A Skill That Reviews Your Agent Architecture
I recently passed the Claude Certified Architect - Professional exam (here’s the practice exam I built for it). Studying for it made me want a consistent way to check that my agents, and my team’s agents, follow industry best practices. So I built a Claude Code skill that reviews agent architectures and implementation plans against them. It combines industry best practices with what I learned studying for the certification, and it installs as a Claude Code plugin in two commands.
What it does
Give it a design doc, an implementation plan, or a codebase. It reviews the architecture, checks whether a plan delivers it, or both. If you only ask for feedback, it doesn’t touch your files.
View accessible text version
| Area | What it checks |
|---|---|
| Real system | Diagrams and docs verified against the code, dependencies, and deployment |
| Topology | Each component labeled code, model call, workflow, tool, evaluator, or agent; simplest pattern that works; every agent justified |
| Autonomy | Freedom matched to how the output is verified; error analysis before adding agents |
| Infrastructure | Hand-built checkpointing, retries, or tracing your runtime already provides; duplicated code |
| Data and memory | Sources of truth, provenance, stale versus missing data, whether a knowledge graph earns its cost |
| Runtime | Checkpoints, idempotent writes, approvals that survive restarts |
| Permissions | Server-enforced scope, untrusted content, approval before external writes |
| Tracing | One trace through model calls, tools, retries, and approvals |
| Golden dataset | Frozen evidence, labeled cases, edge cases, holdout sets |
| Evals | Component, workflow, and security evals; calibrated judges; release gates |
It leads with a recommendation to keep, revise, or replace the design. Each finding is marked satisfied, partial, missing, or not applicable, with a citation to your docs or code and dated primary sources.
How to install it
Add my plugin marketplace in Claude Code, then install the skill from it:
Then point it at a design doc or plan. A shortened example of a review:
If it gets your system wrong, open an issue.