Spec Driven Dev

Spec-Driven Development

Developer = spec author (defines WHAT + constraints). AI = code generator (implements HOW within guardrails).

Scope and When to Use

Use /spec-driven-dev when starting any feature larger than a trivial fix — especially multi-file changes, new modules, or features with unclear requirements. The spec becomes the contract between human intent and AI implementation.

Skip this for trivial fixes (Out-of-Loop tasks): formatting, lint fixes, import cleanup, single-line bug fixes.

Decision Loop Classification

Before starting, classify the feature using Three Loops:

• Out-of-Loop (AI autonomous): Skip spec — trivial fixes, formatting, import organization
• On-the-Loop (AI proposes, human approves): Write spec, get approval, implement with oversight
• In-the-Loop (human decides): Write spec, human drives design, AI assists implementation

Workflow

1. Understand

• Explore relevant codebase areas (Glob, Grep, Read)
• Clarify requirements with user (AskUserQuestion)
• Reference DOMAIN.md for entity definitions and invariants (if present)

2. Specify

Write spec.md in project root or feature directory using the template below.

3. Plan

• Enter Plan Mode to generate implementation plan from spec
• Plan must reference spec acceptance criteria
• Identify files to create/modify

4. Implement

• Code to spec in iterations (On-the-Loop: AI proposes, human approves)
• Each iteration should be reviewable
• Follow existing patterns from codebase

5. Verify

• Check implementation against each acceptance criterion
• Run tests (unit, integration, E2E as appropriate)
• Validate domain invariants from DOMAIN.md are preserved (if present)

---

Spec Template

Create this as spec.md (or .spec.md) in the appropriate directory:

# Spec: [Feature Name]

## Overview

- **Motivation**: Why this feature exists
- **Success criteria**: How we know it's done

## Requirements

### Functional (MUST)

- [ ] Requirement 1
- [ ] Requirement 2

### Functional (SHOULD)

- [ ] Nice-to-have 1

### Non-Functional

- Performance: [targets]
- Security: [considerations]

## Domain Impact

- **Entities affected**: (reference DOMAIN.md if available)
- **New invariants**: any new constraints introduced
- **API changes**: request/response schema changes

## Constraints

- **Anti-requirements**: what NOT to do
- **Patterns to follow**: existing codebase patterns
- **Dependencies**: external services, libraries
- **Compatibility**: backwards compatibility requirements

## Acceptance Criteria

1. Given [context], when [action], then [result]
2. Edge case: [scenario] -> [expected behavior]
3. Error case: [scenario] -> [expected error handling]

## Verification

- **Unit tests**: [what to test]
- **Integration tests**: [API/service interactions]
- **E2E tests**: [user flows, if applicable]
- **Production validation**: [how to verify in prod]

---

Governance Checklist

Before marking the spec as ready:

• [ ] Decision loop classified (Out/On/In-the-Loop)
• [ ] Domain impact assessed
• [ ] Anti-requirements defined (what NOT to do)
• [ ] Acceptance criteria are testable
• [ ] Breaking changes identified and flagged

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: pitimon
• Source: pitimon/claude-governance
• License: MIT

Install and usage instructions live in the source repository linked above.