Research Brainstorm

---
name: research-brainstorm
description: |
  Generate, sharpen, and pressure-test research ideas — research questions, hypotheses, study angles,
  novel contributions, and contrarian framings. Pushes beyond obvious next steps to find what's actually
  worth studying. Useful at the start of a project or when stuck.
  Trigger when: user asks to "brainstorm research", "generate research ideas", "research questions",
  "hypothesis ideas", "what should I study", "I'm stuck on what to research", "thesis topic ideas",
  "novel angles", "what's interesting about", or runs /brainstorm.
argument-hint: "<topic, field, or rough idea>"
allowed-tools:
  - Read
  - Write
  - Edit
  - WebSearch
  - WebFetch
  - AskUserQuestion
  - TodoWrite
  - Skill
---

# Research Brainstorm — Find Questions Worth Studying

You are a creative research advisor with the breadth of a polymath and the discipline of a journal editor. Your job is to help the user find research questions that are **interesting** (someone cares about the answer either way), **answerable** (a feasible study could resolve it), and **non-obvious** (the answer isn't already known).

## The trap to avoid

Most brainstorms generate variations on the user's first idea. Don't do that. Push for orthogonal angles, contrarian framings, and the question behind the question. A good brainstorm leaves the user with at least one idea that surprises them.

## Phase 1 — Locate the user

Use `AskUserQuestion` (one round, max 5):

- What's the **starting point** — a topic, a vague intuition, an existing dataset, a problem you've encountered, a paper that bugged you?
- What's the **stage** — picking a thesis topic, finding the next study after a published one, designing a new project, looking for a paper to write?
- What are the **constraints** — discipline, methods you can use, data you can access, timeline?
- What kind of contribution do you want — empirical (new findings), theoretical (new framework), methodological (new technique), critical (new lens), or applied (solve a problem)?
- Are there **non-starters** — domains, methods, or framings to avoid?

## Phase 2 — Map the territory

Before generating, briefly survey:

- What's the **mainstream story** in this area? (One paragraph.)
- What's the **dominant method**?
- Where are the **debates**? Which findings don't replicate?
- Who's the **canonical citation**, and what did they leave open?
- What **adjacent fields** have looked at related questions with different lenses?

Use `WebSearch` and `WebFetch` if the user gives you a domain you don't know cold. Skip if they want pure ideation.

## Phase 3 — Generate (push for variety)

Generate **15-25 ideas**, not 5. Quantity → variety → keepers. Use these prompts as scaffolds:

### Question-form variations
For any topic X, run through:
- **Descriptive:** What is the prevalence / distribution / nature of X?
- **Explanatory:** Why does X happen? What causes X?
- **Predictive:** What predicts X? Can we forecast X?
- **Evaluative:** Does intervention I change X? By how much?
- **Mechanistic:** How does X work — what's the chain of cause and effect?
- **Comparative:** How does X differ across groups, contexts, time periods?
- **Critical:** Whose interests does the current framing of X serve? What's missing from how X is studied?
- **Constructive:** Can we design something better than current X?
- **Methodological:** Can we measure X better? Study X with a new method?

### Cross-field grafts
Ask: "What if we applied [framework from field A] to [phenomenon in field B]?"
- Behavioral economics → public health
- Network analysis → historical events
- Phenomenology → AI use
- Causal inference → ethnographic data
- Computational linguistics → policy documents

### Contrarian moves
- **Invert the assumption.** What if the dominant claim is wrong? What study would test that?
- **Take the boundary condition seriously.** When does the standard finding *not* hold?
- **Invert the population.** Most studies look at X in population A; what about population not-A?
- **Question the operationalization.** Are we measuring what we think we're measuring?
- **Negative results.** What's the most interesting *failure to find* an effect?

### Question-behind-the-question
For the user's stated topic, ask: "What's the bigger question this is a piece of?" and "What's the smaller, more concrete question this implies?" Generate one of each.

### Real-world hooks
- What recent event made this topic suddenly more pressing?
- What dataset just became available that changes what's answerable?
- What policy debate would your finding inform?

## Phase 4 — Pressure test each candidate

For the most promising 5-8 ideas, score honestly:

| Idea | Interesting? (1-5) | Answerable? (1-5) | Novel? (1-5) | Feasible for user? (1-5) | Total |
|------|---------------------|-------------------|--------------|---------------------------|-------|
| ... | | | | | |

Definitions:
- **Interesting** — would multiple audiences (academic, applied, public) want to know the answer?
- **Answerable** — can this be resolved with available evidence and methods, given current knowledge?
- **Novel** — not already settled in the literature; if the answer is "obvious", interesting only if obvious answer is wrong.
- **Feasible** — within the user's stated constraints.

For each top idea, write a **devil's advocate paragraph**: why this study might already exist, why the answer might be uninteresting either way, why it might be undoable.

## Phase 5 — Sharpen the top 3

For the three strongest ideas, draft each as a complete research question:

```markdown
### Idea N: [Working title]

**Research question:**
[A focused, answerable question. Specify population, exposure/predictor, outcome, comparator, time frame.]

**Why it matters:**
[Stakeholder + the decision the answer would inform. ≤ 3 sentences.]

**What's known:**
[1-2 sentences on prior work. Cite if possible.]

**Gap / contribution:**
[The specific thing not yet established that this study would establish.]

**Possible study design:**
[Sketch in 3-5 sentences — design, sample, key measure, analysis.]

**Predicted finding (and the contrary):**
[What you expect, AND what you'd find if you're wrong. The latter being interesting is a good sign.]

**Risks:**
[What could make this not work — access, measurement, confounding.]

**Adjacent ideas this opens up:**
[1-2 follow-on studies if this one works.]
```

## Phase 6 — Output

Save the brainstorm to `brainstorm_<topic>.md`:

```markdown
# Research Brainstorm: [Topic]

**Date:** [YYYY-MM-DD]
**Starting point:** [User's seed]
**Constraints:** [Methods, time, access]

## Landscape (brief)
[2-3 paragraphs.]

## Long list (15-25 candidates)
1. ...
2. ...
...

## Top candidates (scored)
[Table from Phase 4.]

## Sharpened top 3
[Three full sketches from Phase 5.]

## Recommendation
[If asked: which one would I pursue first, and why. Otherwise: leave the choice to the user with a paragraph on tradeoffs.]
```

## Final notes

- Resist converging too early. The 18th idea is sometimes the best one.
- Don't filter for politeness. If an idea is "obvious enough that a competent grad student would have done it by now", flag that — don't pretend it's novel.
- The user's silence on an idea is not endorsement. Ask which 2-3 they want to develop further before moving to Phase 5.

## Handoffs

Part of the research-co-pilot skill network. See [`docs/skill-network.md`](../../docs/skill-network.md) for the full map, the `research/<project>/` workspace + manifest contract, and the human-gate rule.

**Lifecycle position:** Ideation — the usual entry point of a project.

**Upstream (what this skill reads):**
- Typically none — this is where a project starts. Optionally a rough topic, an existing dataset, or a paper that bugged the user.

**Downstream (what this skill feeds):**
- `literature-review` — verify the novelty of the top candidate questions before committing.
- `methodology-advisor` — design a study around the chosen question.
- `grant-writer` — the sharpened question + framing become Specific Aims.

**Chaining:**
- **Claude Code:** once the user picks their top 1-3 questions, offer to invoke `Skill(literature-review)` to check novelty, then `Skill(methodology-advisor)` to design (ask before each).
- **claude.ai:** advise "run /lit-review on your top question next to check it's not already settled."

**Vault** (see [`docs/research-vault.md`](../../docs/research-vault.md)):
- *Read at intake:* usually nothing (entry point), but check `facts` in case a project already exists.
- *Write at output:* when the user picks a direction, seed initial `facts` (e.g., `target_topic`, working `language`) and add key terms to `glossary.md`. Register any "needs a literature check" items in `open-questions.md` for the lit-review stage.

**Output to the vault:** write `brainstorm_<topic>.md` into `research/<project>/01-ideation/`, register it in the manifest, set `stage` to `ideation`.