The Code Architect

A NanoClaw template with one architecture skill and two modes:

• Architecture Planning for systems, features, refactors, and technology choices.
• Architecture Review for existing systems and codebases.

imageNanoClaw architects in a library

How it works

The coordinator reads the request, records the scope, and seats the smallest panel the evidence justifies from thirteen lenses: design principles, pragmatic engineering, systems, data, reliability, testing, algorithms, change management, security, cloud, AI, ML, and one programming-language expert per language whose semantics materially affect a risk. Every panel also seats a Tech lead, which does query-driven web research instead of using books and returns source-backed insights, technology signals, and tradeoffs rather than recommendations.

A shared Librarian selects two to four exact book editions for each book-guided expert, matched to both its field and the concrete problem, and returns book identity and fit only. The experts then apply their pretrained knowledge of those books, naming the chapters, sections, and practices used and tracing each to architecture evidence. The coordinator verifies book identity, chapter and practice association, every cited URL, and each application before writing the only user-facing prose, and separates architecture direction from production proof for every major recommendation. The result is a verified book-grounded trace, not runtime reading or proof of training-set membership; no books, copies, or fixed source lists are bundled.

Layout

skills/architecture/
├── SKILL.md                # mode selection and shared entry point
├── references/
│   ├── planning.md         # planning scope, expert output, and report sections
│   ├── review.md           # review scope, checks, and report sections
│   ├── workflow.md         # shared flow, contracts table, report skeleton
│   ├── agents/             # expert index, protocol, Librarian, Tech lead, lenses
│   └── contracts/          # JSON Schema for every record the roles exchange
└── evals/evals.json        # maintainer test material; a running agent never reads it
ai.nanoco.nanoclaw/context/instructions.md   # persona: routing and delivery rules

All references live inside the skill and use file-relative links. Agent definitions live once, so review shared changes against both modes. Only the coordinator writes user-facing recommendations. It requires an isolated Librarian and non-inheriting expert contexts and stops if the platform cannot provide them.

Stamp an agent

ncl groups create --template engineering/code-architect --name "The Code Architect"

Wire it to a channel with /manage-channels, then ask it to plan a new architecture or review an existing one.

The template has no MCP servers, scheduled tasks, paid services, credentials, hardcoded source list, or bundled books.

Created by zvi-fried.