brycewang-stanford/discover

Discover

Launch the Discovery phase of research. Routes to the appropriate agents based on the mode specified.

Input: $ARGUMENTS — a mode keyword followed by a topic or query.

---

Modes

Default (no mode specified)

If no mode keyword is given, start with an interactive interview to build the research specification.

/discover interview [topic] — Research Interview

Conduct a structured conversational interview to formalize a research idea.

This is conversational. Ask questions directly in your text responses, one or two at a time. Wait for the user to respond before continuing. Do NOT use AskUserQuestion.

Agents: Direct conversation (no agent dispatch) Output: Research specification + domain profile

Interview structure:

1. Big Picture (1-2 questions): "What phenomenon are you trying to understand?" "Why does this matter?"
2. Theoretical Motivation (1-2 questions): "What's your intuition for why X happens?" "What would standard theory predict?"
3. Data and Setting (1-2 questions): "What data do you have access to?" "Is there a specific institutional setting?"
4. Identification (1-2 questions): "Is there a natural experiment or policy change you can exploit?" "What's the biggest threat to causal interpretation?"
5. Expected Results (1-2 questions): "What would you expect to find?" "What would surprise you?"
6. Contribution (1 question): "How does this differ from what's been done? What gap are you filling?"

Interview style:

• Be curious, not prescriptive. Draw out the researcher's thinking, don't impose your own ideas.
• Probe weak spots gently. "What would a skeptic say about...?" not "This won't work because..."
• Build on answers. Each question should follow from the previous response.
• Know when to stop. If the researcher has a clear vision after 4-5 exchanges, move to the specification.

After interview (5-8 exchanges), produce:

Output 1: Research Specification → quality_reports/research_spec_[topic].md

# Research Specification: [Title]
## Research Question — [one sentence]
## Motivation — [why this matters, theoretical context, policy relevance]
## Hypothesis — [testable prediction with expected direction]
## Empirical Strategy — [method, treatment, control, identifying assumption, robustness]
## Data — [primary dataset, key variables, sample, unit of observation]
## Expected Results — [what the researcher expects and why]
## Contribution — [how this advances the literature]
## Open Questions — [issues needing further thought]

Output 2: Domain Profile → .claude/references/domain-profile.md (if still template) Fill in field, target journals, common data sources, identification strategies, field conventions, seminal references, and referee concerns based on the interview.

/discover lit [topic] — Literature Review

Search and synthesize academic literature.

Agents: Librarian (collector) → librarian-critic (reviewer) Output: Annotated bibliography + BibTeX entries + frontier map

Workflow:

1. Read .claude/references/domain-profile.md for field journals and seminal references
2. Check master_supporting_docs/ for uploaded papers
3. Read bibliography_base.bib for papers already in the project
4. Dispatch Librarian to search:
   • Top-5 journals (AER, Econometrica, QJE, JPE, REStud)
   • Field journals from domain-profile.md
   • NBER/SSRN/IZA working papers
   • Citation chains — forward and backward citation tracking from key papers. Follow: (a) backward citations (what do the key papers cite?), and (b) forward citations (who cites the key papers?). This is often the most productive search vector.
5. Assign proximity scores to each paper:
   • 1 — Directly competes (same question, similar method)
   • 2 — Closely related (same question, different method or setting)
   • 3 — Related (overlapping topic, different angle)
   • 4 — Background (provides theory, method, or context)
   • 5 — Tangentially related (useful framing only)
6. Dispatch librarian-critic to check coverage, gaps, recency, scope
7. If gaps found, re-dispatch Librarian for targeted search (max 1 round)
8. Save to quality_reports/lit_review_[topic].md

Unverified citations: If you cannot verify a citation, mark the BibTeX entry with % UNVERIFIED. Do NOT fabricate or guess citation details. Note when working papers have been published — cite the published version.

Output format for each paper:

### [Author (Year)] — [Short Title]
- **Journal:** [venue]
- **Proximity:** [1-5 score]
- **Main contribution:** [1-2 sentences]
- **Identification strategy:** [DiD / IV / RDD / SC / descriptive]
- **Key finding:** [result with effect size]
- **Relevance:** [why it matters for our research]

/discover data [requirements] — Data Discovery

Find and assess datasets for the research question.

Agents: Explorer (finder) → explorer-critic (assessor) Output: Ranked data sources with feasibility grades

Workflow:

1. Read research spec and strategy memo if they exist
2. Read .claude/references/domain-profile.md for common data sources in the field
3. Understand what variables are needed: treatment, outcome, controls, time period, geography
4. Dispatch Explorer to search across source categories:
   • Public microdata (CPS, ACS, NHIS, MEPS, etc.)
   • Administrative data (Medicare claims, tax records, court records)
   • Survey data (RAND HRS, PSID, Add Health, NLSY)
   • International (World Bank, OECD, Eurostat)
   • Novel/alternative (satellite imagery, web scraping, proprietary)
5. For each dataset found, report:
   • Name, provider, access level (public/restricted)
   • Key variables available
   • Coverage (time period, geography, sample size)
   • Feasibility grade:
     • A — Ready to use (public download, documented, standard format)
     • B — Accessible with effort (application required, moderate cost, needs cleaning)
     • C — Restricted but obtainable (FSRDC, data use agreement, IRB approval)
     • D — Very difficult (proprietary, requires partnership, rare access)
   • Strengths and limitations
6. Dispatch explorer-critic to critique each proposed dataset using the 5-point assessment:
   1. Measurement validity — Does the variable actually measure what we need?
   2. Sample selection — Who is in the data? Who is missing?
   3. External validity — Can we generalize from this sample?
   4. Identification compatibility — Does this data support the proposed design?
   5. Known issues — Documented problems with this dataset in the literature
7. Save exploration to quality_reports/data_exploration_[topic].md

Rejected datasets: Include a rejection table:

| Dataset | Reason for Rejection | Deal-breaker? | |---------|---------------------|---------------| | [Name] | [explorer-critic's finding] | [Yes/No] |

/discover ideate [topic] — Research Ideation

Generate structured research questions and hypotheses from a topic or dataset.

Agents: Direct generation (no agent dispatch) Output: Research questions with empirical strategies

Generate:

1. 3-5 research questions with clear hypotheses
2. For each: potential identification strategy, data requirements, expected contribution
3. Rank by feasibility and novelty
4. Save to quality_reports/research_ideas_[topic].md

---

Principles

• Interview style: Be curious, not prescriptive. Draw out the researcher's thinking.
• Literature honesty: Never fabricate citations. Mark unverified as % UNVERIFIED.
• Proximity scoring: Always assign 1-5 proximity scores to papers found.
• Citation chains: Forward and backward citation tracking is an explicit search vector — do not skip it.
• Effect sizes matter: Report magnitudes, not just signs. Note identification strategy for every paper.
• Data feasibility matters: A perfect dataset you can't access is useless. Always assign A/B/C/D grades.
• 5-point data critique: Measurement validity, sample selection, external validity, identification compatibility, known issues. Never skip this.
• Domain-profile aware: Always read .claude/references/domain-profile.md first for field calibration.
• Worker-critic pairing: Librarian + librarian-critic, Explorer + explorer-critic. Never skip the critic.