A multi-agent skill system that turns five prompts into a compounding productivity loop. Each skill feeds the next. The system gets smarter the longer you run it.
AUDIT → ARCHITECT → ANALYST → REFINERY → COMPOUNDER → back to AUDIT
| Step | Skill | What It Does |
|---|---|---|
| 1 | /audit |
Maps your workflow, scores tasks by time cost + energy drain, builds a 4-week plan |
| 2 | /architect |
Creates implementation blueprints — approach maps, phased steps, rollback points |
| 3 | /analyst |
Reviews through 4 lenses, delivers APPROVE / REVISE / REJECT |
| 4 | /refinery |
Convergence loop — iterates until all dimensions hit 8/10 |
| 5 | /compounder |
Weekly review — tracks velocity, spots patterns, feeds back into audit |
Plus /architect-system — the orchestrator that runs the full loop or individual steps.
npx skills add vxcozy/architect-systemOr clone and use directly:
git clone https://github.com/vxcozy/architect-system.git
cd architect-system
# Open a Claude Code session — skills are auto-discoveredSee Installation Guide for all options.
Week 1: Run /audit. Map your workflow. Automate the easiest high-scoring task.
Week 2: Run /architect on your second target. Get a blueprint before building.
Week 3: Run /analyst to review what you built. Fix what's broken.
Week 4: Run /refinery on your most important output. Watch the quality converge.
Week 5: Run /compounder. The feedback loop closes and the system starts compounding.
Or skip the gradual approach: run /architect-system for the full orchestrator.
See Getting Started Tutorial for a complete walkthrough.
File-based state. Skills communicate through files in system/. No in-memory state. Pause for days and resume exactly where you left off.
Feedback bridge. The compounder writes a ## Feed to Audit section each week. The next audit reads this, incorporating real-world data into its scoring. Each loop builds on the last.
Approval gates. Every step presents results and waits for approval before writing.
See Architecture & Design for the full rationale.
| Doc | Purpose |
|---|---|
| Tutorial | Step-by-step guide to running your first loop |
| How-To Guides | Recipes for specific tasks |
| Reference | Complete specification of every skill, file, and format |
| Explanation | Architecture decisions and design rationale |
/architect-system → System status + options
/architect-system full loop → Run all 5 steps with approval gates
/architect-system resume → Continue from where you left off
/audit → Map workflow, score tasks, build plan
/architect → Design implementation blueprint
/analyst → 4-lens review with verdict
/refinery → Convergence loop until quality threshold
/compounder → Weekly review + feedback bridge
- Claude Code CLI (or any Agent Skills-compatible tool)
- Claude Pro, Max, or Team plan
MIT