Wicked Garden
Tell Claude to build a feature. It assembles a crew, runs enforced phases, ships clean code.
/wicked-garden:crew:start "add OAuth login with role-based access"
Claude analyzes the request, detects security and architecture signals, assembles the right specialists, and runs gated phases — clarify, design, build, test, review, operate. No skipped steps. No hallucinated shortcuts. Every decision is remembered for the next project.
claude plugins add mikeparcewski/wicked-garden
What You Do
Commands you run when you need them. No setup required.
| I want to... | Command |
|---|
| Build a feature end-to-end | crew:start "add payment processing" — assembles specialists, runs enforced phases, adapts to complexity |
| Teach Claude my codebase | smaht:onboard — indexes structure, traces flows, maps architecture. Claude remembers it across sessions. |
| Review code with domain expertise | engineering:review — security checks OWASP Top 10, architecture detects scope creep, QE catches test manipulation |
| Test before I code | qe:scenarios "checkout flow" — generates happy paths, edge cases, error conditions, maps to CLI tools |
| Brainstorm with real perspectives | jam:council "monorepo or polyrepo?" — pipes the question to every LLM CLI on your machine (Codex, Gemini, Copilot), synthesizes agreements and dissent |
| Find what breaks if I change this | search:blast-radius "handlePayment" — dependency analysis across your codebase, not string matching |
| Design an A/B test | delivery:experiment — hypothesis, sample size calculation, instrumentation plan, decision criteria |
| Collect audit evidence | platform:audit — per-control evidence collection (SOC2, HIPAA, GDPR) with PASS/PARTIAL/FAIL and auditor notes |
What the Plugin Does Automatically
These run in the background via lifecycle hooks. You don't invoke them — they make everything above smarter.
| What happens | How it works |
|---|
| Every prompt gets project context | smaht assembles memory, search results, project state, and event history before Claude responds — 3-tier routing keeps it fast for simple prompts |
| Decisions persist across sessions | mem stores what you decided and surfaces it when relevant — search "auth" and find the JWT decision from three months ago via auto-generated synonym tags |
| Working memory consolidates over time | Session noise drops away, important patterns get promoted to durable knowledge. 3 tiers: working (transient) → episodic (sprint-level) → semantic (permanent) |
| Tasks sync to a kanban board | Claude's task tools auto-sync to the kanban via hooks — no manual tracking |
| Complexity shapes the workflow | 7-dimension scoring (impact, reversibility, novelty, test complexity, documentation, coordination, operational) — a config change gets 2 phases, a migration gets 7 |
| Quality gates enforce standards | Gates actually reject work. High-complexity gates require multi-perspective consensus from specialist proposers. Bad coverage blocks advancement. |
| Traceability links deliverables across phases | Requirements → designs → code → tests → evidence → gate approvals. After deploy, incidents link back to the requirements that caused them. |
| Session teardown captures learnings | Stop hook prompts memory storage, consolidates working memories, and persists session metadata for next time |
Why This Works
Most AI coding sessions start from scratch. Session 47 has no idea what session 1 decided. Claude skips steps, hallucinates shortcuts, and ships without review.
Wicked Garden fixes that with three layers:
- Enforced workflow (
crew) — Signal analysis detects what your project needs, scores complexity across 7 dimensions, assembles the right specialists, and runs gated phases. Quality gates reject work that doesn't meet the bar. Consensus review at complexity 5+ brings multiple specialist perspectives before advancement.
- Cross-session memory (
mem) — Decisions, patterns, and gotchas persist and surface automatically. Auto-generated search tags mean "auth" finds the "JWT session token" decision. Working memories consolidate into durable knowledge over time.
- Context assembly (
smaht) — Every prompt is enriched with memory, search results, project state, and brainstorm outcomes. Three-tier routing (hot/fast/slow) keeps simple prompts fast and complex ones thorough.
How It Works
You type a prompt
|
[smaht assembles context: mem + search + kanban + crew state]
|
Claude responds with full project context
|
[hooks track decisions, sync tasks, store learnings]
|
Next session starts with everything remembered
For complex work, crew:start orchestrates the full lifecycle:
Signal Analysis → Phase Selection → Specialist Routing → Quality Gates → Memory Storage
mikeparcewski