Elicitation

<objective> Elicitation lets an expert agent self-review and self-correct its own code before external validation, drawing on 75 elicitation techniques across 12 categories (code quality, security, performance, architecture, testing, docs, UX, data, concurrency, integration, observability, maintainability) inspired by BMAD-METHOD. Three modes control how techniques are chosen: MANUAL (default, user picks from 5 presented options), AUTO (auto-detected and applied silently), and SKIP (bypass straight to sniper).

It sits between Execute and eXamine in the APEX flow, scores itself against an exit threshold (>=90% proceed, 70-89% document gaps, <70% iterate), and persists its findings to .claude/apex/docs/elicit-{task-slug}.json so a later pass can diff against prior verdicts instead of restarting. </objective>

Elicitation Skill

Purpose

Enable expert agents to self-review and self-correct their code before external validation (sniper). Based on BMAD-METHOD's 75 elicitation techniques.

---

3 Execution Modes

Mode 1: MANUAL (default)

Expert presents 5 relevant techniques → User chooses → Expert applies

Mode 2: AUTO (--auto)

Expert auto-detects code type → Auto-selects techniques → Applies silently

Mode 3: SKIP (--skip)

Skip elicitation → Go directly to sniper validation

---

Quick Start

After Execute phase, expert runs:

# Manual mode (default)
> Apply elicitation skill

# Auto mode (no prompts)
> Apply elicitation skill --auto

# Skip self-review
> Apply elicitation skill --skip

---

Workflow Overview

┌─────────────────────────────────────────────────────────┐
│              ELICITATION WORKFLOW                       │
│                                                         │
│  Step 0: Init           → Load context                 │
│  Step 1: Analyze Code   → Detect code type             │
│  Step 2: Select         → Choose techniques (or auto)  │
│  Step 3: Apply Review   → Execute techniques           │
│  Step 4: Self-Correct   → Fix own issues               │
│  Step 5: Report         → Summary before sniper        │
└─────────────────────────────────────────────────────────┘

---

Auto-Detection Matrix

Code Type Detected	Auto-Selected Techniques
Authentication/Security	Security Audit, OWASP Check, Input Validation
API Endpoints	Error Handling, Type Coverage, API Contracts
Database/ORM	N+1 Detection, Migration Safety, Data Integrity
UI Components	Accessibility, Edge Cases, Loading States
Business Logic	SOLID Compliance, Unit Test Coverage, Edge Cases
Refactoring	Breaking Changes, Regression Analysis, Backward Compat
Performance Critical	Profiling, Memory Analysis, Complexity Check
Config/Docs/Plugin files (.md agents/skills, hooks.json, frontmatter YAML)	CQ-01, DOC-01, INT-01 + validation by the strictest parser in the consumption chain (e.g. js-yaml strict for frontmatter, json.tool for JSON) -- never "looks well-formed"

---

Technique Categories (12)

Full catalog: references/techniques-catalog.md

1. Code Quality (7): Code review, Pattern detection, Complexity analysis...
2. Security (7): OWASP audit, Input validation, Auth check...
3. Performance (6): Profiling, N+1 detection, Memory analysis...
4. Architecture (6): SOLID check, Dependency analysis, Coupling review...
5. Testing (6): Edge cases, Boundary testing, Error paths...
6. Documentation (6): API review, Comment check, Type coverage...
7. UX (6): Accessibility, Error messages, Loading states...
8. Data (6): Schema validation, Migration safety, Data integrity...
9. Concurrency (6): Race conditions, Deadlock analysis, State sync...
10. Integration (7): API contracts, Backward compat, Breaking changes...
11. Observability (6): Logging, Metrics, Error tracking...
12. Maintainability (6): Readability, Naming, File organization...

Total: 75 techniques

---

Integration with APEX

A-nalyze → P-lan → E-xecute → [ELICIT] → X-amine
                       │          │           │
                       ▼          ▼           ▼
                    Expert    Expert       sniper
                     code    self-review   (final)

Benefits:

• Expert catches own mistakes before sniper
• Faster validation (less sniper corrections)
• Knowledge retention (expert learns from self-review)

---

Forbidden

• ❌ Skip init step (must load context)
• ❌ Apply techniques without understanding code type
• ❌ Self-correct without documenting changes
• ❌ Report without listing applied techniques
• ❌ Use techniques outside expertise domain

---

Exit Criteria (Step 5: Report)

Score = techniques applied / techniques selected, weighted by critical category (Security, Architecture).

Score	Status	Action
≥ 90%	🟢	Proceed to sniper
70-89%	🟡	Document gaps, then proceed
< 70%	🔴	Iterate before sniper

Self-correction failure: If a self-correction breaks the code → revert that correction and keep the finding as a report item instead.

---

Artifact Contract

Step 5 persists .claude/apex/docs/elicit-{task-slug}.json so a later pass diffs against prior verdicts instead of restarting from scratch. Full contract, {task-slug} derivation, and JSON schema: references/artifact-contract.md.

---

Steps Reference

Step	File	Purpose
0	steps/step-00-init.md	Load context, detect mode, load prior artifact if present
1	steps/step-01-analyze-code.md	Analyze written code
2	steps/step-02-select-techniques.md	Select techniques
3	steps/step-03-apply-review.md	Apply review
4	steps/step-04-self-correct.md	Self-correct
5	steps/step-05-report.md	Generate report, persist elicit-{task-slug}.json