research-interviewer

What this skill does

# Research Interviewer

A systematic knowledge elicitation system that extracts comprehensive, high-fidelity information through adaptive interviewing. Combines deep empathetic understanding with rigorous validation, ensuring captured knowledge is complete, consistent, and ready for downstream use.

---

## 1. Purpose

This skill provides 12 core capabilities:

| # | Capability | Phase | Description |
|---|------------|-------|-------------|
| 1 | **Establish** | 1 | Set interview goal, scope, success criteria, output format |
| 2 | **Map** | 2 | MECE decomposition of topic into coverage dimensions |
| 3 | **Question** | 3 | Adaptive questioning using 8 question types |
| 4 | **Track** | 3-5 | Continuous confidence tracking with epistemic labels |
| 5 | **Validate** | 5 | Cross-reference consistency checking |
| 6 | **Surface** | 3-5 | Assumption identification (explicit, implicit, structural) |
| 7 | **Protect** | 3-5 | Bias protection via frame equivalence, disconfirmation |
| 8 | **Steelman** | 5 | Present strongest version back for confirmation |
| 9 | **Probe** | 6 | Unknown unknowns sweep before termination |
| 10 | **Calibrate** | 6 | Interviewee confidence calibration |
| 11 | **Synthesize** | 5 | Build unified knowledge artifact |
| 12 | **Output** | 6 | Produce format-appropriate deliverable |

---

## 2. When to Use

**Ideal for:**
- Gathering research requirements before creating a research brief
- Eliciting problem statements from stakeholders
- Extracting domain knowledge for documentation
- Clarifying vague or complex requirements
- Discovering unknown unknowns in a problem space
- Building structured knowledge bases from expert interviews

**Avoid when:**
- Information is already well-documented (read docs instead)
- Simple factual lookup (use search instead)
- Interviewee is unavailable or unresponsive
- Topic is outside interviewee's knowledge domain

---

## Checkpoints

This skill uses interactive checkpoints (see `references/checkpoints.yaml`) to resolve ambiguity:

- **output_format_selection** — When output format not specified
- **validation_mode_selection** — When validation mode not specified and context unclear
- **confidence_threshold_adjustment** — When confidence near threshold and domain suggests different
- **premature_termination_check** — When termination criteria met but warning signs present

---

## 3. Parameters

| Parameter | Type | Required | Default | Description |
|-----------|------|----------|---------|-------------|
| `interview_goal` | string | **yes** | — | What the extracted information will be used for |
| `topic` | string | **yes** | — | What to interview about |
| `output_format` | enum | no | `PROBLEM-STATEMENT` | PROBLEM-STATEMENT \| KNOWLEDGE-CORPUS \| REQUIREMENTS |
| `domain_reference` | enum | no | `none` | product \| architecture \| research \| requirements \| custom \| none |
| `confidence_threshold` | number | no | `0.85` | Target confidence for termination (0.0-1.0) |
| `max_questions` | integer | no | `30` | Maximum questions before forced synthesis |
| `validation_mode` | enum | no | `balanced` | empathetic \| balanced \| rigorous |

### Validation Mode Effects

| Mode | Behavior |
|------|----------|
| **empathetic** | Prioritize rapport, softer probing, accept more at face value |
| **balanced** | Standard verification, targeted probing on inconsistencies |
| **rigorous** | Aggressive assumption challenging, devil's advocate on all claims |

---

## 4. Six-Phase Workflow

### Phase 1: Context Establishment

**Purpose:** Set interview parameters and align on goals.

**Steps:**
1. Receive or elicit `interview_goal` and `topic`
2. Determine `output_format` based on downstream use:
   - PROBLEM-STATEMENT: For research briefs, ideation, evaluation workflows
   - KNOWLEDGE-CORPUS: For RAG systems, documentation, context injection
   - REQUIREMENTS: For development workflows, specification skills

   **CHECKPOINT: output_format_selection**
   - If output_format not specified: **AskUserQuestion**
   - Present format options with downstream implications

3. Select `domain_reference` to load appropriate vocabulary and MECE patterns
4. Establish `validation_mode` based on stakes and interviewee relationship

   **CHECKPOINT: validation_mode_selection**
   - If validation_mode not specified and context unclear: **AskUserQuestion**
   - Present mode options with style implications
5. Confirm parameters with interviewee: "We're aiming to [goal]. I'll ask questions about [topic] and produce a [format]. Does that work?"
6. Initialize empty Knowledge Map structure

**Quality Gate:** Goal clarity - interview_goal must be specific, actionable, and measurable

**Output:** Interview contract (parameters confirmed)

---

### Phase 2: MECE Scope Mapping

**Purpose:** Decompose topic into mutually exclusive, collectively exhaustive coverage dimensions.

**Steps:**
1. Load MECE pattern appropriate to domain (see `mece-decomposition-guide.md`):
   - Product domain → Market Research or Technology Evaluation pattern
   - Architecture domain → Technology Evaluation pattern
   - Research domain → Strategic Research pattern
   - Requirements domain → Custom decomposition
2. Adapt pattern to specific topic
3. Create 3-7 coverage dimensions (fewer for narrow topics, more for broad)
4. For each dimension, identify:
   - What knowledge is needed
   - What "complete" looks like
   - Priority level (critical, important, nice-to-have)
5. Present scope map to interviewee: "I've identified [N] areas to cover: [list]. Does this capture everything important?"
6. Refine based on feedback, add missing dimensions

**Quality Gates:**
- Scope definition: All boundaries explicitly stated (in-scope, out-of-scope)
- MECE structure: Categories non-overlapping and collectively exhaustive

**Output:** Coverage Map with dimensions and sub-areas

---

### Phase 3: Adaptive Elicitation (Iterative)

**Purpose:** Extract knowledge through adaptive questioning.

**CRITICAL CONSTRAINT:** Ask ONE question per turn. Wait for response before next question.

**Workflow Per Turn:**

```
1. SELECT DIMENSION
   └─ Choose highest-priority uncovered area

2. SELECT QUESTION TYPE (see Question Taxonomy)
   └─ Based on what's known/unknown about dimension

3. FORMULATE QUESTION
   ├─ Clear and specific
   ├─ Single focus (not compound)
   └─ Non-leading

4. AWAIT RESPONSE
   └─ DO NOT proceed without interviewee input

5. INTEGRATE RESPONSE
   ├─ Update Knowledge Map
   ├─ Link to related findings
   └─ Note any contradictions

6. TRACK CONFIDENCE
   ├─ Assign confidence score (0.0-1.0)
   └─ Tag uncertainty type (EPISTEMIC | ALEATORY | MODEL)

7. SURFACE ASSUMPTIONS
   ├─ Explicit: Directly stated
   ├─ Implicit: Inferred from response
   └─ Structural: About framing itself

8. APPLY BIAS PROTECTION (if needed)
   ├─ Frame equivalence test for critical claims
   └─ Disconfirmation hunt for confident assertions

9. EVALUATE CONTINUATION
   ├─ More questions needed for this dimension?
   └─ Move to next dimension?
```

### Question Type Selection Logic

```
IF dimension is new AND context unknown:
    → GRAND TOUR (establish landscape)

ELIF need to understand organization/hierarchy:
    → STRUCTURAL

ELIF need to differentiate similar concepts:
    → CONTRAST

ELIF response was abstract, need illustration:
    → EXAMPLE

ELIF response was vague or incomplete:
    → PROBING

ELIF need to stress-test assumption or claim:
    → DEVIL'S ADVOCATE

ELIF statement is ambiguous:
    → CLARIFYING

ELIF synthesizing understanding for dimension:
    → CONFIRMING
```

### Adaptive Rules

- **Follow energy:** Pursue topics where interviewee shows engagement
- **Notice gaps:** Track what hasn't been said as carefully as what has
- **Connect threads:** Build on earlier responses, reference previous answers
- **Respect scope:** Stay within established boundaries
- **Vary types:** Don't use same question type consecutively
- **One question per turn:** Never ask multiple