Working with agents

A coding agent can build a report from a conversation: it reads the component catalog, writes a render request, checks it and renders the PDF. The repository ships a skill that describes this route.

The skill

The skill is one file, skills/renderreport/SKILL.md. It says when to use renderreport, the order of the CLI calls, how a report is structured (title page, summary first, findings, appendix), how to use templates and data and renderreport from, and the mistakes that agents make most often.

To use it, copy the folder skills/renderreport into the skills directory of your agent, for Claude Code .claude/skills/ in a project or ~/.claude/skills/ for all projects. The agent needs the renderreport binary on its PATH and nothing else.

Why the loop is short

The agent does not need to know the components. Three commands carry it:

renderreport list-components                      # what exists
renderreport schema --component metric-strip      # the fields of one component
renderreport validate report.json                 # what is wrong, by component and field

validate names the component index, the field and the expected type, so the agent corrects the request and runs it again until it reports Request is valid. Then render produces the PDF, and --format png --pages 1 gives a page image the agent can look at to check what validate cannot see, such as a missing section or a bad page break.

MCP

There is no MCP server yet. The CLI covers the same steps, and every agent that can run a shell can use it. An MCP server (list_components, component_schema, validate, render) is planned only if agents without a shell turn out to need it.