Draw.io Architecture Studio#

Produce editable .drawio artifacts, not flattened pictures. The preferred entrypoint is scripts/diagramctl.py, which unifies generation, incremental sync, multi-view projection, semantic queries/tests/reviews, failure analysis, and accessible publishing over a shared Diagram IR.

Choose the workflow#

Unified CLI#

Run from this skill directory, or replace scripts/ with the absolute path to this skill's scripts directory:

python3 scripts/diagramctl.py doctor
python3 scripts/diagramctl.py build model.json --from ir -o architecture.drawio
python3 scripts/diagramctl.py build ./infra --from terraform --group \
  --ir-output architecture.ir.json -o architecture.drawio
python3 scripts/diagramctl.py sync architecture.drawio ./infra --from terraform \
  -o architecture.next.drawio
python3 scripts/diagramctl.py views architecture.ir.json \
  --views executive,system,deployment,dataflow,security -o views.drawio
python3 scripts/diagramctl.py test architecture.drawio --rules policy.yml
python3 scripts/diagramctl.py review architecture.drawio -o review.md
python3 scripts/diagramctl.py query architecture.drawio --from internet --to orders-db
python3 scripts/diagramctl.py whatif architecture.ir.json --fail kafka \
  --drawio kafka-failure.drawio -o impact.json
python3 scripts/diagramctl.py story architecture.ir.json -o walkthrough.html

doctor does not launch GUI tools unless --probe is passed. Core semantic commands are offline and stdlib-only.

Creation workflow#

1. Infer the diagram type, audience, scope, output format, and location from the request. Ask only when a missing choice materially changes the result; default to PNG plus .drawio in the working directory.

2. Select the authoring route from the table above. For a data-backed diagram, prefer Diagram IR and preserve provenance. For a large graph, use an importer or autolayout.py; do not hand-place more than roughly fifteen nodes.

3. Resolve an explicitly named style preset, or the user's default preset, as documented in references/style-presets.md. Structural diagram conventions and visual presets compose; they do not replace each other.

4. Generate the .drawio, then run structural validation:

   python3 scripts/validate.py diagram.drawio --score
   

   When semantic metadata or an architecture policy is in scope, also run diagramctl.py test. Do not present inferred semantic findings as verified runtime facts.

5. Export a draft PNG without embedded XML and inspect it visually. Fix obvious overlap, clipping, disconnected edges, edge-through-node routing, stacked edges, and unreadable labels. Stop automatic vision repair after two rounds.

6. Show the draft and apply targeted edits. Preserve existing geometry for local changes. Use sync for source-backed changes and write a reviewable output; use --prune only when deletion was requested.

7. After approval, create final requested formats and report both editable source and export paths.

Export invariants#

Resolve the available binary once (drawio, draw.io, the macOS app path, or the Windows executable) and use that exact binary for the run.

# Draft for visual inspection: never use -e here
drawio -x -f png --width 2000 -o diagram.png diagram.drawio

# Final editable PNG
drawio -x -f png -e -s 2 -o diagram.drawio.png diagram.drawio
python3 scripts/repair_png.py diagram.drawio.png

# Final editable SVG/PDF
drawio -x -f svg -e --embed-svg-images -o diagram.svg diagram.drawio
drawio -x -f pdf -e -o diagram.pdf diagram.drawio

Do not combine --width and -s. Embedded PNG exports require repair_png.py; draft PNGs used by vision must not use -e. On Linux headless, follow references/troubleshooting.md rather than improvising Electron flags. If the CLI crashes in a macOS sandbox, try one permitted escalated run, then use encode_drawio_url.py or deliver XML; do not repeatedly launch it.

Editing and identity#

• Use stable semantic IDs and never reuse reserved IDs 0 or 1.
• Every edge requires <mxGeometry relative="1" as="geometry"/>.
• For a local edit, change the matching cell only; for a global direction change, regenerate/re-layout the page.
• Keep provenance, data-model-id, semantic properties, manual geometry, and manual styles intact unless the user requests otherwise.
• When reconciling, retain removals as reviewable faded elements by default.
• For edges stacked at a boundary, run edgeports.py; add waypoints when an edge still crosses an unrelated shape. There is no CLI-only edge rerouter that preserves node positions.

Quality and trust#

An attractive diagram can still be wrong. Prefer source-backed relationships, show provenance where useful, distinguish exact extraction from AI inference, and keep architecture review findings framed as prompts. Story HTML must remain self-contained, keyboard usable, and include a text alternative. Never include secrets in node properties or provenance because they are embedded in outputs.

For all focused scripts and composition patterns, read references/toolbox.md; load only the task-specific reference needed for the current request.