Give AI room to create without making the build probabilistic
Clodsite is an experiment in dividing work according to what each tool does well.
AI is good at understanding goals, finding themes in source material, drafting language, and exploring alternatives. Scripts are good at enforcing contracts, repeating transformations, tracking files, and talking to deployment APIs.
Clodsite puts each on the side of the boundary where it belongs.
The principles
The plan is the product contract
The approved site is visible in one human-readable document. There is no hidden prompt state required to rebuild it.
AI proposes; people review; scripts execute
The customer remains in the loop where meaning and presentation are decided. Approval is a real handoff, not a ceremonial click before more invisible generation.
Components are capabilities, not decoration
A component says what a section is for and what information it requires. The catalog can grow without turning every build plan into handwritten frontend code.
Deployment belongs in the workflow
A website is not finished when files appear in dist/. Hosting, secrets, bot protection, domains, status, and teardown are part of the system.
The agent is replaceable
Claude Code commands, Codex collaboration, direct YAML editing, and MCP designer agents can all meet at the same build boundary.
Small enough to understand
Clodsite is built from Bash, Node.js, Nunjucks, Eleventy, Wrangler, and a catalog of local components. Its behavior lives in scripts and files you can inspect.
That modest architecture is a feature. It makes the project useful as a working tool and as a laboratory for questions that matter beyond websites:
- What should an AI-native build contract look like?
- Where should inference end and deterministic execution begin?
- How can an agent discover a constrained capability vocabulary?
- How do people review and govern generated artifacts?
- When does a local workflow become a callable service?
Clodsite is building answers one shipped site at a time.