",
  model: "gemini-3.1-pro-preview"  // fallback chain applies
)

Codex Round 2 — Defend/Concede/Revise:

mcp__codex-cli__ask-codex(
  prompt: "[Persona: {codex_persona_name}]\n\nYou reviewed Gemini's ideas and Gemini reviewed yours. Here are the top 3 points of disagreement:\n\n@{output_dir}/brainstorm/disagreements.md\n\nFor each disagreement:\n1. **Defend** your position if you believe it is correct, providing additional evidence or reasoning\n2. **Concede** if the opposing argument is stronger, explaining why\n3. **Revise** your assessment to a new position if appropriate\n\nYour original review:\n@{output_dir}/brainstorm/codex_review_of_gemini.md\n\nGemini's review of your ideas:\n@{output_dir}/brainstorm/gemini_review_of_codex.md"
)

If --claude-only: Replace both debate calls above with two Agent subagents, executed simultaneously:

Subagent A (Creative-Divergent, replaces Gemini Round 2):

Agent(
  subagent_type: "general-purpose",
  prompt: "You are a Creative-Divergent debater. Prioritize unconventional reasoning and questioning assumptions.

[Persona: {gemini_persona_name}]

You reviewed the opposing side's ideas and they reviewed yours. Use the Read tool to read:
- {output_dir}/brainstorm/disagreements.md
- {output_dir}/brainstorm/gemini_review_of_codex.md
- {output_dir}/brainstorm/codex_review_of_gemini.md

For each disagreement:
1. **Defend** your position if you believe it is correct, providing additional evidence or reasoning
2. **Concede** if the opposing argument is stronger, explaining why
3. **Revise** your assessment to a new position if appropriate

Save to {output_dir}/brainstorm/debate_round2_gemini.md. Start with:
> Source: Claude Agent subagent (claude-only mode, Creative-Divergent)"
)

Subagent B (Analytical-Convergent, replaces Codex Round 2):

Agent(
  subagent_type: "general-purpose",
  prompt: "You are an Analytical-Convergent debater. Prioritize evidence-based reasoning and practical constraints.

[Persona: {codex_persona_name}]

You reviewed the opposing side's ideas and they reviewed yours. Use the Read tool to read:
- {output_dir}/brainstorm/disagreements.md
- {output_dir}/brainstorm/codex_review_of_gemini.md
- {output_dir}/brainstorm/gemini_review_of_codex.md

For each disagreement:
1. **Defend** your position if you believe it is correct, providing additional evidence or reasoning
2. **Concede** if the opposing argument is stronger, explaining why
3. **Revise** your assessment to a new position if appropriate

Save to {output_dir}/brainstorm/debate_round2_codex.md. Start with:
> Source: Claude Agent subagent (claude-only mode, Analytical-Convergent)"
)

Save results to:

• brainstorm/debate_round2_gemini.md
• brainstorm/debate_round2_codex.md

Step 1-max-a: Layer 1 — Parallel Persona Subagents (--depth max only)

If --depth is not max: Skip Steps 1-max-a through 1-max-d entirely. Use Steps 1a/1b/1b+/1c instead.

Spawn N Task subagents simultaneously (one per persona, subagent_type: general-purpose). Each subagent receives the persona definition and executes a self-contained mini-MAGI pipeline:

Each subagent prompt includes:

1. The persona definition (name, expertise, guiding question, primary lens) from brainstorm/personas.md

2. The research topic and domain template (if available)

3. The following 5-step execution plan:

   A. Gemini Brainstorm — Call mcp__gemini-cli__brainstorm with the persona's viewpoint. Save to brainstorm/persona_{i}/gemini_ideas.md.

   B. Codex Brainstorm — Call mcp__codex-cli__brainstorm with the persona's viewpoint. Save to brainstorm/persona_{i}/codex_ideas.md.

   If --claude-only: Replace A and B above with two Agent sub-subagents, executed simultaneously within the persona subagent:

   A'. Expansive Explorer (replaces Gemini):

   Agent(
     subagent_type: "general-purpose",
     prompt: "You are an Expansive Explorer. Push boundaries, explore emerging approaches, and propose ideas at the frontier of what's possible. Challenge conventional thinking and seek novel combinations.
   
   [Persona: {persona_i_name} — {persona_i_expertise}]
   Guiding question: {persona_i_guiding_question}
   
   Use the Read tool to read the domain template at {domain_template_path} (skip if none).
   
   Research topic: {topic}
   
   Generate 12 creative research ideas from this persona's perspective. Save to {output_dir}/brainstorm/persona_{i}/gemini_ideas.md. Start with:
   > Source: Claude Agent subagent (claude-only mode, Expansive Explorer)"
   )
   

   B'. Grounded Builder (replaces Codex):

   Agent(
     subagent_type: "general-purpose",
     prompt: "You are a Grounded Builder. Focus on proven approaches with clear implementation paths, established tools, and demonstrable results. Prioritize what can be built and validated now.
   
   [Persona: {persona_i_name} — {persona_i_expertise}]
   Guiding question: {persona_i_guiding_question}
   
   Use the Read tool to read the domain template at {domain_template_path} (skip if none).
   
   Research topic: {topic}
   
   Generate 12 implementation-focused research ideas from this persona's perspective. Save to {output_dir}/brainstorm/persona_{i}/codex_ideas.md. Start with:
   > Source: Claude Agent subagent (claude-only mode, Grounded Builder)"
   )
   

   C+D. Cross-Review (simultaneous):

   • Gemini reviews Codex ideas using @{output_dir}/brainstorm/persona_{i}/codex_ideas.md → save to brainstorm/persona_{i}/gemini_review_of_codex.md
   • Codex reviews Gemini ideas using @{output_dir}/brainstorm/persona_{i}/gemini_ideas.md → save to brainstorm/persona_{i}/codex_review_of_gemini.md

   If --claude-only: Replace C+D above with two Agent sub-subagents, executed simultaneously:

   C'. Expansive Explorer reviews Grounded Builder's ideas:

   Agent(
     subagent_type: "general-purpose",
     prompt: "You are an Expansive Explorer reviewer. Question constraints and look for hidden potential.
   
   [Persona: {persona_i_name}]
   
   Use the Read tool to read: {output_dir}/brainstorm/persona_{i}/codex_ideas.md
   
   Review these ideas for novelty, scientific rigor, and potential impact. Identify strengths, weaknesses, and improvements.
   
   Save to {output_dir}/brainstorm/persona_{i}/gemini_review_of_codex.md. Start with:
   > Source: Claude Agent subagent (claude-only mode, Expansive Explorer)"
   )
   

   D'. Grounded Builder reviews Expansive Explorer's ideas:

   Agent(
     subagent_type: "general-purpose",
     prompt: "You are a Grounded Builder reviewer. Evaluate practical feasibility and implementation viability.
   
   [Persona: {persona_i_name}]
   
   Use the Read tool to read: {output_dir}/brainstorm/persona_{i}/gemini_ideas.md
   
   Review these ideas for implementation feasibility, computational practicality, and timeline realism. Identify strengths, weaknesses, and improvements.
   
   Save to {output_dir}/brainstorm/persona_{i}/codex_review_of_gemini.md. Start with:
   > Source: Claude Agent subagent (claude-only mode, Grounded Builder)"
   )
   

   E. Persona Conclusion — The subagent synthesizes its top 3 research directions, noting areas of internal agreement and disagreement between the two models (or two cognitive styles in claude-only mode). Save to brainstorm/persona_{i}/conclusion.md.

Wait for all N subagents to complete before proceeding.

Step 1-max-b: Layer 1 — Output Collection

1. Use Glob to verify that all N brainstorm/persona_{i}/conclusion.md files exist.
   • If any are missing, re-spawn the failed subagent(s) and wait for completion.
2. Read all N conclusion.md files.
3. Construct a cross-persona summary identifying:
   • Recurring themes — directions proposed by 2+ personas
   • Unique directions — ideas that appeared in only one persona's output
   • Explicit disagreements — contradictory assessments across personas
4. Consolidate conclusions for Layer 2: Create brainstorm/all_conclusions.md by concatenating all N conclusion files with clear separators:

   # Persona 1: {persona_1_name} — {persona_1_lens}
   
   {persona_1_conclusion_content}
   
   ---
   
   # Persona 2: {persona_2_name} — {persona_2_lens}
   
   {persona_2_conclusion_content}
   
   ---
   
   ... (for all N personas)
   

   This single consolidated file replaces multiple @-references in Layer 2 calls.

Step 1-max-c: Layer 2 — Meta-Review + Adversarial Debate

Phase A — Parallel Meta-Reviews:

Execute simultaneously:

Gemini Meta-Review:

mcp__gemini-cli__ask-gemini(
  prompt: "You are reviewing the outputs of {N} domain-specialist research personas who independently analyzed: {topic}\n\nHere are all persona conclusions:\n@{output_dir}/brainstorm/all_conclusions.md\n\nProvide a meta-review covering:\n1. **Coverage analysis** — Which aspects of the research space are well-covered vs. underexplored?\n2. **Quality assessment** — Rate each persona's conclusions (depth, rigor, creativity) on a 1-10 scale\n3. **Cross-persona synthesis** — What emerges when combining all perspectives that no single persona captured?\n4. **Top 3 disagreements** — Identify the 3 most significant points where personas contradict each other, with specific quotes\n5. **Recommended directions** — Your top 5 research directions considering all perspectives",
  model: "gemini-3.1-pro-preview"  // fallback chain applies
)

Save to brainstorm/meta_review_gemini.md.

Codex Meta-Review:

mcp__codex-cli__ask-codex(
  prompt: "You are reviewing the outputs of {N} domain-specialist research personas who independently analyzed: {topic}\n\nHere are all persona conclusions:\n@{output_dir}/brainstorm/all_conclusions.md\n\nProvide a meta-review covering:\n1. **Coverage analysis** — Which aspects of the research space are well-covered vs. underexplored?\n2. **Quality assessment** — Rate each persona's conclusions (depth, rigor, creativity) on a 1-10 scale\n3. **Cross-persona synthesis** — What emerges when combining all perspectives that no single persona captured?\n4. **Top 3 disagreements** — Identify the 3 most significant points where personas contradict each other, with specific quotes\n5. **Recommended directions** — Your top 5 research directions considering all perspectives"
)

Save to brainstorm/meta_review_codex.md.

If --claude-only: Replace both meta-review calls above with two Agent subagents, executed simultaneously:

Subagent A (Creative-Divergent, replaces Gemini Meta-Review):

Agent(
  subagent_type: "general-purpose",
  prompt: "You are a Creative-Divergent meta-reviewer. Look for emergent patterns, underexplored spaces, and unconventional combinations across perspectives.

Use the Read tool to read: {output_dir}/brainstorm/all_conclusions.md

You are reviewing the outputs of {N} domain-specialist research personas who independently analyzed: {topic}

Provide a meta-review covering:
1. **Coverage analysis** — Which aspects are well-covered vs. underexplored?
2. **Quality assessment** — Rate each persona's conclusions (depth, rigor, creativity) on a 1-10 scale
3. **Cross-persona synthesis** — What emerges when combining all perspectives?
4. **Top 3 disagreements** — Identify the 3 most significant contradictions with specific quotes
5. **Recommended directions** — Your top 5 research directions

Save to {output_dir}/brainstorm/meta_review_gemini.md. Start with:
> Source: Claude Agent subagent (claude-only mode, Creative-Divergent)"
)

Subagent B (Analytical-Convergent, replaces Codex Meta-Review):

Agent(
  subagent_type: "general-purpose",
  prompt: "You are an Analytical-Convergent meta-reviewer. Focus on methodological rigor, feasibility, and practical implementation paths across perspectives.

Use the Read tool to read: {output_dir}/brainstorm/all_conclusions.md

You are reviewing the outputs of {N} domain-specialist research personas who independently analyzed: {topic}

Provide a meta-review covering:
1. **Coverage analysis** — Which aspects are well-covered vs. underexplored?
2. **Quality assessment** — Rate each persona's conclusions (depth, rigor, creativity) on a 1-10 scale
3. **Cross-persona synthesis** — What emerges when combining all perspectives?
4. **Top 3 disagreements** — Identify the 3 most significant contradictions with specific quotes
5. **Recommended directions** — Your top 5 research directions

Save to {output_dir}/brainstorm/meta_review_codex.md. Start with:
> Source: Claude Agent subagent (claude-only mode, Analytical-Convergent)"
)

Phase B — Disagreement Extraction:

Claude reads both meta-reviews and extracts the top 3 cross-persona disagreements — prioritizing disagreements identified by both reviewers. For each disagreement, produce a structured summary: the claim, which personas support each side, and the core tension. Save the meta-disagreement summary to brainstorm/meta_disagreements.md before the debate calls.

Phase C — Consolidate Debate Context + Adversarial Debate:

Before the debate calls, create consolidated context files (each containing exactly what the opposing model needs):

• brainstorm/debate_context_for_gemini.md — concatenate meta_disagreements.md + meta_review_codex.md
• brainstorm/debate_context_for_codex.md — concatenate meta_disagreements.md + meta_review_gemini.md

Then execute the debate calls simultaneously:

mcp__gemini-cli__ask-gemini(
  prompt: "[Meta-Reviewer]\n\nYou reviewed {N} persona conclusions and identified top disagreements. Below is the disagreement summary followed by Codex's meta-review for context:\n\n@{output_dir}/brainstorm/debate_context_for_gemini.md\n\nFor each disagreement:\n1. **Defend** your position with additional evidence or reasoning\n2. **Concede** if the opposing argument is stronger, explaining why\n3. **Revise** your assessment to a new synthesized position if appropriate",
  model: "gemini-3.1-pro-preview"  // fallback chain applies
)

Save to brainstorm/meta_debate_gemini.md.

mcp__codex-cli__ask-codex(
  prompt: "[Meta-Reviewer]\n\nYou reviewed {N} persona conclusions and identified top disagreements. Below is the disagreement summary followed by Gemini's meta-review for context:\n\n@{output_dir}/brainstorm/debate_context_for_codex.md\n\nFor each disagreement:\n1. **Defend** your position with additional evidence or reasoning\n2. **Concede** if the opposing argument is stronger, explaining why\n3. **Revise** your assessment to a new synthesized position if appropriate"
)

Save to brainstorm/meta_debate_codex.md.

If --claude-only: Replace both debate calls above with two Agent subagents, executed simultaneously:

Subagent A (Creative-Divergent, replaces Gemini Debate):

Agent(
  subagent_type: "general-purpose",
  prompt: "You are a Creative-Divergent debater. Defend unconventional positions with imaginative reasoning, but concede when evidence demands it.

[Meta-Reviewer]

Use the Read tool to read: {output_dir}/brainstorm/debate_context_for_gemini.md

You reviewed {N} persona conclusions and identified top disagreements. The file contains the disagreement summary followed by the opposing meta-review.

For each disagreement:
1. **Defend** your position with additional evidence or reasoning
2. **Concede** if the opposing argument is stronger, explaining why
3. **Revise** your assessment to a new synthesized position if appropriate

Save to {output_dir}/brainstorm/meta_debate_gemini.md. Start with:
> Source: Claude Agent subagent (claude-only mode, Creative-Divergent)"
)

Subagent B (Analytical-Convergent, replaces Codex Debate):

Agent(
  subagent_type: "general-purpose",
  prompt: "You are an Analytical-Convergent debater. Defend positions with evidence and practical reasoning, but concede when creative arguments reveal genuine blind spots.

[Meta-Reviewer]

Use the Read tool to read: {output_dir}/brainstorm/debate_context_for_codex.md

You reviewed {N} persona conclusions and identified top disagreements. The file contains the disagreement summary followed by the opposing meta-review.

For each disagreement:
1. **Defend** your position with additional evidence or reasoning
2. **Concede** if the opposing argument is stronger, explaining why
3. **Revise** your assessment to a new synthesized position if appropriate

Save to {output_dir}/brainstorm/meta_debate_codex.md. Start with:
> Source: Claude Agent subagent (claude-only mode, Analytical-Convergent)"
)

Step 1-max-d: Layer 3 — Final Enriched Synthesis (--depth max)

1. Read all available documents:
   • weights.json, personas.md
   • All N persona_{i}/conclusion.md files
   • meta_review_gemini.md, meta_review_codex.md
   • meta_debate_gemini.md, meta_debate_codex.md
2. Load weights.json and extract the "weights" object. Compute weighted scores for each research direction (same 0-10 rating per dimension, weighted sum).
3. Produce an enriched brainstorm/synthesis.md with the following structure:
   1. Personas Used — table of N personas with name, expertise summary, and primary lens
   2. Scoring Weights — the weights used for ranking (from weights.json), including _meta.method to document how weights were selected (explicit, adaptive-recommended, domain-default, or custom)
   3. Top Research Directions — ranked by weighted score, with:
      • Score breakdown per dimension
      • Which personas supported this direction
      • Key arguments for and against
   4. Cross-Persona Consensus — ideas where 3+ personas independently converged
   5. Unique Contributions — valuable ideas that only a single persona identified
   6. Debate Resolution — for each of the 3 debated disagreements: the original tension, how each meta-reviewer responded (defend/concede/revise), and the synthesized resolution
   7. Emergent Insights — patterns or connections visible only from the cross-persona vantage point that no individual persona captured
   8. Recommended Path Forward — Claude's top recommendation with reasoning grounded in the multi-persona analysis
   9. MAGI Process Traceability — table mapping each conclusion to its source persona, layer, and artifact file path
4. Save to brainstorm/synthesis.md.

Step 1c: Claude Synthesis

If --depth max: Skip — synthesis is produced by Step 1-max-d above.

1. Read all available documents:
   • Always: gemini_ideas.md, codex_ideas.md
   • If --depth medium or high: gemini_review_of_codex.md, codex_review_of_gemini.md
   • If --depth high: debate_round2_gemini.md, debate_round2_codex.md
   • Always: weights.json, personas.md
2. Load weights.json and extract the "weights" object. Use the weights to compute a weighted score for each research direction:
   • For each candidate direction, rate it on each weight dimension (0-10 scale)
   • Compute the weighted sum: score = Σ(weight_i × rating_i)
   • Rank directions by weighted score
3. Synthesize into a coherent research direction document that includes:
   • Personas Used — brief summary of assigned Gemini and Codex personas
   • Scoring Weights — the weights used for ranking (from weights.json), including _meta.method to document how weights were selected
   • Top Research Directions (ranked by weighted score, showing score breakdown)
   • Key Technical Approaches for each direction
   • Consensus Points — ideas both models agreed on
   • Divergence Points — areas of disagreement and how to resolve them
   • Debate Resolution (--depth high only) — for each of the 3 debated disagreements, document the final resolution: who conceded, what was revised, and the synthesized position
   • Recommended Path Forward — Claude's recommendation with reasoning
4. Save to brainstorm/synthesis.md.

Step 2: User Feedback

Present the synthesis to the user with:

• A concise summary of the top 3-5 research directions
• Clear options for the user to choose, modify, or combine directions
• Ask for any additional constraints or preferences

Wait for user input before proceeding.

Output Files

--depth low|medium|high:

brainstorm/
├── weights.json                  # Scoring weights + selection metadata (always)
├── personas.md                   # Assigned expert personas (always)
├── gemini_ideas.md               # Gemini brainstorm output (always)
├── codex_ideas.md                # Codex brainstorm output (always)
├── gemini_review_of_codex.md     # Cross-review (--depth medium|high)
├── codex_review_of_gemini.md     # Cross-review (--depth medium|high)
├── disagreements.md              # Disagreement summary for debate (--depth high only)
├── debate_round2_gemini.md       # Adversarial debate (--depth high only)
├── debate_round2_codex.md        # Adversarial debate (--depth high only)
└── synthesis.md                  # Claude synthesis (always)

--depth max:

brainstorm/
├── weights.json                  # Scoring weights + selection metadata
├── personas.md                   # N domain-specialist personas
├── persona_1/                    # Persona 1 mini-MAGI output
│   ├── gemini_ideas.md
│   ├── codex_ideas.md
│   ├── gemini_review_of_codex.md
│   ├── codex_review_of_gemini.md
│   └── conclusion.md
├── persona_2/                    # Persona 2 mini-MAGI output
│   └── ...                       # (same 5 files)
├── persona_N/                    # Persona N mini-MAGI output
│   └── ...
├── all_conclusions.md            # Consolidated persona conclusions for Layer 2
├── meta_review_gemini.md         # Gemini meta-review of all conclusions
├── meta_review_codex.md          # Codex meta-review of all conclusions
├── meta_disagreements.md         # Meta-disagreement summary for debate
├── debate_context_for_gemini.md  # Consolidated: disagreements + Codex meta-review
├── debate_context_for_codex.md   # Consolidated: disagreements + Gemini meta-review
├── meta_debate_gemini.md         # Adversarial debate — Gemini
├── meta_debate_codex.md          # Adversarial debate — Codex
└── synthesis.md                  # Enriched final synthesis