Hand the whole loop to an AI agent

Shojiku templates are designed to be read, written and checked by agents. An agent works by reading and editing text — so a document is two YAML files, errors are stable diagnostic codes, and the layout can be read back as a tree.

Register the MCP server

shojiku-mcp is a stdio server exposing validate / render_preview / inspect_layout / capabilities, shipped in the same Docker image as the CLI. For Claude Code:

claude mcp add shojiku -- \
  docker run --rm -i --entrypoint shojiku-mcp \
  -v "$PWD:/work" -w /work ghcr.io/kengos/shojiku:edge

Other clients take the same command as JSON; only the config file's location and the mount spelling differ. Details in the quickstart's MCP section.

Install the skills

Four product skills ship in the repository: template authoring, definitions-from-your-schema, render debugging, and Thinreports migration.

npx skills add kengos/shojiku

shojiku-template-author is the core one: it carries the full authoring procedure — write from requirements, validate, actually look at the preview images, iterate until the diagnostics are clean — plus the trap list the reference can't fit. shojiku-definitions-author starts one step earlier: point it at your database schema, ORM models or an API payload and it derives definitions.yml plus the code that builds params from the real rows, then proves the mapping against the engine's params_* diagnostics.

The loop an agent runs

1. Declare the data items in definitions.yml
2. Write templates.yml
3. validate — fail machine-readably, with codes
4. render_preview — look at the page PNGs
5. inspect_layout — read the resolved geometry back
6. Return to 2 until clean

The hero banner on this site was authored through exactly this loop.

Feed the pages to an AI

Every page here has a raw .md twin; /llms.txt is the map and /llms-full.txt the full blob, including the reference index and the complete diagnostic-code registry. To teach an agent Shojiku, one URL is enough.