Research Interviewer

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:

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

Validation Mode Effects

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 questions at once

Quality Gate: Epistemic labeling - every finding tagged with uncertainty type

Output: Growing Knowledge Map with confidence scores

Phase 4: Continuous Confidence Tracking

Purpose: Maintain real-time epistemic status of all gathered knowledge.

Runs parallel to Phase 3.

Mechanism:

1. Classify each finding using uncertainty taxonomy:

   • EPISTEMIC: We don't know, but COULD find out with more questions
   • ALEATORY: Inherent uncertainty that CANNOT be reduced (future events, randomness)
   • MODEL: Answer depends on definitions/framework choices

2. Assign confidence score (0.0-1.0) based on:

   • Clarity of response
   • Consistency with other findings
   • Evidence quality (direct experience vs. hearsay)
   • Specificity (concrete vs. vague)

3. Track coverage per dimension:

   • Questions asked
   • Findings captured
   • Confidence level
   • Remaining gaps

4. Calculate overall confidence:

   • Per-dimension: Average of finding confidences
   • Overall: Weighted average (critical dimensions weighted 1.5x)

5. Identify high-value targets:

   • EPISTEMIC gaps in critical dimensions → ask more questions
   • MODEL uncertainties → surface for explicit choice
   • ALEATORY factors → document and move on

Quality Gate: Confidence threshold - overall confidence >= confidence_threshold

Phase 5: Validation and Synthesis

Purpose: Verify consistency and build unified artifact.

Steps:

5.1 Consistency Matrix

Cross-reference all findings for contradictions:

• Compare related claims across dimensions
• Flag any inconsistencies
• If conflict found: Present both versions to interviewee for resolution

"Earlier you mentioned [X]. Just now you said [Y].
These seem to conflict. Can you help me understand?"

5.2 Assumption Inventory

Compile all surfaced assumptions:

Validate critical assumptions: "It sounds like we're assuming [X]. Is that right? What would change if that assumption were wrong?"

5.3 Steelmanning

Present the strongest version of gathered knowledge:

"Let me play back what I've understood. The core issue is [X],
driven by [Y], with the key constraint being [Z]. The main
stakeholders are [A, B, C], and success looks like [criteria].

Is this an accurate and complete representation?"

Iterate until interviewee confirms.

5.4 Synthesis

Build unified knowledge structure:

• Organize findings by dimension
• Ensure all MECE dimensions covered
• Apply output format structure (see Output Specifications)
• Calculate final confidence scores

Quality Gates:

• Consistency verified: No unresolved contradictions
• Assumptions surfaced: All critical assumptions documented and validated

Output: Synthesized knowledge ready for formatting

Phase 6: Calibrated Termination

Purpose: Ensure completeness and produce deliverable.

Steps:

6.1 Unknown Unknowns Probe

Ask these five questions before concluding:

1. "What haven't I asked about that I should have?"
2. "What would someone who knows this domain well be surprised I didn't ask?"
3. "Are there adjacent areas that connect to this topic we should touch on?"
4. "What assumptions might I be making that aren't valid?"
5. "What could go wrong that we haven't discussed?"

6.2 Coverage Assessment

Review Coverage Map:

• All dimensions at minimum confidence?
• Any CRITICAL gaps remaining?
• Epistemic gaps that could be closed with one more question?

6.3 Interviewee Calibration

Capture the interviewee's confidence:

"How confident are you in the completeness of what we've covered?"
"Which areas are you most certain about? Least certain?"

Map to final confidence report.

6.4 Termination Decision

CHECKPOINT: confidence_threshold_adjustment

• If confidence near threshold and domain suggests different: AskUserQuestion
• Example: "Confidence is 0.82 (threshold 0.85). For exploratory, 0.80 is often sufficient."

CHECKPOINT: premature_termination_check

• If termination criteria met but warning signs present: AskUserQuestion
• Warning signs: new areas surfaced, critical dimension gaps, unexplored topics mentioned

TERMINATE IF:
  - confidence_threshold met (default 0.85)
  - max_questions reached
  - Interviewee signals completion
  - No new significant information in last 3 questions

CONTINUE IF:
  - Critical gaps remain
  - Unresolved contradictions exist
  - Unknown unknowns probe surfaced new areas

6.5 Output Generation

Select template based on output_format:

• PROBLEM-STATEMENT → See Section 6.1 of Output Specifications
• KNOWLEDGE-CORPUS → See Section 6.2 of Output Specifications
• REQUIREMENTS → See Section 6.3 of Output Specifications

Quality Gates:

• Confidence threshold met
• Interviewee calibration complete

Output: Final deliverable in specified format

5. Question Taxonomy

The 8 Question Types

Typical Flow Within a Dimension

Grand Tour → Structural → Example → Probing → Contrast → Devil's Advocate → Confirming

Reference: See references/question-taxonomy.md for detailed examples and templates.

6. Output Specifications

6.1 PROBLEM-STATEMENT Format

Aligns with CONTRACT-01 from artifact-contracts.yaml.

<problem_statement contract="CONTRACT-01">
  <metadata>
    <artifact_id>[PS-YYYY-MM-DD-XXXXX]</artifact_id>
    <contract_type>PROBLEM-STATEMENT</contract_type>
    <created_at>[ISO 8601]</created_at>
    <created_by>research-interviewer</created_by>
    <confidence>[0.0-1.0]</confidence>
  </metadata>

  <statement>[Clear, actionable problem statement]</statement>

  <jtbd_format>
    <situation>[When/context in which the problem arises]</situation>
    <motivation>[What the user wants to do]</motivation>
    <outcome>[Desired result/benefit]</outcome>
  </jtbd_format>

  <context>
    <domain>[product | architecture | strategy | research | ...]</domain>
    <stakeholders>
      <stakeholder role="[role]">[Who]</stakeholder>
    </stakeholders>
    <constraints>
      <constraint>[Hard constraint]</constraint>
    </constraints>
    <assumptions>
      <assumption type="[explicit|implicit|structural]" validated="[true|false]">
        [Assumption text]
      </assumption>
    </assumptions>
  </context>

  <success_criteria>
    <criterion measurable="[true|false]" priority="[must_have|should_have|nice_to_have]">
      [Criterion text]
    </criterion>
  </success_criteria>

  <epistemic_status>
    <overall_confidence>[0.0-1.0]</overall_confidence>
    <uncertainty_breakdown>
      <epistemic_gaps>[What we don't know but could find out]</epistemic_gaps>
      <aleatory_factors>[Inherent uncertainties]</aleatory_factors>
      <model_dependencies>[Framework-dependent answers]</model_dependencies>
    </uncertainty_breakdown>
  </epistemic_status>
</problem_statement>

6.2 KNOWLEDGE-CORPUS Format

Optimized for RAG systems and context injection.

<knowledge_corpus>
  <metadata>
    <corpus_id>[KC-YYYY-MM-DD-XXXXX]</corpus_id>
    <topic>[Interview topic]</topic>
    <created_at>[ISO 8601]</created_at>
    <created_by>research-interviewer</created_by>
    <overall_confidence>[0.0-1.0]</overall_confidence>
  </metadata>

  <coverage_map>
    <dimension id="D1" name="[Name]" confidence="[0.0-1.0]">
      <finding id="D1F1" confidence="[0.0-1.0]"
               uncertainty_type="[EPISTEMIC|ALEATORY|MODEL]">
        <statement>[What was learned]</statement>
        <evidence>[How we know this]</evidence>
        <source_question>[Question that elicited this]</source_question>
      </finding>
    </dimension>
  </coverage_map>

  <relationships>
    <relationship from="[finding_id]" to="[finding_id]"
                  type="[depends_on|contradicts|supports|refines]">
      [Description]
    </relationship>
  </relationships>

  <assumption_inventory>
    <assumption id="A1" type="[explicit|implicit|structural]"
                validated="[true|false]" confidence="[0.0-1.0]">
      <statement>[Assumption]</statement>
      <implications>[What depends on this]</implications>
    </assumption>
  </assumption_inventory>

  <gaps_registry>
    <gap dimension="[dimension_id]" severity="[critical|significant|minor]">
      <description>[What's missing]</description>
      <suggested_resolution>[How to close]</suggested_resolution>
    </gap>
  </gaps_registry>
</knowledge_corpus>

6.3 REQUIREMENTS Format

Job stories with acceptance criteria.

<requirements>
  <metadata>
    <requirements_id>[REQ-YYYY-MM-DD-XXXXX]</requirements_id>
    <topic>[Interview topic]</topic>
    <created_at>[ISO 8601]</created_at>
    <created_by>research-interviewer</created_by>
    <overall_confidence>[0.0-1.0]</overall_confidence>
  </metadata>

  <job_stories>
    <job_story id="JS1" priority="[must_have|should_have|nice_to_have]"
               confidence="[0.0-1.0]">
      <situation>When [context/trigger]</situation>
      <motivation>I want to [action/capability]</motivation>
      <outcome>So that [benefit/result]</outcome>
      <acceptance_criteria>
        <criterion id="JS1AC1" testable="[true|false]">[Criterion]</criterion>
      </acceptance_criteria>
    </job_story>
  </job_stories>

  <constraints>
    <constraint id="C1" type="[technical|business|regulatory]"
                non_negotiable="[true|false]">
      <description>[Constraint]</description>
      <rationale>[Why this constraint exists]</rationale>
    </constraint>
  </constraints>

  <non_functional_requirements>
    <nfr id="NFR1" category="[performance|security|scalability|...]">
      <description>[NFR description]</description>
      <measurement>[How to verify]</measurement>
    </nfr>
  </non_functional_requirements>

  <traceability>
    <finding_to_requirement from="[finding_id]" to="[requirement_id]">
      [How finding led to requirement]
    </finding_to_requirement>
  </traceability>
</requirements>

Reference: See references/output-templates.md for complete templates with examples.

7. Quality Gates

8. Workflow Integration

This skill serves as the upstream elicitation component in the research workflow:

┌─────────────────────────┐
│  research-interviewer   │  ◀── THIS SKILL
│                         │  Elicit research requirements
└───────────┬─────────────┘
            │
            │ Produces: PROBLEM-STATEMENT | KNOWLEDGE-CORPUS | REQUIREMENTS
            │
            ▼
┌─────────────────────────┐
│  create-research-brief  │  Design multi-LLM research strategy
│       (Phase 1)         │
└───────────┬─────────────┘
            │
            ▼
┌─────────────────────────┐
│   Execute Research      │  Run prompts across models
│   (Manual or Agent)     │
└───────────┬─────────────┘
            │
            ▼
┌─────────────────────────┐
│  create-research-brief  │  Consolidate into report
│       (Phase 2)         │
└─────────────────────────┘

Artifact Flow

9. Behavioral Guidelines

• Curious not interrogating: Maintain genuine interest, avoid feeling like an interrogation
• Patient with complexity: Allow time for complex explanations
• Humble about understanding: Check assumptions, don't assume you understand
• Goal-anchored: Keep interview goal visible throughout
• Efficient with time: Respect interviewee's time, don't over-ask covered areas
• One question at a time: Never bundle multiple questions

10. References

External References

11. Templates

12. Examples

Example 1: Problem Statement Elicitation

input:
  interview_goal: "Understand authentication pain points for research brief"
  topic: "User authentication system"
  output_format: PROBLEM-STATEMENT
  domain_reference: product
  validation_mode: balanced

flow:
  phase_1: Confirmed goal, selected PROBLEM-STATEMENT output
  phase_2: Created 5 MECE dimensions - UX, security, maintenance, scalability, integration
  phase_3: 18 questions across dimensions
           - Grand Tour on current auth flow
           - Structural on user types and permissions
           - Example on recent authentication failure
           - Probing on "complicated" password reset
           - Devil's Advocate on SSO assumption
           - Confirming on core pain point
  phase_4: Tracked confidence per dimension
           - UX: 0.90, Security: 0.85, Maintenance: 0.75
           - Scalability: 0.60 (EPISTEMIC gap identified)
           - Integration: 0.80
  phase_5: Built consistency matrix, no contradictions
           Surfaced 4 assumptions (2 explicit, 2 implicit)
           Steelmanned: "Users abandon MFA due to friction in recovery flow"
  phase_6: Unknown unknowns probe revealed compliance requirement
           Final confidence: 0.87

output: PROBLEM-STATEMENT artifact per CONTRACT-01 schema

Example 2: Knowledge Corpus Extraction

input:
  interview_goal: "Document data pipeline architecture for context"
  topic: "Data engineering infrastructure"
  output_format: KNOWLEDGE-CORPUS
  domain_reference: architecture
  validation_mode: rigorous

flow:
  phase_1: Established goal for RAG context generation
  phase_2: 5 dimensions - ingestion, processing, storage, serving, monitoring
  phase_3: 24 questions, heavy Structural and Example types
           Rigorous mode triggered more Devil's Advocate probes
  phase_4: 3 MODEL uncertainties flagged (definition-dependent)
           12 EPISTEMIC findings, 9 ALEATORY factors documented
  phase_5: Built knowledge graph with 47 findings
           12 inter-finding relationships documented
  phase_6: Confidence: 0.91, interviewee highly confident

output: KNOWLEDGE-CORPUS with 47 findings across 5 dimensions

Example 3: Requirements Discovery

input:
  interview_goal: "Gather requirements for executive reporting feature"
  topic: "Executive reporting dashboard"
  output_format: REQUIREMENTS
  domain_reference: requirements
  validation_mode: empathetic

flow:
  phase_1: Empathetic mode for executive stakeholder
  phase_2: 5 dimensions - data sources, visualizations, access control, export, scheduling
  phase_3: 15 questions, softer probing style
           Heavy use of Example and Confirming questions
  phase_4: High ALEATORY uncertainty around future metrics
           Documented as "subject to change"
  phase_5: Generated 8 job stories with acceptance criteria
           Traced each to source findings
  phase_6: Confidence: 0.86, stakeholder confirmed completeness

output: REQUIREMENTS with 8 job stories, 4 constraints, 3 NFRs

Quick Start

/research-interviewer
interview_goal: "[What you'll use the information for]"
topic: "[Subject to interview about]"
output_format: PROBLEM-STATEMENT