uml-mcp: Skill for Claude Code

Diagram batches

Follow skill://uml-mcp-diagrams/SKILL.md. Build one inventory of named components, boundaries and directed labelled edges before converting formats. Keep IDs stable where possible. Compare each source against that inventory; record deliberate scope changes between C4 levels. Discover all requested types and formats, then read the appropriate focused skill.

Call generate_uml_batch with an items array. Each item uses the documented generate_uml fields: diagram_type, code, output_format, optional theme and scale. Keep transport fields such as user_intent out of items.

{"items":[{"diagram_type":"mermaid","output_format":"svg","code":"flowchart LR
  web[Frontend] -->|HTTPS| api[Backend]"},{"diagram_type":"d2","output_format":"svg","code":"direction: right
web: Frontend
api: Backend
web -> api: HTTPS"}]}

Validate complex or unfamiliar sources before the batch. Validation checks syntax, while rendering checks the actual backend; neither proves that the host has displayed a readable preview. For four C4 levels, use context/container/ component libraries and ordinary UML for code detail, as the C4 guide explains.

Inspect every returned index for errors, MIME type, bytes or a usable render link. Check each gallery entry, including labels and fit, when the host supports visual inspection. Partial success remains partial success. Retry failed items individually with corrected source and preserve successful results. Never call all items successful based only on HTTP status or valid: true.

Prefer SVG per item for Apps. Initial thumbnails share a response-size budget; selecting an item retrieves full vectors on demand. Do not treat a tiny thumbnail or a Link placeholder as proof of a full-resolution image. If full retrieval fails, report the affected item and offer its fresh link only if returned.

In an App host, confirm how many items rendered and identify failures briefly; avoid duplicating every diagram, URL and source in chat. For plain clients, provide a compact list of named, returned render/playground links and one image per requested view where supported. Do not reuse URLs from earlier failed runs.