Brain storming strategy tailored for scientific discovery

sci-brainstorm.md

name	sci-brainstorm
description	Use when exploring a research idea, starting from a paper, concept, or question — conducts deep literature survey via arxiv and web search, discovers cross-field connections, then challenges the idea through Socratic and adversarial discussion

Scientific Research Brainstorming

Research-first brainstorming. Every phase begins with autonomous search — never ask the human what the literature says; go find out. Staged posture: Socratic while the idea is forming, adversarial once it has shape. Produces a research direction document.

Entry

digraph {
    "Paper / link?" [shape=diamond];
    "Vague interest?" [shape=diamond];
    "Extract topic" [shape=box];
    "Ask one clarifying\nquestion" [shape=box];
    "Phase 1: Survey" [shape=box];
    "Phase 3: Crystallize" [shape=box];

    "Paper / link?" -> "Extract topic" [label="yes"];
    "Paper / link?" -> "Vague interest?" [label="no"];
    "Extract topic" -> "Phase 1: Survey";
    "Vague interest?" -> "Ask one clarifying\nquestion" [label="yes"];
    "Vague interest?" -> "Phase 3: Crystallize" [label="no, well-formed question"];
    "Ask one clarifying\nquestion" -> "Phase 1: Survey";
}

Process

Run phases sequentially. Search autonomously at the start of each phase. Show findings before asking questions.

One question at a time. Never ask multiple questions in one message.

Announce posture shifts. Tell the human when switching from Socratic to adversarial.

digraph sci_brainstorm {
    "User states interest" [shape=box];
    "Phase 1: Survey" [shape=box];
    "Phase 2: Expand" [shape=box];
    "Phase 3: Crystallize\n(Socratic)" [shape=box];
    "Phase 4: Stress-test\n(Adversarial)" [shape=box];
    "Idea survives?" [shape=diamond];
    "Phase 5: Refine" [shape=box];
    "Research Direction Doc" [shape=doublecircle];

    "User states interest" -> "Phase 1: Survey";
    "Phase 1: Survey" -> "Phase 2: Expand";
    "Phase 2: Expand" -> "Phase 3: Crystallize\n(Socratic)";
    "Phase 3: Crystallize\n(Socratic)" -> "Phase 4: Stress-test\n(Adversarial)";
    "Phase 4: Stress-test\n(Adversarial)" -> "Idea survives?";
    "Idea survives?" -> "Phase 3: Crystallize\n(Socratic)" [label="no, pivot"];
    "Idea survives?" -> "Phase 5: Refine" [label="yes"];
    "Phase 5: Refine" -> "Research Direction Doc";
}

Phase 1 — Survey (automated)

Map the landscape before any discussion.

Autonomous research:

1. arxiv MCP — search topic, pull 10-15 recent papers (last 2-3 years), read abstracts
2. paper-search-mcp — same query across PubMed, bioRxiv, CrossRef for non-CS hits
3. Semantic Scholar MCP — top 5 papers: pull citation graphs, identify clusters and seminal works
4. WebSearch — blog posts, talks, open problem lists

Synthesize: key papers clustered by sub-theme, active groups, open questions, citation graph shape (converging vs diverging).

Ask: "What surprises you here? What did you already know?" — answer calibrates Phase 2.

Phase 2 — Expand (automated)

Push beyond the user's known territory.

Strategy A — Adjacent subfield: From Phase 1 clusters, pick the one user knows least about. arxiv MCP deep search + Semantic Scholar MCP citation chains outward.

Strategy B — Cross-vocabulary: Extract the structural problem (abstract away jargon, e.g., "compressing a high-dimensional transformation" not "LLM attention compression"). paper-search-mcp across all databases + WebSearch non-academic contexts + Semantic Scholar MCP to trace cross-field hits to their home community.

Present: adjacent subfield findings, cross-field matches (same structure, different vocabulary). Ask: "Which connections feel worth exploring?" User picks 1-2 → raw material for Phase 3.

Phase 3 — Crystallize (Socratic)

Help user articulate a specific research angle. Posture: Socratic — only questions, no judgments.

Targeted search: Semantic Scholar MCP — check if angle has been explored. arxiv MCP — find closest existing work.

Socratic questions (one at a time):

• "What specifically is new about combining [X] with [Y]?"
• "What would a successful result look like?"
• "What's the minimal experiment that would tell you this works?"
• "Who would care? Which venue?"

Exit criterion: User states idea in one sentence with clear novelty claim.

Phase 4 — Stress-test (Adversarial)

Try to kill the idea with evidence. Whatever survives is worth pursuing.

Posture shift: Announce — "Now I'm going to challenge this. My job is to find reasons this doesn't work."

Autonomous research (adversarial): Search for prior art via Semantic Scholar MCP (citation chains) + arxiv MCP (novelty claim, negative results) + paper-search-mcp (cross-database) + WebSearch (blog posts, workshop papers).

Challenge on three axes:

Axis	Challenge	Evidence Source
Novelty	"I found [paper X] very similar. How is yours different?"	Semantic Scholar + arxiv
Rigor	"State the core claim as a theorem or testable hypothesis."	Socratic (no tool)
Impact	"If this works perfectly, what improvement? Enough for [venue]?"	WebSearch

Outcomes:

• Survives → Phase 5
• Dies → "Here's what killed it: [evidence]. Here's what's still interesting: [salvageable]. Want to pivot?" → loop to Phase 3

Phase 5 — Refine (collaborative)

Produce a structured research direction document.

Autonomous research (gap-filling):

• Semantic Scholar MCP — full reference list
• arxiv MCP — methodology papers for planned approach
• WebSearch — code repos, datasets, benchmarks

Output: Save to docs/plans/YYYY-MM-DD-<topic>-research-direction.md

Structure (draft each section, show, get feedback):

• Research Question — one sentence
• Novelty Claim — what's new (survived stress-test)
• Key References — from survey + expansion
• Cross-field Connections — unexpected links from Phase 2
• Proposed Approach — method outline
• Minimum Viable Experiment — from Phase 3
• Open Risks — unresolved from stress-test
• Target Venue

Edge Cases

Situation	Handling
User already has a well-formed research question	Skip to Phase 3
Survey reveals idea is already published	Present prior art, ask if user sees a different angle
No cross-field connections found	Proceed with within-field survey
MCP tool unavailable	Fall back to WebSearch only
User disagrees with adversarial challenge	Present evidence, let user decide

Guardrails

• Never fabricate citations — only present what tools actually found.
• Never assert novelty judgments — present evidence, let user evaluate.
• Always preserve pivot path — show what's salvageable when stress-test kills an idea.
• Cite sources inline — every literature claim includes paper title or URL.