Nexcanvas Drawio

NexCanvas Draw.io

Create evidence-grounded technical diagrams that remain editable, portable, and visually reviewable. Use the contract-first pipeline in this skill instead of drawing unsupported boxes directly from a thin prompt.

Select the workflow

Read exactly one primary workflow:

• New diagram from a request, repository, document, or mixed evidence: workflows/generate.md
• Update an existing repository-backed V3 project after source changes: workflows/sync-repository.md
• Repair, restyle, or extend an existing .drawio: workflows/repair.md
• Recreate a screenshot, slide, Mermaid/PlantUML result, or visual reference: workflows/convert-reference.md

Always read:

• references/intake-and-discovery.md
• references/project-contract.md
• references/semantic-model-v3.md
• references/asset-policy.md
• references/qa-contract.md

Load only the references needed for the selected route:

• Semantic intent and repository-backed work: references/semantic-intents-and-repository-evidence.md
• Route selection: references/route-catalog.md
• Software, runtime, and cloud: references/software-cloud-notation.md
• Data, behavior, security, delivery, and product: references/data-behavior-security-notation.md
• AI/ML, RAG, agents, evaluation, and governance: references/ai-ml-notation.md
• Visual system and connector rules: references/visual-system-v2.md and references/connector-label-routing.md
• Microsoft/AWS/Google reference architecture: references/enterprise-reference-style.md, references/reference-image-patterns.md, and references/provider-icon-packs.md
• Layout selection: references/layout-brainstorming.md

Intake contract

Use brief-first intake: resolve the brief and diagram language, then proceed. Select route, detail, style, orientation and canvas internally from the content. Do not offer preset type/style menus or ask how many diagram types the user wants. Ask only for missing language or a material content ambiguity. A concise brief is a decision record, not a routine approval gate. Visual references and named standards already established in the conversation remain authoritative across later examples and revisions; brief-first intake does not reset them to the default theme.

Standard project location

Unless the user chooses another path, write generated work to:

<current-working-directory>/nexcanvas-output/<project-slug>/

This is the user's repository or working directory, never the installed skill directory. The output project contains semantic contracts, local assets, the editable artifact, previews, and QA reports.

Required pipeline

1. Run nexcanvas doctor and record whether Draw.io Desktop rendering is available.
2. Infer one dominant viewIntent: architecture, workflow, sequence, data-flow, or lifecycle. Do not ask the user to choose when the brief is clear.
3. Investigate only the source evidence required by that intent. For repository-backed facts, pin Git origin/revision and exact file/line ranges; separate confirmed facts, assumptions, and exclusions in source_model.json.
4. Select one of the 48 profiles in config/route-registry.json as the specialized notation beneath the intent.
5. Build diagram_model.json schema 3.0: keep stable groups, entities, relationships, and fact provenance in semantics; keep icons, emphasis, geometry, label placement, connector lanes, and routes in presentation. Compare plausible layout alternatives before locking orientation and composition; the orchestrator persists the scored brainstorm report.
6. Confirm intent, audience, delivery target, route, notation, layout strategy, theme, canvas, source hash, and asset policy in diagram_lock.json.
7. Resolve exact official logos from verified catalogs or provider-owned packs. Use neutral native glyphs for internal concepts. Never substitute a neighboring product logo.
8. Run nexcanvas generate <project-dir> to plan, build native uncompressed mxGraph XML, run strict diagram QA, and render. Do not paste a screenshot onto a Draw.io canvas.
9. Treat exit code 3 as an external visual-review gate. Inspect the rendered image at target size and connector terminals at enlarged scale; never approve from XML or command output alone.
10. Iterate from model/layout inputs. When the current render is actually approved, rerun generate with --approve-visual, reviewer, and specific notes. Completion requires the orchestrated postflight stage.

Visual rules that cannot be waived silently

• Derive landscape, portrait, columns, rows, hub-and-spoke, or hybrid composition from content density and flow direction.
• Use boundaries only for ownership, runtime, trust, lifecycle, phase, or lane meaning.
• Route primary connectors orthogonally and attach them to shape perimeters.
• Give every independent fan-in, fan-out, and relay relationship a distinct perimeter port. A service must never look like an unlabeled continuation point between unrelated edges.
• Separate request/response, publish/consume, success/failure, and data/control when meanings differ.
• Render a relationship label only when it adds an action, payload, state, or result not obvious from the endpoints.
• Keep every label clear of nodes, icons, step badges, unrelated connector strokes, and every visible container or hub outline. Clear interior space is allowed; cutting through a border is not.
• Preserve independent icon and text bands inside icon-led nodes. Never allow a node kind, description, or manual layout override to shrink the title or caption into the icon box.
• Use opaque callouts only to break their own long rail. Never use a white label box to hide an unrelated crossing.
• Give independent flows separate lanes. Share a busId only when the diagram represents a real semantic bus.
• Embed synced SVG data so logos survive cloning and offline use.
• Never call XML-only output visually verified.

Unified CLI

Prefer the installed nexcanvas command. If the host has cloned the skill but has not installed the Python package, use python <skill-root>/scripts/nexcanvas_cli.py as the command prefix. Both invoke the same CLI and resolve bundled configuration and assets independently of the current working directory. Do not call the deprecated one-file script entry points in new workflows.

Initialize at the standard output location:

nexcanvas init --name "<title>" --brief "<user brief>" --language <language>

An explicit project directory remains supported:

nexcanvas init <project-dir> --name "<title>" --brief "<user brief>" --language <language>

Build, gate, and safely resume after completing the contracts:

nexcanvas generate <project-dir>

Only after actually viewing the preview:

nexcanvas generate <project-dir> --approve-visual --reviewer "<reviewer>" --notes "<specific observations>"

For repository-backed evidence, pass --repo-root <repo-root> on every generate invocation. Read project_state.json or the JSON command result to distinguish reused, rerun, failed, and awaiting-review stages. Do not bypass the orchestrator with stale reports when claiming completion.

For a V2 project, keep the original model and create a separate V3 candidate:

nexcanvas migrate v2-to-v3 <project-dir>/diagram_model.json --output <project-dir>/diagram_model.v3.json

Inspect and validate the candidate before adopting it as diagram_model.json. Never put coordinates, styles, icons, label placement, or connector routes back into V3 semantic records.

For a repository-backed V3 project whose source has evolved, read workflows/sync-repository.md and preview the three-way reconciliation before applying it:

nexcanvas sync <project-dir> --repo-root <repo-root> --dry-run
nexcanvas sync <project-dir> --repo-root <repo-root> --apply

Never confirm a removal without tracing its stable ID to the repository change. Sync preserves current presentation records and reports same-field conflicts instead of overwriting user edits.

If Draw.io Desktop is unavailable, produce and structurally validate the editable .drawio, leave visual approval pending, and state the limitation explicitly.

Delivery contract

For a completed project, deliver the editable .drawio, rendered preview, source/model/lock contracts, asset manifest, and QA reports. State the selected route, evidence snapshot, assumptions, output directory, and whether postflight passed. Do not claim completion while actionable warnings or unreviewed renders remain.