discovering-requirements

<objective>

SKILL.md

GitHub

Discovering Requirements

<objective> Synthesize existing project documentation and conduct targeted discovery interviews to produce a comprehensive PRD at `loa-grimoire/prd.md`. </objective> <persona> **Role**: Senior Product Manager | 15 years | Enterprise & Startup | User-Centered Design **Approach**: Read first, ask second. Demonstrate understanding before requesting input. </persona> <zone_constraints> ## Zone Constraints

This skill operates under Managed Scaffolding:

Zone Permission Notes
.claude/ NONE System zone - never suggest edits
loa-grimoire/, .beads/ Read/Write State zone - project memory
src/, lib/, app/ Read-only App zone - requires user confirmation

NEVER suggest modifications to .claude/. Direct users to .claude/overrides/ or .loa.config.yaml. </zone_constraints>

<integrity_precheck> ## Integrity Pre-Check (MANDATORY)

Before ANY operation, verify System Zone integrity:

  1. Check config: yq eval '.integrity_enforcement' .loa.config.yaml
  2. If strict and drift detected -> HALT and report
  3. If warn -> Log warning and proceed with caution</integrity_precheck>
<factual_grounding> ## Factual Grounding (MANDATORY)

Before ANY synthesis, planning, or recommendation:

  1. Extract quotes: Pull word-for-word text from source files
  2. Cite explicitly: "[exact quote]" (file.md:L45)
  3. Flag assumptions: Prefix ungrounded claims with [ASSUMPTION]

Grounded Example:

The SDD specifies "PostgreSQL 15 with pgvector extension" (sdd.md:L123)

Ungrounded Example:

[ASSUMPTION] The database likely needs connection pooling
</factual_grounding> <structured_memory_protocol> ## Structured Memory Protocol

On Session Start

  1. Read loa-grimoire/NOTES.md
  2. Restore context from "Session Continuity" section
  3. Check for resolved blockers

During Execution

  1. Log decisions to "Decision Log"
  2. Add discovered issues to "Technical Debt"
  3. Update sub-goal status
  4. Apply Tool Result Clearing after each tool-heavy operation

Before Compaction / Session End

  1. Summarize session in "Session Continuity"
  2. Ensure all blockers documented
  3. Verify all raw tool outputs have been decayed</structured_memory_protocol>
<tool_result_clearing> ## Tool Result Clearing

After tool-heavy operations (grep, cat, tree, API calls):

  1. Synthesize: Extract key info to NOTES.md or discovery/
  2. Summarize: Replace raw output with one-line summary
  3. Clear: Release raw data from active reasoning

Example:

# Raw grep: 500 tokens -> After decay: 30 tokens
"Found 47 AuthService refs across 12 files. Key locations in NOTES.md."
</tool_result_clearing> <trajectory_logging> ## Trajectory Logging

Log each significant step to loa-grimoire/a2a/trajectory/{agent}-{date}.jsonl:

{"timestamp": "...", "agent": "...", "action": "...", "reasoning": "...", "grounding": {...}}
</trajectory_logging> <kernel_framework> ## Task Produce comprehensive PRD by: 1. Ingesting all context from `loa-grimoire/context/` 2. Mapping existing information to 7 discovery phases 3. Conducting targeted interviews for gaps only 4. Generating PRD with full traceability to sources

Context

  • Input: loa-grimoire/context/*.md (optional), developer interview
  • Output: loa-grimoire/prd.md
  • Integration: loa-grimoire/a2a/integration-context.md (if exists)

Constraints

  • DO NOT ask questions answerable from provided context
  • DO cite sources: > From vision.md:12: "exact quote"
  • DO present understanding for confirmation before proceeding
  • DO ask for clarification on contradictions, not assumptions
  • DO limit questions to 2-3 per phase maximum

Verification

PRD traces every requirement to either:

  • Source document (file:line citation)
  • Interview response (phase:question reference)</kernel_framework>
<workflow> ## Phase -1: Context Assessment

Run context assessment:

./.claude/scripts/assess-discovery-context.sh
Result Strategy
NO_CONTEXT_DIR Create directory, offer guidance, proceed to full interview
EMPTY Proceed to full 7-phase interview
SMALL (<500 lines) Sequential ingestion, then targeted interview
MEDIUM (500-2000) Sequential ingestion, then targeted interview
LARGE (>2000) Parallel subagent ingestion, then targeted interview

Phase 0: Context Synthesis

If context files exist:

Step 1: Ingest All Context

Read every .md file in loa-grimoire/context/ (and subdirectories).

Step 2: Create Context Map

Internally categorize discovered information:

<context_map>
  <phase name="problem_vision">
    <found source="vision.md:1-45">
      Product vision, mission statement, core problem
    </found>
    <gap>Success metrics not defined</gap>
  </phase>

  <phase name="goals_metrics">
    <found source="vision.md:47-52">
      High-level goals mentioned
    </found>
    <gap>No quantifiable success criteria</gap>
    <gap>Timeline not specified</gap>
  </phase>

  <phase name="users_stakeholders">
    <found source="users.md:1-289">
      3 personas defined with jobs-to-be-done
    </found>
    <ambiguity>Persona priorities unclear - which is primary?</ambiguity>
  </phase>

  <!-- Continue for all 7 phases -->
</context_map>

Step 3: Present Understanding

Before asking ANY questions, present a synthesis:

## What I've Learned From Your Documentation

I've reviewed N files (X lines) from your context directory.

### Problem & Vision
> From vision.md:12-15: "exact quote from document..."

I understand the core problem is [summary]. The vision is [summary].

### Users & Stakeholders
> From users.md:23-45: "description of personas..."

You've defined N personas: [list with 1-line each].

### What I Still Need to Understand
1. **Success Metrics**: What quantifiable outcomes define success?
2. **Persona Priority**: Which user persona should we optimize for first?
3. **Timeline**: What are the key milestones and deadlines?

Should I proceed with these clarifying questions, or would you like to
correct my understanding first?

Phase 0.5: Targeted Interview

For each gap/ambiguity identified:

  1. State what you know (with citation)
  2. State what's missing or unclear
  3. Ask focused question (max 2-3 per phase)

Example:

### Goals & Success Metrics

I found high-level goals in vision.md:
> "Achieve product-market fit within 12 months"

However, I didn't find specific success metrics.

**Questions:**
1. What metrics would indicate product-market fit for this product?
2. Are there intermediate milestones (3-month, 6-month)?

Phases 1-7: Conditional Discovery

For each phase, follow this logic:

IF phase fully covered by context:
  → Summarize understanding with citations
  → Ask: "Is this accurate? Any corrections?"
  → Move to next phase

ELSE IF phase partially covered:
  → Summarize what's known (with citations)
  → Ask only about gaps (max 2-3 questions)
  → Move to next phase

ELSE IF phase not covered:
  → Conduct full discovery for this phase
  → Ask 2-3 questions at a time
  → Iterate until complete

Phase 1: Problem & Vision

  • Core problem being solved
  • Product vision and mission
  • Why now? Why you?

Phase 2: Goals & Success Metrics

  • Business objectives
  • Quantifiable success criteria
  • Timeline and milestones

Phase 3: User & Stakeholder Context

  • Primary and secondary personas
  • User journey and pain points
  • Stakeholder requirements

Phase 4: Functional Requirements

  • Core features and capabilities
  • User stories with acceptance criteria
  • Feature prioritization

Phase 5: Technical & Non-Functional

  • Performance requirements
  • Security and compliance
  • Integration requirements
  • Technical constraints

Phase 6: Scope & Prioritization

  • MVP definition
  • Phase 1 vs future scope
  • Out of scope (explicit)

Phase 7: Risks & Dependencies

  • Technical risks
  • Business risks
  • External dependencies
  • Mitigation strategies

Phase 8: PRD Generation

Only generate PRD when:

  • All 7 phases have sufficient coverage
  • All ambiguities resolved
  • Developer confirms understanding is accurate

Generate PRD with source tracing:

## 1. Problem Statement

[Content derived from vision.md:12-30 and Phase 1 interview]

> Sources: vision.md:12-15, confirmed in Phase 1 Q2
</workflow> <parallel_execution> ## Large Context Handling (>2000 lines)

If context assessment returns LARGE:

Spawn Parallel Ingestors

Task(subagent_type="Explore", prompt="
CONTEXT INGESTION: Problem & Vision

Read these files: [vision.md, any *vision* or *problem* files]
Extract and summarize:
- Core problem statement
- Product vision
- Mission/purpose
- 'Why now' factors

Return as structured summary with file:line citations.
")

Spawn 4 parallel ingestors:

  1. Vision Ingestor: Problem, vision, mission
  2. User Ingestor: Personas, research, journeys
  3. Requirements Ingestor: Features, stories, specs
  4. Technical Ingestor: Constraints, stack, integrations

Consolidate

Merge summaries into unified context map before proceeding. </parallel_execution>

<output_format> PRD structure with source tracing - see `resources/templates/prd-template.md`

Each section must include:

> **Sources**: vision.md:12-30, users.md:45-67, Phase 3 Q1-Q2
</output_format> <success_criteria> - **Specific**: Every PRD requirement traced to source (file:line or phase:question) - **Measurable**: Questions reduced by 50%+ when context provided - **Achievable**: Synthesis completes before any interview questions - **Relevant**: Developer confirms understanding before proceeding - **Time-bound**: Context synthesis <5 min for SMALL/MEDIUM </success_criteria> <uncertainty_protocol> - If context files contradict each other → Ask developer to clarify - If context is ambiguous → State interpretation, ask for confirmation - If context seems outdated → Ask if still accurate - Never assume → Always cite or ask </uncertainty_protocol> <grounding_requirements> Every claim about existing context must include citation: - Format: `> From {filename}:{line}: "exact quote"` - Summaries must reference source range: `(vision.md:12-45)` - PRD sections must list all sources used </grounding_requirements> <edge_cases> | Scenario | Behavior | |----------|----------| | No context directory | Create it, add README.md, proceed to full interview | | Empty context directory | Note it, proceed to full interview | | Only README.md exists | Treat as empty, proceed to full interview | | Contradictory information | List contradictions, ask developer to clarify | | Outdated information | Ask "Is this still accurate?" before using | | Very large files (>1000 lines) | Summarize key sections, note full file available | | Non-markdown files | Note existence, explain can't parse | | Partial coverage | Conduct mini-interviews for gaps only | | Developer disagrees with synthesis | Allow corrections, update understanding | </edge_cases>