// the find
Agents365-ai/drawio-skill
Agent skill that turns natural language, code, Terraform/K8s, SQL, OpenAPI, AsyncAPI, Protobuf and GraphQL sources into editable, tested draw.io architecture diagrams: incremental sync, multi-view projection, drift diff, CI architecture tests, whiteboard derasterize, interactive HTML/PPTX/Mermaid exports.
A Claude/Cursor/Copilot 'agent skill' (SKILL.md + a big Python script toolbox) that turns prompts, code, Terraform/K8s, SQL, OpenAPI/AsyncAPI, Protobuf and GraphQL into editable .drawio files, with CI linting and drift diffing on top. Aimed at teams who want an AI agent to keep architecture diagrams in sync with actual infra/code instead of hand-drawing them once and letting them rot.
The import side is the real value: 16 extractors (Terraform, K8s, docker-compose, SQL DDL, OpenAPI, AsyncAPI, Protobuf, GraphQL SDL, language import graphs) that produce diagrams from ground truth rather than an LLM guessing boxes and arrows. Layout and sequence/C4 generation are deterministic (Graphviz placement, computed lifelines) rather than left to the model, which is what actually makes diagrams reproducible. The CI story is concrete and shippable: a GitHub Action that enforces architecture rules (internet-to-db access, cycles, trust boundaries) and renders PR diagram diffs, not just a doc describing how you could do that yourself. Core workflows are stdlib-only and work offline with no daemon, so adoption cost for the non-LLM parts is low.
Everything renders through the draw.io desktop CLI, an external GUI binary — headless CI needs xvfb, some features require version >=30, and that's a fragile dependency for something billed as CI-gateable. The 'self-check PNG, auto-fix overlaps, 5-round feedback loop' pipeline is LLM-driven heuristics layered on top of the deterministic core, so diagram quality for anything beyond the scripted importers is as consistent as the model session that day, and that part is hard to regression-test. The surface area is enormous for what's essentially a single-maintainer skill file — dozens of scripts, a 10k+ shape index, 321 bundled brand logos — which is a lot to keep correct as draw.io's own shape library and AWS/Azure icon sets change versions. The README leans hard into comparison tables against competing skills, which is a signal to actually read the code rather than trust the feature checklist at face value.