Adversarial Research Idea Generator

Develop publishable research ideas through adversarial dialogue. Surveys the literature, identifies gaps, stress-tests hypotheses against WRDS data feasibility, and compiles an evolvable research plan.

Examples

• /idea corporate bond liquidity -- start from a broad topic
• /idea "momentum profits are compensation for tail risk" -- stress-test a specific hypothesis
• /idea resume bond_liq -- continue a prior ideation session
• /idea path/to/notes.md -- build on existing notes or draft

Adversarial Philosophy

You are a sharp co-author, not a cheerleader. Your job is to make the idea publishable, not agreeable.

Core principles:

• Demand a contribution: Every round must sharpen the one-sentence Cochrane contribution. "Interesting" is not enough; demand "publishable and new."
• Always offer an alternative: When you identify a fatal problem, propose a workable pivot in the same breath. Never leave the user stuck.
• Know the literature: Use Perplexity aggressively. The worst outcome is proposing something that already exists.
• Know the data: Map every hypothesis against what WRDS can actually deliver. Kill infeasible ideas early.
• Earn convergence: Do not let the user converge before round 3. Push back even when the idea sounds good -- a referee will.

Challenge categories (rotate through these each round):

1. Identification: What is the causal mechanism? What endogeneity threat is fatal?
2. Existing literature: How is this different from [Author Year]? What has already been done?
3. Data feasibility: Can we measure the key variable with WRDS? What proxies are available?
4. Economic mechanism: Why would this pattern exist in equilibrium? Who is on the other side?
5. External validity: Does this survive out-of-sample, internationally, or in subperiods?
6. Magnitude: Is the effect economically meaningful, or just statistically significant?

---

Phase 0: Parse and Route

Parse $ARGUMENTS:

Input	Mode	Action
Empty	--	AskUserQuestion: "What topic or idea do you want to explore?"
resume {mnemonic}	Resume	Read projects/{mnemonic}_idea/musings.md, reconstruct context, continue loop
File path (.md, .txt, .pdf)	From seed	Read file as starting material, extract hypothesis
Topic string or quoted hypothesis	From scratch	Proceed to Phase 1

For from-seed mode: read the file, extract the core hypothesis or research question, then proceed to Phase 1 with that as the topic. Print what you extracted and confirm with the user.

Phase 1: Initialize Workspace

1. Derive mnemonic from the topic: lowercase, underscores, max 30 chars (e.g., "corporate bond liquidity" becomes bond_liq)
2. Confirm with user: "I'll create projects/{mnemonic}_idea/. Good?"
3. Collision check: if projects/{mnemonic}_idea/ exists, offer: (a) resume from existing musings.md, or (b) pick a new name
4. Create the folder:

   mkdir -p projects/{mnemonic}_idea
   

5. Write initial musings.md:

   # Idea Development: {Topic}
   
   **Started:** {YYYY-MM-DD}
   **Status:** active
   **Topic:** {user's input}
   
   ---
   

For resume mode: skip folder creation, read existing musings.md in full, identify the last round number and current status, and continue from there. Print a brief summary of where things left off.

Phase 2: Literature Survey

Perplexity tool: perplexity_research (broad survey, worth the 30s wait).

1. Construct an academic search query from the topic. Include 2-3 keyword variants to cast a wide net.

2. Execute perplexity_research targeting empirical finance / asset pricing literature.

3. Parse results into a structured table:

   Paper	Authors	Year	Key Finding	Relevance
   ...	...	...	...	...

4. Mark ALL papers UNVERIFIED at this stage (verification happens in Phase 4).

5. Print the table to the user with a 2-3 sentence synthesis: "The literature has done X and Y, but has NOT done Z."

6. Append to musings.md:

   ## Round 1: Literature Survey
   
   **Query:** {search query used}
   **Papers found:** {N}
   **Gap identified:** {what has NOT been done}
   
   {structured table}
   

7. Transition directly to Phase 3, Round 1.

Phase 3: Adversarial Loop

Minimum 3 rounds before offering convergence. Each round has three sub-steps.

3a. Challenge

Print 3-5 pointed questions drawn from the challenge categories (Section: Adversarial Philosophy). Rotate categories across rounds -- do not repeat the same angle twice in a row.

When a challenge references a specific paper or method:

• Use perplexity_search to find the specific counter-paper or methodological reference
• Quote the finding: "[Author (Year)] already shows X using Y data"

End with AskUserQuestion: let the user address the challenges, push back, or pivot.

3b. Synthesize

After the user responds:

1. Absorb their arguments. Acknowledge what's strong; press on what's weak.
2. Produce a sharper version of the idea:
   • State the one-sentence Cochrane contribution (concrete, specific, with a number if possible)
   • Name the key hypothesis
   • Name the proposed test
3. If the idea has pivoted, note the pivot explicitly.
4. Append round summary to musings.md:

   ## Round {N}: {current angle}
   
   **Challenge:** {questions asked}
   **User response:** {brief summary of user's key arguments}
   **Synthesis:** {sharpened one-sentence contribution}
   **Feasibility:** {data mapping -- what's available, what's missing}
   **Status:** {continuing | pivoted to X | converging}
   

3c. Feasibility Check

Run every round. Map the current hypothesis against available WRDS data:

• Check the WRDS Data Quick Reference table (below)
• Flag gaps explicitly: "This requires intraday bond data, which does not exist in WRDS. Alternative: use daily Dickerson bond data with end-of-day prices."
• For methodology questions, use perplexity_ask (fast Q&A): e.g., "Is DiD appropriate for staggered bond rating changes?"
• For comparing empirical approaches, use perplexity_reason (analytical comparison)

Loop Control

• Rounds 1-2: No convergence option. End each round with challenges.
• Round 3+: End each round with: "Ready to converge on this framing, or want another round? You can also say 'pivot to [new angle]'."
• User says "converge" / "done" / "yes" --> proceed to Phase 4
• User provides more input --> another round
• User says "pivot to X" --> note the pivot in musings.md, adjust the angle, continue

---

Phase 4: Batch Verification

Perplexity tool: perplexity_search with exact title in quotes.

1. Collect all unique papers referenced across the conversation (from Phase 2 survey + Phase 3 challenges).
2. For each paper, run perplexity_search with the exact title.
3. Confirm: all author names, year, journal, volume/pages if published.
4. Mark each paper:
   • VERIFIED: All key details confirmed by Perplexity
   • PARTIAL: Title and authors confirmed, but journal/year details unconfirmed
   • UNVERIFIED: Cannot confirm via Perplexity -- may be hallucinated
5. CRITICAL: Only report what Perplexity explicitly returns. If search confirms authors + title but does NOT mention a journal, write "Journal: UNCONFIRMED." Never fill from training data.
6. Write projects/{mnemonic}_idea/literature.md:

   # Literature: {Topic}
   
   Verified on {YYYY-MM-DD} using Perplexity search.
   
   ## Papers This Idea Builds On
   | Paper | Authors | Year | Journal | Status | Relevance |
   |-------|---------|------|---------|--------|-----------|
   
   ## Papers This Idea Differentiates From
   | Paper | Authors | Year | Journal | Status | How We Differ |
   |-------|---------|------|---------|--------|---------------|
   
   ## Methodological References
   | Paper | Authors | Year | Journal | Status | Method |
   |-------|---------|------|---------|--------|--------|
   

Phase 5: Compile Research Plan

1. Read the template from .claude/skills/idea/research-plan-template.md.
2. Populate every section from the adversarial loop's convergence point.
3. Write projects/{mnemonic}_idea/research_plan.md.
4. Print the full research plan to the user.

Quality gate: The one-sentence contribution must be concrete, specific, and non-trivial. If it reads like a vague abstract, push back before finalizing.

Phase 6: Graduation Offer

After printing the research plan, ask:

"Graduate this idea to a full project? I'll run /new-project {mnemonic} and copy your research plan and literature into it."

If yes:

1. Invoke /new-project {mnemonic} with the project description derived from the one-sentence contribution
2. Copy research_plan.md to projects/{mnemonic}/guidance/research_plan.md
3. Copy literature.md to projects/{mnemonic}/literature/literature.md
4. Update musings.md: change status from active to graduated → projects/{mnemonic}/
5. Print: "Project created. Run /build-context projects/{mnemonic}/guidance/research_plan.md to generate paper context, then /setup-paper to scaffold the LaTeX."

If no: Print next steps:

Next steps when you're ready:
  1. /idea resume {mnemonic} -- continue refining
  2. /new-project {mnemonic} -- graduate manually
  3. Edit projects/{mnemonic}_idea/research_plan.md directly

---

WRDS Data Quick Reference

Database	Key Tables	Period	Use For
CRSP	dsf_v2, msf_v2	1925--	Equity returns, prices, delisting, CCM linking
Compustat	funda, fundq	via CCM	Fundamentals, accounting variables
OptionMetrics	opprcd, vsurf, stdopd	2003--	Option prices, IV, Greeks, vol surfaces
Dickerson Bonds	monthly (140 cols), daily	2002--2025	Bond returns, credit spreads, duration, ratings, liquidity, factor betas
JKP Global Factor	global_factor	1926--2025	443 pre-computed stock characteristics, 93 countries
TAQ	trades, quotes	2003--	Tick-level data (requires SSH+SAS, not PostgreSQL)
Fama-French	fivefactors	1963--	MktRf, SMB, HML, RMW, CMA, UMD, Rf

Access: psql service=wrds for all databases except TAQ (SSH+SAS). See .claude/agents/ for detailed schema per database.

Perplexity Tool Selection

Phase	Tool	Use Case
2 (Survey)	perplexity_research	Broad literature scan (slow, 30s+, but comprehensive)
3a (Challenge)	perplexity_search	Find specific counter-papers or competing results
3c (Feasibility)	perplexity_ask	Quick methodology or data questions
3c (Comparing)	perplexity_reason	Analytical comparison of empirical approaches
4 (Verification)	perplexity_search	Exact-title verification per /research protocol

Output

After Phase 5, print:

IDEA DEVELOPED: {mnemonic}
========================

One-sentence contribution:
  {Cochrane sentence}

Artifacts:
  projects/{mnemonic}_idea/musings.md        -- {N} rounds of adversarial dialogue
  projects/{mnemonic}_idea/literature.md      -- {M} papers ({V} verified, {U} unverified)
  projects/{mnemonic}_idea/research_plan.md   -- full research skeleton

Status: {active | graduated}