Making templates with an AI agent

Shojiku templates are designed to be read and written by AI agents. A document is YAML and JSON text, errors come back as stable diagnostic codes, and the layout result can be read back as a tree. With these, an agent can iterate on its own: write, then check.

Register the MCP server

shojiku-mcp is a stdio server with validate / render_preview / inspect_layout / capabilities, plus list_examples / get_example for reading the bundled documents, shipped in the same Docker image as the CLI. In Claude Code, one command registers it:

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 format differ. See the quickstart's MCP section for the details.

Install the skills

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

npx skills add kengos/shojiku

The central one is shojiku-template-author. It writes a template from requirements, validates, actually looks at the preview images, and iterates until the diagnostics are clean; it carries the whole procedure plus a list of pitfalls the reference does not cover. shojiku-definitions-author handles the step before that: give it your database schema, ORM models or an API response and it produces definitions.yml plus the code that builds params from real data, checking the mapping against the engine's params_* diagnostics.

The loop an agent runs

1. list_examples — find the bundled document closest to the job, and get_example to read its source
2. Declare the data items in definitions.yml
3. Write templates.yml
4. validate — mistakes come back as diagnostic codes
5. render_preview — look at the page PNGs
6. inspect_layout — read the resolved layout back
7. Return to 3 until the diagnostics are clean

The hero banner on this site was made through this loop.

Feed the pages to an AI

Every page on this site has a plain .md version. /llms.txt is the table of contents, and /llms-full.txt carries the whole template reference — every feature page, not just the index and the diagnostic codes — so an agent asking about flex has it without a second fetch. To teach an agent Shojiku, handing it that URL is enough.