Expert Knowledge Elicitation

SKILL: Expert Knowledge Elicitation and Representation

license: Apache-2.0

Metadata

• Name: expert-knowledge-elicitation
• Source: Eliciting and Representing the Knowledge of Experts (Hoffman & Lintern)
• Domain: Knowledge engineering, cognitive systems, organizational learning
• Triggers:
  • Designing systems that require expert-level decision-making
  • Capturing knowledge from retiring or departing experts
  • Understanding why documented procedures fail in practice
  • Building human-agent collaborative systems
  • Analyzing expert performance for training or automation
  • Investigating failures in complex sociotechnical systems

When to Use This Skill

Load this skill when facing:

• The Documentation Gap: Official procedures exist but systems still fail; need to understand "true work"
• Knowledge Preservation: Critical expertise at risk due to turnover, retirement, or organizational change
• Intelligent System Design: Building agents/tools that must reason like domain experts
• Performance Analysis: Understanding why experts succeed where novices fail
• Collaboration Design: Creating effective human-agent or human-human knowledge sharing
• Training Development: Need to teach not just facts but expert reasoning strategies
• Failure Investigation: System failures despite procedure compliance; need deeper understanding

Anti-use cases: Simple procedural tasks, well-documented routine work, contexts where expertise doesn't exist yet (pure innovation vs. refined practice)

Core Mental Models

1. The Three-Level Architecture

Expert knowledge exists at three distinct levels, each requiring different methods and serving different purposes:

Level 1: Work Domain Structure (What the world contains)

• Functional relationships, constraints, physical laws governing the domain
• Example: In power grid management—electrical physics, grid topology, load dynamics
• Captured via: Work Domain Analysis, abstraction-decomposition matrices
• Purpose: Understanding task constraints, enabling principled reasoning about domain

Level 2: Conceptual Knowledge (What experts know about the domain)

• Domain concepts, categories, relationships, mental models
• Example: "Load shedding," "cascade failure," "frequency stability"—expert terminology and concept networks
• Captured via: Concept mapping, card sorting, structured interviews
• Purpose: Shared vocabulary, semantic understanding, knowledge organization

Level 3: Reasoning Strategies (How experts think through problems)

• Decision-making heuristics, perceptual cues, problem-solving strategies
• Example: "When frequency drops below X and Y is rising, consider Z before standard procedure"
• Captured via: Critical Decision Method, think-aloud protocols, case studies
• Purpose: Adaptive intelligence, handling non-routine situations, expert-level performance

Critical insight: Most failed systems conflate these levels—they capture concepts (L2) but miss reasoning strategies (L3), or document procedures without understanding functional constraints (L1). Effective systems require all three.

2. Scaffolding Over Extraction

The "tacit knowledge problem" is a methodological failure, not a fundamental barrier:

The myth: Experts possess inexpressible, unconscious knowledge that resists articulation The reality: Poor elicitation methods fail to provide adequate cognitive scaffolding

Effective scaffolding principles:

• Ground discussion in concrete cases, not abstract generalizations
• Use visual/spatial representations (concept maps, diagrams) to externalize thinking
• Multi-pass retrospection: revisit the same case with different analytical lenses
• Collaborative discovery stance: "Let's figure this out together" not "Tell me what you know"
• Challenge and probing: respectful questioning that surfaces unstated assumptions

Diagnostic marker: When experts say "I've never thought about it this way, but..." you've achieved good scaffolding—they're discovering their own knowledge through the elicitation process.

3. The Gold Is Not in the Documents

Critical expertise exists outside official procedures—adaptive heuristics, workarounds, "naughty" departures:

Why procedures miss expertise:

• Written for liability/compliance, not actual practice
• Cannot anticipate all contextual variations experts handle
• Experts develop efficient shortcuts through experience
• Productive rule-bending represents evolved wisdom, not negligence

What's missing from documentation:

• Perceptual cues that trigger strategy selection
• Context-dependent trade-offs and judgment calls
• Workarounds for known system/process limitations
• Pattern recognition shortcuts ("smells like X situation")
• Metacognitive awareness ("this is where novices typically err")

Implication for system design: Systems built only from documented procedures will be brittle, inefficient, and miss the actual intelligence that makes work succeed.

4. Differential Utility, Not Differential Access

Methods differ in what they're useful for, not what knowledge types they can "access":

Failed hypothesis: Different methods tap different knowledge types (declarative vs. procedural, conscious vs. unconscious) Validated finding: All methods can elicit all knowledge types given sufficient effort, but efficiency and naturalness differ dramatically

Method-to-purpose matching:

• Critical Decision Method: Excellent for reasoning strategies, decision cues, uncertainty handling
• Concept Mapping: Efficient for domain knowledge structure, terminology, relationships
• Work Domain Analysis: Superior for functional constraints, system relationships, goal hierarchies
• Simulated Tasks: Good for procedural knowledge, revealing implicit assumptions through errors
• Contrasting Cases: Effective for discrimination criteria, category boundaries

Selection principle: Match method to project goals and expert's natural articulation style, not to assumed "knowledge type."

5. Knowledge Elicitation as Ongoing Practice

Shift from one-time extraction project to continuous organizational capability:

Traditional model (limited): KE as discrete phase in system development—identify need → extract knowledge → build system → done

Mature model (sustainable): KE as ongoing practice integrated into organizational culture:

• Regular capture sessions integrated into project retrospectives
• Mentoring relationships structured around knowledge articulation
• Collaborative concept maps as living organizational knowledge bases
• Post-incident CDM sessions for continuous learning
• Lightweight capture tools embedded in normal workflow

Organizational contexts requiring ongoing KE:

• Distributed expertise (knowledge scattered across specialists)
• Rare-event domains (limited learning opportunities)
• High-turnover environments (continuous knowledge loss risk)
• Rapid domain evolution (expertise constantly updating)
• Safety-critical systems (failure lessons must be captured and shared)

Decision Frameworks

When selecting elicitation methods:

IF you need to understand expert reasoning during challenging decisions
THEN use Critical Decision Method with concrete past incidents
BECAUSE abstract hypotheticals miss context-dependent factors experts actually use

IF building shared vocabulary or knowledge bases
THEN start with concept mapping sessions
BECAUSE visual representation scaffolds articulation and reveals organizational knowledge gaps

IF designing systems for complex domains with multiple goals
THEN conduct Work Domain Analysis first
BECAUSE understanding functional structure prevents brittle procedural automation

IF experts struggle to articulate knowledge in interviews
THEN switch to case-based or simulation methods
BECAUSE performance reveals knowledge that introspection misses

IF documented procedures exist but systems still fail
THEN investigate the gap between "work-as-imagined" and "work-as-done"
BECAUSE critical adaptive expertise lives in undocumented practice

When designing intelligent systems:

IF the system must handle non-routine situations
THEN capture Level 3 reasoning strategies, not just Level 2 concepts
BECAUSE brittle rule-following fails when conditions change

IF building human-agent collaboration
THEN design bidirectional scaffolding interfaces
BECAUSE collaboration is knowledge co-construction, not information transfer

IF experts will work alongside the system
THEN capture expert perceptual cues and decision triggers
BECAUSE agents need to understand what experts monitor and why

IF domain knowledge evolves rapidly
THEN build continuous learning mechanisms, not static deployment
BECAUSE one-time capture creates frozen knowledge that decays

When facing organizational knowledge challenges:

IF critical expert is retiring
THEN conduct CDM sessions on their most challenging past cases
BECAUSE difficult cases reveal strategic knowledge that routine work doesn't

IF team needs shared understanding
THEN create collaborative concept maps, not individual interviews
BECAUSE shared artifact construction reveals and resolves conceptual conflicts

IF investigating system failure
THEN use multi-pass retrospection with involved practitioners
BECAUSE single-pass incident reports miss evolving understanding and contextual factors

IF knowledge exists across distributed specialists
THEN use contrasting cases to elicit discrimination criteria
BECAUSE specialists best articulate expertise by comparing domains

Reference Documentation

Anti-Patterns

1. Checklist Reductionism

• Mistake: Reducing expert knowledge to simple checklists or decision trees
• Why harmful: Misses context-dependent judgment, adaptive strategies, exception handling
• Reality: Experts use checklists as memory aids within rich strategic understanding
• Detection: System handles routine cases but fails on edge cases experts routinely manage

2. The Documentation Delusion

• Mistake: Assuming critical knowledge exists in procedures, manuals, or documentation
• Why harmful: Misses adaptive heuristics, workarounds, and contextual judgment that make work succeed
• Reality: Documentation captures work-as-imagined; expertise lives in work-as-done
• Detection: Complaints that "the system doesn't understand how we really work"

3. Single-Method Monogamy

• Mistake: Using only interviews, or only observation, or only documentation review
• Why harmful: No single method captures all three knowledge levels effectively
• Reality: Comprehensive understanding requires methodological triangulation
• Detection: Knowledge base has gaps (rich concepts but no reasoning strategies, or procedures without functional understanding)

4. Abstract Hypothetical Questioning

• Mistake: Asking "What would you do if..." or "How do you generally handle X?"
• Why harmful: Produces textbook answers, not actual expertise; misses contextual factors
• Reality: Expert knowledge is case-based and context-dependent
• Detection: Experts give generic answers; knowledge doesn't match observed performance

5. One-and-Done Extraction

• Mistake: Treating KE as discrete project phase—capture once, then deploy
• Why harmful: Domain knowledge evolves; expertise develops; organizational memory decays
• Reality: Effective KE is ongoing organizational practice, not one-time event
• Detection: Deployed systems become increasingly disconnected from current practice

6. The Tacit Knowledge Excuse

• Mistake: Attributing elicitation failure to "unconscious/inexpressible knowledge"
• Why harmful: Blames fundamental barriers instead of methodological inadequacy
• Reality: Poor scaffolding, not tacit knowledge, causes most articulation failures
• Detection: Switching methods or improving scaffolding suddenly reveals "inexpressible" knowledge

7. Conflating Knowledge Levels

• Mistake: Not distinguishing domain structure, conceptual knowledge, and reasoning strategies
• Why harmful: Produces confused representations that serve no purpose well
• Reality: Each level requires different methods and supports different system capabilities
• Detection: Knowledge base contains mixed content types without clear organizational principle

Shibboleths: Distinguishing Deep Understanding

Someone has truly internalized this book if they:

1. Immediately ask "Which level?"

• When discussing knowledge capture, they instinctively distinguish: "Are we talking about domain structure, conceptual knowledge, or reasoning strategies?"
• They don't treat "expert knowledge" as monolithic

2. Recognize productive rule-bending

• They don't assume deviation from procedures indicates negligence
• They actively investigate workarounds as potential sources of evolved wisdom
• They understand "naughty" departures may be more intelligent than compliance

3. Diagnose elicitation problems methodologically

• When experts struggle to articulate, they change scaffolding, not blame tacit knowledge
• They say things like "Let's try grounding this in a specific case" or "Can we map this visually?"

4. Match methods to purposes, not knowledge types

• They don't say "We'll use CDM for procedural knowledge and interviews for declarative"
• They say "CDM for reasoning strategies, concept mapping for domain structure, WDA for functional constraints"

5. Spot documentation gaps automatically

• When reviewing procedures, they immediately ask "But what do experts actually do when X happens?"
• They assume documentation misses critical adaptive strategies

6. Think in terms of continuous practice, not one-time extraction

• They design ongoing capture mechanisms, not just project-based elicitation
• They integrate KE into normal organizational workflow

7. Use the transactional view of knowledge

• They look for knowledge in interactions, artifacts, tools, and practices—not just individual minds
• They design for emergent collective intelligence, not knowledge transfer

Surface-level understanding reveals itself when someone:

• Treats all knowledge elicitation as "interviewing subject matter experts"
• Assumes documentation captures critical knowledge
• Blames "tacit knowledge" when elicitation produces shallow results
• Thinks procedures and expertise are the same thing
• Designs one-time extraction projects instead of ongoing capability
• Doesn't distinguish between knowing what the domain contains vs. how experts reason about it
• Believes a single method (usually interviews) can capture everything important

The ultimate shibboleth:

Deep understanding: "We need to understand not just what experts know about power grids (L2 concepts), but also how the grid itself functions (L1 domain structure) and how experts reason when standard operating conditions break down (L3 strategies). We'll use WDA for functional relationships, concept mapping for terminology and models, and CDM on the blackout incident for decision-making under uncertainty. And we'll build this as ongoing practice, not one-time capture."

Surface understanding: "We should interview the senior engineers to document their knowledge of power grid operations, then build a knowledge base from the transcripts."

The difference is immediate, structural, and reveals whether someone grasps the fundamental architecture of expert knowledge or is still operating in the failed extraction paradigm.