Capture Requirements

Dual-mode requirements capture: extract from conversation or elicit through semi-structured questions. Produces plans/<job>/requirements.md artifact consumed by /design and /runbook.

Target Model

Opus required. Extract mode requires synthesizing nuanced conversation context; sonnet misses implicit requirements.

Usage

/requirements <job-name>

Skill automatically detects appropriate mode based on conversation history.

Mode Detection

Primary signal: Glob: plans/<job>/requirements.md — if file exists, Read it and use as base for update/refinement (extract mode with existing artifact). If absent, fall through to conversation heuristic.

Conversation heuristic (fallback):

• Extract mode — conversation contains substantive discussion (feature descriptions, constraints, scope)
• Elicit mode — fresh conversation or minimal context (no prior feature discussion, cold-start)

Recall Pass

Runs after mode detection, before either procedure. Ground extraction in existing infrastructure, patterns, and decisions — prevents naive requirements.

Process

Invoke /recall all (deep + broad, topic-scoped). This is a skill invocation — /recall handles memory-index scanning, batch resolution, and tail-recursion. Derive topic from job name, conversation context, and any existing plans/<job>/ artifacts.

Gate anchor: The /recall all Skill invocation is the structural anchor — a tool call on both paths. When no relevant entries are found, write the recall artifact with an explicit null entry (see format below). Downstream consumers batch-resolve it via claudeutils _recall resolve null — the tool call fires on both paths without consumer-side conditional logic.

Boundaries:

• No agent delegation, no Context7, no web research — those belong to /design A.1
• Purpose: ground the extraction, not exhaustive documentation loading

Recall Artifact

After recall completes, write plans/<job>/recall-artifact.md. Entry keys only — downstream consumers resolve fresh content at consumption time.

Format:

# Recall Artifact: <Job Name>

Resolve entries via `claudeutils _recall resolve` — do not use inline summaries.

## Entry Keys

<trigger phrase> — <1-line relevance note>
<trigger phrase> — <1-line relevance note>

Null artifact (no relevant entries): Write null — no relevant entries found as the sole entry. Downstream consumers batch-resolve it via claudeutils _recall resolve null (silent exit) — the tool call anchors the gate without consumer-side empty-section handling. Augmenting consumers (/design A.1, /runbook Phase 0.5) remove the null entry when adding real ones.

Selection criteria: Include entries that informed requirements or constrain implementation. Exclude entries read but proved irrelevant — the artifact is curated, not exhaustive.

Output: plans/<job>/recall-artifact.md

Extract Mode Procedure

When conversation contains requirements signals:

1. Scan Conversation

Extract from conversation history:

• Functional requirements: What should the system do?
• Non-functional requirements: Performance, usability, constraints
• Explicit constraints: Technology, compatibility, architecture
• Scope boundaries: What's in vs out of scope
• Dependencies: Blocking work, prerequisites
• Open questions: Unresolved items mentioned

Critical rule: Extract from what was said, do NOT infer unstated requirements. Hallucination risk — if not discussed, flag as Open Question rather than inventing.

2. Lightweight Codebase Discovery

Quick scan to ground requirements (runs after extraction, so scan is targeted):

Discovery actions (guideline: ~5 tool calls, flexible based on complexity):

1. Glob for plans/*/requirements.md and plans/*/design.md — find related work
2. Grep for key terms from extracted requirements — check what exists
3. Read existing patterns in relevant area — inform constraints

Boundaries:

• Direct tool use only (Glob, Grep, Read, Bash, Write)
• No agent delegation
• No Context7, no web research, no scout
• Purpose: prevent naive requirements (e.g., "add X" when X exists)
• Tool count is a guideline, not a hard limit — use judgment for complex domains
• Save exploration prototypes to plans/prototypes/ (not tmp/) — they are referenced artifacts, not ephemera

Post-Explore Recall Gate

Discovery via Glob/Grep may surface domains not anticipated during the initial recall pass. Re-scan memory-index (already in context from recall pass) for entries relevant to areas discovered during codebase exploration.

Gate anchor (mandatory tool call on both paths):

• New entries found: claudeutils _recall resolve "when <trigger>" ... — resolve into context, update recall artifact with new entry keys
• No new entries: claudeutils _recall resolve null — no-op, proves gate was reached

3. Structure Requirements

Format into standard artifact (see Standard Format section below):

• Number items sequentially within each section (FR-1, FR-2, ...)
• Include acceptance criteria for functional requirements
• Capture rationale for constraints
• Omit empty sections entirely

4. Gap Detection

Check for empty critical sections:

Critical sections (ask if empty):

• Functional Requirements — "What should the system do?"
• Out of Scope — "What should it NOT do?"

Non-critical sections (note absence, don't ask):

• NFRs, Constraints, Dependencies — many tasks genuinely have none

Question budget (applies to both Extract and Elicit modes):

• Extract mode: Maximum 3 gap-fill questions for empty critical sections
• Elicit mode: 4-6 total questions (section questions + adaptive follow-ups)

Use AskUserQuestion tool for gap-fill:

questions: [
  {
    question: "What should the system do? (Functional requirements)",
    header: "Functional",
    options: [
      // Generate context-relevant options from job name and conversation
      // Example for job "git-hook-validator":
      {label: "Validate hook scripts before commit", description: "Parse and check hooks for common errors"},
      {label: "Block commits with invalid hooks", description: "Fail commit if hooks malformed"},
      {label: "Report hook validation errors", description: "Show clear error messages for failures"},
      ...
    ],
    multiSelect: true
  }
]

5. Present Draft

Show user the structured requirements:

• Display artifact path
• Summarize sections populated
• Note Open Questions if any
• Ask: "Does this capture your requirements accurately?"

6. Write Artifact

After user validation:

• Write to plans/<job>/requirements.md (create directory if needed)
• Scan for skill dependencies (see Skill Dependency Scanning section)
• Suggest next step (see Default Exit section)

Elicit Mode Procedure

When conversation lacks requirements context:

1. Semi-Structured Elicitation

Use AskUserQuestion for each critical section (Functional Requirements, Scope Boundaries) with domain-relevant options. Add optional Constraints/Dependencies questions if job suggests them. Adaptive follow-ups: 1-2 clarifying questions based on responses. Total budget: 4-6 questions.

2-6. Shared Steps

After elicitation, follow Extract mode steps 2-6 (Codebase Discovery, Structure, Gap Detection, Present Draft, Write Artifact).

Standard Format

# <Job Name>

## Requirements

### Functional Requirements
**FR-1: <requirement title>**
<description, acceptance criteria>

**FR-2: <requirement title>**
<description, acceptance criteria>

### Non-Functional Requirements
**NFR-1: <requirement title>**
<description>

### Constraints
**C-1: <constraint title>**
<rationale>

### Out of Scope
- <item> — <rationale>

### Dependencies
- <dependency> — <impact on implementation>

### Open Questions
- Q-1: <unresolved item — what needs investigation>

### References
- `plans/<job-name>/reports/explore-<topic>.md` — codebase exploration findings
- `plans/reports/<topic>.md` — grounding research synthesis
- [External Paper Title](url) — informed FR-2 scope decision

Section rules:

• Omit empty sections entirely (don't scaffold structure)
• Number identifiers sequentially within each section (FR-1, FR-2, ... not FR-1, FR-3)
• Open Questions captures items conversation didn't resolve → become Design Phase A research targets
• Dependencies flags blocking work, prerequisite plans, or external dependencies
• References track provenance — what research, external sources, or project artifacts informed the requirements. Include exploration reports, grounding research, external papers, prior plan reports. Omit if requirements came entirely from conversation.

Skill Dependency Scanning

After writing artifact, scan requirements text for skill dependency indicators:

Indicator	Skill to Flag
"sub-agent", "delegate to agent", "agent definition"	plugin-dev:agent-development
"skill", "invoke skill", "skill preloaded"	plugin-dev:skill-development
"hook", "PreToolUse", "PostToolUse"	plugin-dev:hook-development
"plugin", "MCP server"	plugin-dev:plugin-structure

If indicators found, append to artifact:

### Skill Dependencies (for /design)
- Load `plugin-dev:hook-development` before design (hooks mentioned in FR-2)

Rationale: Artifact-scoped scanning reduces A.0 scan time — planner reads requirements.md first, loads indicated skills immediately rather than scanning entire codebase.

Default Exit

Present artifact path and suggest next step based on completeness:

Requirements written to: plans/<job>/requirements.md

Next steps (decision criteria):
- 0-2 open questions, all critical sections populated → `/design plans/<job>/`
- 3+ open questions or empty critical sections → Standalone (revisit when clearer)
- Very clear scope + simple (Tier 1/2) → `/runbook plans/<job>/requirements.md`
- User stated explicit next step → Use that path

Workflow positioning:

/requirements <job> → /design plans/<job>/ (seeds A.0)
/requirements <job> → /handoff (document intent for later)
/requirements <job> → /runbook plans/<job>/requirements.md (direct to planning)
/requirements <job> (standalone — capture intent, no immediate follow-up)

Key Principles

• Extract from conversation, do not infer (hallucination risk)
• Gap-fill, not interrogation: max 3 questions (extract) or 4-6 (elicit)
• Omit empty sections — do not scaffold structure nobody will fill
• See references/empirical-grounding.md for research basis