Skip to content

Latest commit

 

History

History
620 lines (448 loc) · 30.5 KB

File metadata and controls

620 lines (448 loc) · 30.5 KB

ADD (Agent Driven Development) Plugin - Product Requirements Document

Project: ADD Plugin v0.1.0 → v1.0.0 Author: abrooke Organization: MountainUnicorn Repository: https://github.com/MountainUnicorn/add Document Version: v0.9.5 Last Updated: 2026-04-22


1. Problem Statement

AI code generation has fundamentally changed how software is built, yet development practices remain ad-hoc. Developers and AI agents operate without structured collaboration frameworks, leading to:

  • Specification drift: Features are built without clear, agreed-upon requirements
  • Quality unpredictability: No consistent testing or verification practices
  • Knowledge loss: Learnings from projects aren't captured or reused across teams
  • Collaboration friction: Humans and agents work in isolation or with unclear handoffs
  • Environment blindness: Development practices don't adapt to local/staging/production contexts

ADD (Agent Driven Development) solves this by bringing structured, specification-driven SDLC practices to AI-native development—just as TDD revolutionized testing and DDD revolutionized domain modeling, ADD provides a proven methodology for human-AI collaboration at scale.

This PRD dog-foods ADD on itself: the ADD plugin is being built using ADD principles.


2. Target Users

Primary: AI-augmented development teams (1-100+ engineers)

  • Individual developers using Claude Code for daily development
  • Small teams (2-5 engineers) looking to standardize AI collaboration
  • Larger teams (10+) scaling AI agents across projects

Secondary: Open-source communities

  • Projects adopting Claude/Claude Code as primary development tool
  • Teams wanting to document and persist project-specific AI conventions

Tertiary: Enterprises

  • Organizations piloting AI-assisted development
  • Teams needing reproducible, auditable AI-driven workflows

3. Success Metrics

Metric Target Measurement
Adoption 100+ GitHub stars in Y1 Public GitHub analytics
Plugin installs 500+ Claude Code installations Marketplace metrics
Project coverage Add adopted by 20+ real projects Dogfooding + community reports
Developer feedback ≥4.5/5 stars in marketplace User reviews
Time-to-first-value ≤5 min setup for new project User interviews
Quality gates efficacy 90%+ of gate violations caught pre-deploy Project retrospectives
Knowledge persistence ≥70% learned patterns reapplied cross-project Learnings library analysis
Adoption gradient 50% adoption on legacy projects (via /add:init --adopt) Tracked adoptions

Note on telemetry-derived metrics: Several measurements above depend on the OpenTelemetry GenAI JSONL stream specified by core/rules/telemetry.md. Telemetry emission verification is a v0.9.7 milestone item; current claim reflects the spec, not yet validated dog-food output.


4. Scope

MVP Status: v0.1.0 ✓ (COMPLETE)

Core infrastructure for spec-driven development:

  • ✓ 6 commands: /add:init, /add:spec, /add:away, /add:back, /add:retro, /add:cycle
  • ✓ 8 skills: tdd-cycle, test-writer, implementer, reviewer, verify, plan, optimize, deploy
  • ✓ 10 rules: spec-driven, tdd-enforcement, human-collaboration, agent-coordination, source-control, environment-awareness, quality-gates, learning, project-structure, maturity-lifecycle
  • ✓ 10 templates: PRD, spec, plan, config, settings, CLAUDE.md, learnings, profile, library, milestone
  • ✓ 1 hooks file: auto-lint on Write/Edit
  • ✓ 2 manifests: plugin.json, marketplace.json

Delivered via:

  • Pure markdown + JSON (no compiled code, no agent-side runtime dependencies; jq required for hook scripts — see docs/runtime-dependencies.md)
  • Claude Code plugin format (.claude-plugin/plugin.json)
  • Local-only Tier 1 environment

NOT in v0.1.0 (backlog):

  • Marketplace submission flow
  • Rich interactive CLI (beyond markdown/JSON)
  • CI/CD integration (lint on commit, auto-test)
  • Sync between user-level and project-level learnings
  • Adoption wizard for legacy projects (/add:init --adopt logic)

v0.2.0: Adoption & Polish (Q1 2026)

Goal: Make ADD easy to adopt on existing projects; refine v0.1.0 based on dogfooding.

Features:

  • /add:init --adopt command with legacy project detection
  • Enhanced spec interview workflow (iterative refinement)
  • Better away/back mode context preservation
  • Integration with dossierFYI dogfooding project
  • Comprehensive user documentation + video walkthrough
  • Retro template automation (extract lessons from project)
  • Learning library cross-project search

Success: 10+ projects using ADD via /add:init --adopt


v0.5.0: Installation Reliability & Interview Safety Nets (Q1 2026)

Goal: Make ADD installable via the standard plugin command and harden the human-collaboration interview protocol.

Features:

  • Plugin installation reliability — fix claude plugin install add (currently broken, only manual clone works)
  • Interview safety nets (Nick Barger's additions to human-collaboration.md):
    • Question Complexity Check
    • Confusion Protocol
    • Confirmation Gate
    • Cross-Spec Consistency Check
  • 5 new NEVER rules for interview guardrails

Success: claude plugin install add works end-to-end; interview protocol catches complexity/confusion issues before they propagate


v0.6.0: Project Management & Documentation (Q1–Q2 2026)

Goal: Ship project visibility tooling and comprehensive user-facing documentation.

Features:

  • Project Dashboard (/add:dashboard — already implemented, 6-panel HTML report)
  • Timeline Events (spec drafted)
  • Comprehensive user documentation + video walkthrough

Success: Dashboard and timeline provide at-a-glance project health; new users can onboard from docs alone


v1.0.0: Production-Credible GA (target ~2026-07-18)

Goal: Earn the GA tag. v1.0 is a credibility cycle, not a feature cycle — the methodology surface is broad enough; v1.0's job is to make the public claim surface match reality, prove the install path actually works on both runtimes, and reposition around ADD's durable moats (maturity ladder, multi-runtime methodology, TDD-deletion guardrail).

The v1.0 plan that originally lived in this section is largely shipped (multi-environment support, advanced learnings with auto-checkpoints, CI/CD hooks, template/profile coverage all landed across v0.7.x–v0.9.x). The canonical, current scope is owned by docs/milestones/v1.0-ga.md; the synthesis that produced it — three parallel research swarms across current state, unimplemented inventory, and market review — lives at docs/v1.0-roadmap.md. The path is six releases over 8–10 weeks (v0.9.5 → v0.9.6 → v0.9.7 → v0.10.0 → v0.11.0 → v1.0.0), with seven decisions (D1–D7) locked on 2026-04-30.

v1.0 GA criteria (from .add/config.json:next_promotion_criteria, reworded per D1):

  1. Guardrail suite running in CI and release-blocking.
  2. Real Claude + Codex install smoke in CI.
  3. Per-runtime capability matrix in release notes.
  4. Beta exercise evidence: ≥3 minor or patch releases past v0.9.0 shipped with no post-tag rollback; ≥1 community PR merged during the beta cycle; AC-025 release-evidence bundle generated for the v1.0 candidate release.
  5. Marketplace submission approved.
  6. ADD verified in active use beyond the maintainer's solo work — maintainer attests via private signal (direct contact, observed adoption, adopter feedback) that ADD methodology is in production on multiple external projects. Adopter naming only with explicit consent.

See the milestone doc for per-release scope, hill chart, parallelism analysis, and risks. See the roadmap synthesis for the Part 4 capability-family architecture (what stays in ADD, what delegates to external tools, what gets added) and the M4 / v1.1 deferrals (host-neutral kernel, Cursor + Cline runtime adapters, ADD as MCP server).


5. Architecture

5.1 Tech Stack

Component Technology Rationale
Plugin format Claude Code .claude-plugin/plugin.json Fits Claude Code plugin ecosystem
Configuration JSON + YAML frontmatter Human-readable, no runtime deps
Templates Markdown + {PLACEHOLDER} syntax Portable, version-controllable
Persistence Git (local .add/ + user ~/.claude/add/) Audit trail + version control
Command interface Claude Code / command syntax Native to Claude Code
Rules engine YAML frontmatter autoload Declarative, opt-in per project
Hooks Write/Edit tool triggers via manifest Non-invasive file modification
Skills YAML metadata + markdown workflow Reusable, composable SDLC tasks

5.2 Infrastructure

Layer Component Status Purpose
Local (.add/) Project config, specs, plans, logs v0.1.0 ✓ Git-committed, shared with team
User (~/.claude/add/) Global learnings, profiles, settings v0.2.0 Machine-local, auto-initialized
GitHub Repository + releases v0.1.0 ✓ Distribution + version tracking
Marketplace Claude Code plugin registry v1.0.0 Public discoverability
CI/CD Optional: GitHub Actions linter v1.0.0 Pre-commit + pre-push gates

5.3 Environment Strategy

ADD supports tiered deployment contexts:

Tier Scope Environment Config Use Case
Tier 1 Local development only .add/config.json in .add/ No backend Plugin itself, small projects
Tier 2 Local + production staging .add/config.json + .add/env/prod.json SSH/API keys Staging CI/CD pipelines
Tier 3 Full pipeline (local → CI → deploy) Tiered configs + secrets mgmt Cloud infra Enterprise deployments

Current tier: Tier 1 (plugin itself has no backend/frontend—just markdown/JSON files in git).

5.4 Work Hierarchy

ADD structures work into four levels, governed by the project's maturity level:

Level 1: Roadmap — Where are we going? Lives in docs/prd.md milestones section. Uses Now/Next/Later framing (no fake dates). Each milestone has a name, goal, success criteria, and target maturity level. Milestones are waypoints, not timelines.

Level 2: Milestone — What's in this stage? Lives in docs/milestones/M{N}-{name}.md. Defines appetite (effort budget, not estimate), features, and a hill chart. Features progress: SHAPED → SPECCED → PLANNED → IN_PROGRESS → VERIFIED → DONE. Uphill work (figuring out) vs downhill work (executing).

Level 3: Cycle — What's the next batch of work? Lives in .add/cycles/cycle-{N}.md. A scope-boxed selection of features from the current milestone. Planned via /add:cycle command. Includes dependency assessment, parallel strategy, and validation criteria. Not time-boxed — cycles end when validation criteria are met or the human checkpoints.

Level 4: Execution — Doing the work. Existing ADD layer: /add:tdd-cycle, /add:verify, /add:deploy. Individual feature work against specs.

Maturity Cascade: The maturity level (poc → alpha → beta → ga) governs how much rigor each level requires. A POC has 1 informal milestone and no cycle planning. A GA project has full milestones with hill charts, structured cycle plans with parallel strategy, and WIP limits.


6. Key Features

6.1 Core Feature Set (v0.1.0)

Feature: Specification-Driven Development

Description: Every feature begins with a PRD-→-Spec-→-Plan document hierarchy before code is written.

Components:

  • /add:spec command: Interactive PRD/Spec generator
  • PRD template: Structured product requirements (9 sections)
  • Spec template: Detailed feature specification with test cases
  • Plan template: Implementation roadmap with estimation

Why it matters: Specification drift is prevented. All stakeholders align before engineering effort.

Example workflow:

1. Product manager: Create PRD using /add:spec command
2. Human review: Q&A in "spec interview" engagement mode
3. Agent accepts: Generates detailed Spec from approved PRD
4. Engineer reviews: Provides estimation and sign-off
5. Code: Can now only implement spec-approved features

Integration: Spec-driven rule enforces that implementations reference spec sections.


Feature: Test-Before-Implementation (TDD)

Description: Tests are written before code; quality gates ensure all code is tested.

Components:

  • /add:tdd-cycle skill: Orchestrates test-write → implement → verify loop
  • test-writer skill: Generates unit/integration tests from spec
  • implementer skill: Writes code that passes tests
  • verify skill: Semantic testing + regression detection
  • quality-gates rule: Blocks unverified code from merge

Why it matters: TDD is proven to reduce defects. Quality gates make it enforceable.

Example workflow:

1. Feature spec approved
2. Agent runs test-writer skill → generates test file
3. Agent runs implementer skill → writes code to pass tests
4. Pre-commit gate triggers verify skill
5. All tests pass → can commit; fail → blocked

Integration: Verify skill runs at 5 quality gates (pre-commit → pre-deploy).


Feature: Human-AI Collaboration

Description: Humans and agents work together via structured engagement modes; handoffs are explicit.

Engagement modes:

  • Spec interview: Iterative refinement of PRD/Spec (human leads, agent refines)
  • Quick check: Agent asks clarifying questions; human provides direction
  • Decision point: Architectural/design choice; human decides, agent implements
  • Review gate: Human reviews artifact (spec, plan, PR); agent awaits approval
  • Status pulse: Regular checkins on project health

Commands:

  • /add:away: Agent pauses; human takes over (can modify files freely)
  • /add:back: Agent resumes; catches up on changes human made
  • /add:retro: Retrospective meeting; team + agent reflect on learning

Why it matters: Prevents agent hallucination. Ensures human judgment on high-stakes decisions.

Example workflow:

1. Agent generates implementation plan
2. /decision-point: Agent asks human for architectural guidance
3. Human provides decision (async, via message)
4. Agent implements decision
5. /add:away: Human needs to focus; agent pauses
6. Human modifies code, docs, configs
7. /add:back: Agent re-reads changes, identifies learnings
8. Agent resumes with updated context

Integration: Commands map to rules (human-collaboration, agent-coordination).


Feature: Environment Awareness

Description: Development practices adapt to the target environment (local → staging → prod).

Components:

  • Environment config: Tier-based settings (local, staging, production)
  • Deployment skill: Moves code through environments
  • Quality gates: Different rules per environment (e.g., stricter pre-prod gates)

Why it matters: Prevents "works on my machine" bugs. Ensures prod safety.

Example:

Local environment: Unit tests only, fast feedback
Staging: Integration tests + security scans
Production: Full test suite + approval gates

Integration: Environment-awareness rule checks .add/env/ configs; quality-gates rule adapts per tier.


Feature: Structured Project Initialization

Description: /add:init command sets up project structure, rules, and conventions.

Components:

  • .add/config.json: Project metadata (name, tier, rules, team)
  • .add/settings.json: Team conventions (code style, test framework, etc.)
  • .add/learnings/library.json: Persistent patterns learned on project
  • .claude-plugin/ rules autoload in project context

Why it matters: No manual setup. Conventions are enforced from project start.

Example:

$ /add:init
? Project name: dossierFYI
? Environment tier: Tier 1 (local only)
? Rules to enable: [all defaults]
? Test framework: pytest

→ Creates .add/config.json, .add/settings.json, rules autoload
→ Ready to start spec-driven development

Integration: Source-control rule ensures .add/ is git-committed; project-structure rule validates layout.


Feature: Learning System

Description: Agents automatically checkpoint learnings during development; humans reflect during retros. Learnings persist across projects.

Components:

  • Agent auto-checkpoints: During /add:away//add:back, /add:retro, on spec approval
  • Learnings template: Structured format (decision, rationale, context, pattern, reusability)
  • Learnings library: Project-level + user-level (machine-local)
  • Cross-project search: Find similar patterns across projects

Why it matters: Prevents repeating mistakes. Builds institutional knowledge.

Example:

Project dossierFYI:
- Learning: "PostgreSQL UUID columns must be type uuid, not text"
  Context: dossierFYI database schema
  Reusability: HIGH (all SQL projects)

Later, on project X:
- Agent searches learnings: "uuid column"
- Finds dossierFYI pattern, applies it automatically

Integration: Learning rule enforces retro-based reflection; project-structure rule persists learnings.


Feature: Quality Gates

Description: 5 checkpoints ensure code quality before merge/deploy.

Gate Trigger Checks Enforcer
Pre-commit Before Write/Edit auto-lint hook Lint, type check Auto (hooks)
Pre-push Before git push All tests pass Manual (human review required)
CI On push to main Full test suite + security scan GitHub Actions (v1.0.0)
Pre-deploy Before deploy to prod Regression tests + approval Manual (deployment skill)
Post-deploy After deploy to prod Smoke tests + monitoring Manual (deployment skill)

Why it matters: Catches bugs early. Ensures prod stability.

Example:

1. Agent implements feature
2. Pre-commit gate: Lint + type checks (auto, must pass)
3. Human pushes to main
4. Pre-push gate: Requires manual review + test sign-off
5. CI gate: Full test suite (GitHub Actions)
6. If deploying to prod: Pre-deploy gate (approval required)

Feature: Non-Greenfield Adoption (/add:init --adopt)

Description: Make ADD accessible to existing projects via gradual adoption.

Components (v0.2.0):

  • Adoption detection: Scan project structure, identify test framework, linter, etc.
  • Gradual enablement: Start with "monitoring" mode, gradually adopt rules
  • Legacy compatibility: Map existing conventions to ADD templates
  • Migration guide: Walk team through adoption process

Why it matters: Lowers barrier to entry. Existing projects can adopt without massive refactor.

Example:

$ /add:init --adopt
? Detected test framework: pytest ✓
? Detected linter: black ✓
? Adopt spec-driven rule? (Recommended for new features) [Y/n]

→ Enables spec-driven rule for *new* features only
→ Existing code untouched
→ Team can gradually migrate at own pace

6.2 Commands (v0.1.0)

Command Aliases Purpose Output
/add:init /setup Initialize project with ADD infrastructure .add/ directory + config files
/add:spec /spec-interview, /spec-write Start spec-driven development workflow PRD template → Spec template
/add:away /pause, /hands-off Pause agent; human takes control Saves context checkpoint
/add:back /resume, /hands-on Resume agent; catch up on changes Loads context, identifies deltas
/add:retro /retrospective, /learnings Team retrospective + learning capture Learnings template + library update
/add:cycle /sprint, /plan-cycle Plan and execute a work cycle .add/cycles/cycle-{N}.md

6.3 Skills (v0.1.0)

Skill Triggers Input Output Integration
tdd-cycle Manual + /add:spec approval Spec Tests → Code → Verified spec-driven, tdd-enforcement
test-writer tdd-cycle, manual Spec → Test framework Test file with full coverage tdd-enforcement, quality-gates
implementer tdd-cycle, manual Tests + Spec Implementation code tdd-enforcement, source-control
reviewer Manual + pre-push gate Code + Spec Review checklist + feedback human-collaboration, quality-gates
verify Pre-commit, pre-push, CI, pre-deploy, post-deploy Code + Tests Pass/fail + coverage report quality-gates, tdd-enforcement
plan /add:spec, manual Spec Implementation plan + estimation human-collaboration, learning
optimize Manual + post-retro Code + Learnings Optimized code + rationale learning, source-control
deploy Manual Code + Environment Deploy to env + smoke tests environment-awareness, quality-gates

6.4 Rules (v0.1.0)

Rule Applies to Enforces Blockable
spec-driven All code commits Implementation must reference approved spec section Yes (pre-commit)
tdd-enforcement All code commits Tests must exist before code; 80%+ coverage required Yes (pre-commit)
human-collaboration All decisions Architectural/design decisions require human sign-off Yes (manual)
agent-coordination All agent outputs Agent actions logged; human can pause with /add:away Yes (/add:away)
source-control Project init + commits .add/ directory always committed; clean commit messages Yes (pre-push)
environment-awareness Deployment Code built/tested for target environment Yes (pre-deploy)
quality-gates All commits + deploys 5-gate enforcement (pre-commit → post-deploy) Yes (all gates)
learning Specs, /add:away, /add:back, /add:retro Decisions logged; patterns captured; learnings persist No (auto-logging)
project-structure Project init Consistent .add/ layout; templates valid Yes (init)
maturity-lifecycle All ADD behavior Cascades maturity level into process rigor, quality gates, parallelism Yes (all rules)

6.5 Templates (v0.1.0)

Template Purpose Sections Reusable
PRD.md Product requirements Problem, users, success metrics, scope, architecture, features, NFRs, questions, history Yes (all projects)
spec.md Feature specification Overview, requirements, test cases, edge cases, architecture, success criteria Yes (all features)
plan.md Implementation plan Milestones, tasks, estimation, risks, success criteria Yes (all features)
config.json Project metadata name, tier, rules, team, repo, contact Project-specific
settings.json Team conventions test-framework, linter, code-style, doc-format Reusable (team profile)
CLAUDE.md Agent guidelines Context, rules, conventions, tools, knowledge cutoffs Reusable (team profile)
learnings.md Captured pattern Decision, context, rationale, pattern, reusability Global (library)
profile.json Team conventions + learnings Bundled settings.json + learnings.json for easy adoption Reusable (profiles)
library.json Learnings index Index of all learnings (tags, search metadata) Global (org-wide)
milestone.md Milestone tracking Goal, features, hill chart, cycles, retrospective Yes (all projects)

7. Non-Functional Requirements

7.1 Performance

  • Init time: < 1 second (JSON reads + template setup)
  • Spec interview: < 30 seconds per response (conversational flow)
  • Plan generation: < 2 minutes (estimation + task breakdown)
  • Verify gate: < 5 seconds (lint + type check)
  • Learning search: < 2 seconds (library scan)

7.2 Reliability

  • Plugin crash rate: < 0.1% (pure markdown/JSON, no runtime errors)
  • Lost context: 0% (all state in git-committed .add/ + ~/.claude/add/)
  • Command availability: 99.9% (no external dependencies)

7.3 Maintainability

  • Code-free design: No compiled code, no agent-side runtime dependencies, pure markdown + JSON (jq required for hook scripts — see docs/runtime-dependencies.md)
  • Template portability: All templates use {PLACEHOLDER} syntax (portable across projects)
  • Version compatibility: Backward-compatible minor versions; migration guide for major versions
  • Documentation: README + getting started guide + per-command docs

7.4 Security

  • Secrets handling: No secrets stored in templates; environment variables for sensitive data
  • Access control: Project rules can restrict who can approve specs/plans (v1.0.0)
  • Audit trail: All decisions logged in .add/ (git history = audit log)

7.5 Scalability

  • Team size: 1-100+ engineers (no backend bottleneck)
  • Project count: Unlimited (each project is independent .add/ directory)
  • Learnings library: 1000+ entries searchable (JSON scan, < 2s)

7.6 Usability

  • Onboarding: First spec written in < 5 minutes (guided /add:spec command)
  • Commands: ≤ 5 top-level commands; all have aliases
  • Error messages: Clear, actionable (e.g., "spec-driven rule: Implementation must reference spec section; missing reference to spec.md#3.2")

6.6 Maturity Lifecycle (v0.1.0)

ADD projects declare a maturity level that governs all process rigor:

Level PRD Specs TDD Quality Gates Parallel Agents Cycle Planning
poc Paragraph Optional Optional Pre-commit only 1 (serial) Informal
alpha 1-pager Critical paths Critical paths Pre-commit + CI 1-2 Brief doc
beta Full template Required Enforced Pre-commit + CI + pre-deploy 2-4 Full plan
ga Full + architecture Required + acceptance criteria Strict All 5 levels 3-5 (worktrees) Full + risk assessment

Maturity promotion is deliberate — triggered via /add:cycle --complete or /add:retro with gap analysis showing readiness for the next level.


8. Open Questions

Question Impact Status Owner
Should learnings auto-sync between user-level (~/.claude/add/) and project-level (.add/)? Or manual sync via command? System design v0.2.0 abrooke
What's the minimum viable CI/CD integration? (GitHub Actions lint only? Full test suite?) v1.0.0 scope v1.0.0 abrooke
Should quality gates support "override" with human approval (e.g., skip test for urgent hotfix)? Security/agility tradeoff Decision pending abrooke
How to make /add:init --adopt work for truly alien project structures (e.g., monorepos, multilang)? Adoption complexity v0.2.0+ abrooke
Should profiles be versioned? (e.g., "use Profile v2.0 on this project") Long-term maintainability v1.0.0 abrooke
Multi-agent coordination: How do agents coordinate across multiple concurrent specs/plans? Future roadmap Addressed in v0.1.0 — swarm coordination protocol in agent-coordination rule + /add:cycle command abrooke
Context window optimization: 10 autoloaded rules consume ~14K tokens (7% of 200K). Should heavy rules (agent-coordination, maturity-lifecycle) use conditional loading or progressive disclosure? Performance v0.2.0 — dog-food first, optimize based on real usage patterns abrooke
Cowork vs Claude Code interplay: How should ADD bridge Cowork (human decisions, reviews) and Claude Code (agent execution, TDD)? Shared state via .add/ directory? Architecture v0.2.0 abrooke

9. Revision History

Version Date Author Changes
0.1 2026-02-07 abrooke Initial PRD; describes v0.1.0 (complete) + v0.2.0 + v1.0.0 roadmap
0.2 2026-02-07 abrooke Added maturity lifecycle, work hierarchy (roadmap→milestone→cycle→execution), swarm coordination, /add:cycle command

Appendix A: Dog-Fooding This PRD

This PRD was created using ADD methodology:

  1. Spec-driven: This document is the PRD for the ADD plugin. It's the single source of truth for requirements.
  2. Test-before-impl: Test cases (e.g., "v0.1.0 should have 5 commands") are embedded in scope section; implementations verified against them.
  3. Trust but verify: Human (abrooke) approved problem statement; agent (Claude) refined architecture/features; human reviewed final PRD.
  4. Structured collaboration: PRD created via /add:spec workflow (hypothetically, once plugin is live).
  5. Environment awareness: Tier 1 environment appropriate (plugin = markdown/JSON, no backend).
  6. Continuous learning: Learnings from building ADD will be captured in /add:retro and applied to v0.2.0.

Appendix B: Success Criteria for v0.1.0 Completion

Historical record. These were the launch criteria for v0.1.0 (2026-02-07). All were met. They describe the v0.1.0 surface area, not current state — today's plugin (v0.8.0) ships 26 skills, 15 rules, 21 templates, 2 hook scripts, and supports both Claude Code and Codex CLI runtimes. See README.md and /docs/skills for the current reference.

  • ✓ 5 commands implemented and working
  • ✓ 8 skills implemented and composable
  • ✓ 9 rules enforcing ADD principles
  • ✓ 9 templates available for all core artifacts
  • ✓ 1 hooks file (auto-lint on Write/Edit)
  • ✓ 2 manifests (plugin.json, marketplace.json) valid for Claude Code
  • ✓ Dogfooding on dossierFYI project (real project uses ADD for development)
  • ✓ This PRD published and approved

Appendix C: Glossary

Term Definition
ADD Agent Driven Development—structured SDLC methodology for AI-native teams
Agent Claude (or other AI assistant) running under ADD rules + skills
Engagement mode Structured interaction pattern (spec-interview, quick-check, decision-point, review-gate, status-pulse)
Quality gate Checkpoint that must pass before code advances (pre-commit, pre-push, CI, pre-deploy, post-deploy)
Rule Declarative constraint enforced by plugin (e.g., spec-driven, tdd-enforcement)
Skill Reusable SDLC workflow (e.g., tdd-cycle, plan, verify)
Tier Environment context (Tier 1 = local, Tier 2 = local+prod, Tier 3 = full pipeline)
Dogfooding Using ADD on real projects (e.g., dossierFYI) to validate design
Learning Captured pattern/decision + metadata; persists in library
Profile Bundled team conventions + learnings for easy adoption

END OF DOCUMENT