Writing Feature Specs as Agent Skills

Create feature specifications structured as Agent Skills. The PRD becomes a skill that Claude can discover and activate, giving it contextual knowledge about the feature automatically.

Use this skill when defining what to build and why, writing product requirements, user stories, or acceptance criteria.

Supporting files: CONTEXT-LAYERS.md for layer mapping details, EXAMPLES.md for complete examples.

Why Agent Skill?

A PRD as an Agent Skill provides:

• Auto-discovery: Claude finds the spec via skill metadata without searching
• Progressive disclosure: SKILL.md body = core spec (L3), reference files = deep context (L4)
• Token efficiency: Only loaded when the feature is relevant to the conversation

Output Structure

skills/prd-{feature-name}/
├── SKILL.md              # Core spec: TL;DR, requirements, user stories, constraints
└── references/
    ├── BACKGROUND.md     # Deep reference: background, research, edge cases (L4)
    └── RULES.md          # Code-area constraints to extract to L2

Template: skills/prd-{feature-name}/SKILL.md

---
name: prd-{feature-name}
description: PRD for {Feature Name}. {1-sentence summary of capability and who benefits}. Use when implementing, testing, or reviewing code related to {feature area}.
---

# {Feature Name} - PRD

## TL;DR

[1-3 sentences. Must answer: What capability? Who benefits? Why now?
This is the most critical section — Claude reads this to decide relevance.]

## Requirements

### Functional Requirements

- **FR-1**: [Imperative statement: "The system SHALL..."]
  - Acceptance: [Testable condition]
- **FR-2**: [Imperative statement]
  - Acceptance: [Testable condition]

### Non-Functional Requirements

- **NFR-1**: [Measurable quality attribute]
  - Metric: [Threshold, e.g., "p99 latency < 200ms"]

## User Stories

- As a [role], I want [action], so that [benefit]
  - Given [precondition], when [action], then [expected result]
  - Given [alt precondition], when [action], then [alt result]

## Constraints

- [Technical: e.g., must use existing auth middleware]
- [Business: e.g., must work on mobile viewports >= 375px]

## Non-Goals

- [Explicitly excluded scope item]
- [Another exclusion with reason or link to future spec]

## Open Questions

- [ ] [Unresolved question] (@owner)
- [x] [Resolved question] -> Decision: [answer]

## Deep Reference

**Background & research**: See [BACKGROUND.md](references/BACKGROUND.md)
**Code-area constraints**: See [RULES.md](references/RULES.md)

Template: references/BACKGROUND.md

# {Feature Name} - Background & Research

## Background

[Why this feature exists. Keep under 300 words.]

## Research

[User research, data analysis, competitive analysis that
informed requirements.]

## Edge Cases

| Case | Expected Behavior | Requirement |
|------|-------------------|-------------|
| [Edge case 1] | [Behavior] | FR-1 |
| [Edge case 2] | [Behavior] | FR-2 |

Template: references/RULES.md

# {Feature Name} - Code-Area Constraints

These constraints should be extracted to L2 path-conditional rules
after the spec is approved.

## Constraints

- [Constraint derived from requirements]
- [Another constraint]

## L2 Extraction Target

Extract to `.claude/rules/{feature-name}.md` with appropriate path globs:

    ---
    paths:
      - "path/to/relevant/code/**/*.ext"
    ---

    {Feature} constraints (from skills/prd-{feature}/SKILL.md):
    - [constraint 1]
    - [constraint 2]

Writing Guidelines

TL;DR: The Gate for Skill Activation

The TL;DR combined with the skill description determines whether Claude loads this spec. Both must be:

• Self-contained: Understandable without reading anything else
• Decision-enabling: Claude can determine relevance from description alone
• Concrete: Specific capability, not vague aspiration

<!-- BAD description: Too vague -->
description: PRD for notification improvements.

<!-- GOOD description: Specific, includes trigger keywords -->
description: PRD for push notifications on order status changes (FCM/APNs). Use when implementing notification handlers, order status events, or device token management.

Requirements: Atomic and Testable

Each requirement is a unit of implementation. Agents use requirements to:

1. Determine what code to write
2. Generate test cases from acceptance criteria
3. Verify completeness

<!-- BAD: Compound, untestable -->
- **FR-1**: Handle notifications properly with good performance

<!-- GOOD: Atomic, testable, with acceptance criterion -->
- **FR-1**: The system SHALL deliver push notifications within 5 seconds
  of an order status change event.
  - Acceptance: Test device receives notification within 5s of
    status update via API call. Measured over 100 consecutive events.

User Stories: Given-When-Then for Test Derivation

Agents directly convert Given-When-Then into test cases:

- As a customer, I want to receive push notifications when my order ships
  - Given I have a confirmed order AND push notifications enabled,
    when the order status changes to "shipped",
    then I receive a notification with title "Order Shipped"
    and body containing the tracking number.
  - Given I have push notifications DISABLED,
    when the order status changes to "shipped",
    then NO notification is sent
    and the event is logged for later retrieval.

Non-Goals: Preventing Agent Scope Creep

## Non-Goals

- Email notification fallback (planned: skills/prd-email-notifications/)
- Notification preferences UI (Phase 2, not yet specced)
- Rich media notifications (images, action buttons)

Layer Mapping (Skill Edition)

Layer   What goes here                  Location
======================================================================
L1      Skill metadata (auto)           SKILL.md frontmatter
        Claude discovers via description, no manual constitution entry needed

L2      Code-area constraints           .claude/rules/{feature}.md
        Extract from references/RULES.md after approval

L3      Core spec (auto-loaded)         SKILL.md body
        Requirements, user stories, constraints, non-goals

L4      Deep reference (on demand)      references/BACKGROUND.md
        Background, research, edge cases
======================================================================

Key advantage over plain files: L1 is automatic. No need to manually add references to CLAUDE.md — the skill system handles discovery.

L2 Integration: Path-Conditional Rules

After spec approval, extract code-area constraints from references/RULES.md into .claude/rules/:

<!-- .claude/rules/notification-handlers.md -->
---
paths:
  - "internal/notification/**/*.go"
---

Notification handler rules (from skills/prd-notifications/SKILL.md):
- All notifications must go through the NotificationSender interface
- Never send notifications synchronously in HTTP handlers
- Log all notification failures with structured error context

Lifecycle

1. Create skills/prd-{feature}/ directory with references/ subdirectory
2. Write SKILL.md with template (TL;DR first)
3. Write references/BACKGROUND.md for deep context
4. Write references/RULES.md for code-area constraints
5. Review -> refine description for discoverability
6. Create technical design (writing-technical-design skill)
7. Extract L2 rules from RULES.md to .claude/rules/
8. Create implementation task commands (writing-implementation-tasks skill)
9. After completion -> remove skill or archive

Quality Checklist

Feature Spec (Agent Skill) Quality Check:
- [ ] Skill name: prd-{feature-name}, matches directory name
- [ ] Description answers what/who/when with trigger keywords
- [ ] TL;DR answers what/who/why in under 3 sentences
- [ ] Every FR has a testable acceptance criterion
- [ ] User stories use Given-When-Then format
- [ ] Non-goals list at least one exclusion
- [ ] No implementation details (those belong in technical design)
- [ ] references/BACKGROUND.md contains L4 content
- [ ] references/RULES.md contains extractable L2 constraints
- [ ] SKILL.md body under 500 lines
- [ ] Description is specific enough for auto-discovery

Detailed Guides

Context layer mapping details: See CONTEXT-LAYERS.md Complete feature spec examples: See EXAMPLES.md