interview-me（采访我）

Research Interview

Conduct a structured interview to help formalize a research idea into a concrete specification.

Input: $ARGUMENTS — a brief topic description or "start fresh" for an open-ended exploration.

How This Works

This is a conversational skill. Instead of producing a report immediately, you conduct an interview by asking questions one at a time, probing deeper based on answers, and building toward a structured research specification.

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

Interview Structure

Phase 1: The Big Picture (1-2 questions)

• "What phenomenon or puzzle are you trying to understand?"
• "Why does this matter? Who should care about the answer?"
• After the user answers, optionally ask: "Do you have a sense of what *kind* of paper this would be — reduced-form / structural / theory+empirics / descriptive / formal-theory / survey-experiment / unsure?" (See .claude/agents/methods-referee.md for the type definitions and .claude/references/discipline-cards.md for field-default frequencies.) Record the answer in the saved spec under the **Paper type:** header field; "unsure" is fine and is recorded as **Paper type:** unsure.

Phase 2: Theoretical Motivation (1-2 questions)

• "What's your intuition for why X happens / what drives Y?"
• "What would standard theory predict? Do you expect something different?"

Phase 3: Data and Setting (1-2 questions)

• "What data do you have access to, or what data would you ideally want?"
• "Is there a specific context, time period, or institutional setting you're focused on?"

Phase 4: Identification (1-2 questions)

• "Is there a natural experiment, policy change, or source of variation you can exploit?"
• "What's the biggest threat to a causal interpretation?"

Phase 5: Expected Results (1-2 questions)

• "What would you expect to find? What would surprise you?"
• "What would the results imply for policy or theory?"

Phase 6: Contribution (1 question)

• "How does this differ from what's already been done? What's the gap you're filling?"

After the Interview

Once you have enough information (typically 5-8 exchanges), produce a Research Specification Document:

# Research Specification: [Title]

**Date:** [YYYY-MM-DD]
**Researcher:** [from conversation context]
**Paper type:** [reduced-form | structural | theory+empirics | descriptive | formal-theory | survey-experiment | unsure]

## Research Question

[Clear, specific question in one sentence]

## Motivation

[2-3 paragraphs: why this matters, theoretical context, policy relevance]

## Hypothesis

[Testable prediction with expected direction]

## Empirical Strategy

- **Method:** [e.g., Difference-in-Differences with staggered adoption]
- **Treatment:** [What varies]
- **Control:** [Comparison group]
- **Key identifying assumption:** [What must hold]
- **Robustness checks:** [Pre-trends, placebo tests, etc.]

## Data

- **Primary dataset:** [Name, source, coverage]
- **Key variables:** [Treatment, outcome, controls]
- **Sample:** [Unit of observation, time period, N]

## Expected Results

[What the researcher expects to find and why]

## Contribution

[How this advances the literature — 2-3 sentences]

## Open Questions

[Issues raised during the interview that need further thought]

Save to: quality_reports/research_spec_[sanitized_topic].md

Post-Flight Verification (mandatory, CoVe — applies when the spec cites prior work)

The research spec's Motivation and Contribution sections typically reference prior papers by author + year. Those citations are hallucination-prone. Before saving the spec, run the Post-Flight Verification protocol from .claude/rules/post-flight-verification.md if the spec contains any citations.

Steps (skip if the spec cites zero papers)

1. Extract claims: every paper-citation in the Motivation / Contribution sections ("Smith 2019 shows X"), any dataset-structure claims ("the CPS has field educ_attain"), any negative-literature assertions ("nobody has studied Y").
2. Generate verification questions: specific, answerable questions per claim. "Does Smith (2019, *JEL*) Section 3 report finding X? Is the venue correct?"
3. Spawn claim-verifier via Task with subagent_type=claim-verifier and context=fork. Hand it the claims + questions + source pointers (DOIs, arXiv links, master_supporting_docs/ PDFs if the user provided any during the interview). Do NOT include the drafted spec.
4. Reconcile: PASS → attach green block to the spec. PARTIAL → mark unverifiable citations with uncertainty flags. FAIL → rewrite the affected paragraph using the verifier's evidence before saving the spec.

Skip conditions

• Spec contains zero paper citations (pure-methodology specs with no lit references).
• --no-verify flag.
• The user explicitly said during the interview "I'll verify the literature myself."

Decision records (when tradeoffs surface)

If during the interview the researcher explicitly chose among alternatives — identification strategy (DiD vs IV vs RDD), data source (admin vs survey), outcome measure, sample scope, etc. — also write an ADR-style decision record for each choice. Use templates/decision-record.md and save to quality_reports/decisions/YYYY-MM-DD_[short-topic].md. Required fields: Status / Problem / Options considered / Decision + rationale / Consequences / Rejected alternatives.

Skip the ADR if the interview produced a single uncontested direction — ADRs are for *decisions with live alternatives*, not for announcing the default path.

Interview Style

• Be curious, not prescriptive. Your job is to draw out the researcher's thinking, not impose your own ideas.
• Probe weak spots gently. If the identification strategy sounds fragile, ask "What would a skeptic say about...?" rather than "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. Don't over-interview.