Author D2 diagrams (the Terrastruct text-to-diagram language at https://d2lang.com). Use whenever the user wants to produce, edit, or render any kind of diagram with D2 — architecture, flowchart, sequence, ER/SQL schema, UML class, network topology, mind map, deployment, org chart, state machine, grid layout, mermaid-style flow, or anything similar — even when they don't say "D2" explicitly. Trigger on requests for `.d2` files, mentions of `d2`/`d2lang`, or open-ended diagramming tasks ("draw me a system diagram", "make a flowchart of X", "diagram the request flow", "give me an ER diagram for these tables"). Covers the full DSL — shapes, connections, containers, classes, vars, globs/filters, imports, layers/scenarios/steps, styling, themes, sketch mode, icons, markdown/LaTeX/code blocks, layout engines (dagre/elk/tala), CLI flags, and the specialized diagram shapes (sequence_diagram, sql_table, class, grid, hierarchy, c4-person, image, text/code).
Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Une commande directe contourne le prompt de vérification. Examinez la source avant de l'exécuter.
Author D2 diagrams (the Terrastruct text-to-diagram language at https://d2lang.com). Use whenever the user wants to produce, edit, or render any kind of diagram with D2 — architecture, flowchart, sequence, ER/SQL schema, UML class, network topology, mind map, deployment, org chart, state machine, grid layout, mermaid-style flow, or anything similar — even when they don't say "D2" explicitly. Trigger on requests for `.d2` files, mentions of `d2`/`d2lang`, or open-ended diagramming tasks ("draw me a system diagram", "make a flowchart of X", "diagram the request flow", "give me an ER diagram for these tables"). Covers the full DSL — shapes, connections, containers, classes, vars, globs/filters, imports, layers/scenarios/steps, styling, themes, sketch mode, icons, markdown/LaTeX/code blocks, layout engines (dagre/elk/tala), CLI flags, and the specialized diagram shapes (sequence_diagram, sql_table, class, grid, hierarchy, c4-person, image, text/code).
D2 — Authoring Diagrams
D2 is a declarative text-to-diagram language. You write a .d2 file; the d2 CLI compiles it to SVG/PNG/PDF/PPTX/GIF. Most files are short — five to fifty lines — and lean on layout engines (dagre/elk/tala) to position things automatically.
This skill is your reference for writing D2 well. The body below is the working overview; the detailed grammar and per-diagram-type references are in references/. Hand-edit examples are in assets/examples/.
When to reach for which diagram type
Pick the shape that actually matches the data, not the one that's flashy. D2 has eight specialized diagram modes plus the general-purpose flowchart mode that does everything else.
Architecture / system / flow / mind map / org chart / state machine → general D2 (default rectangle/cloud/cylinder/etc. shapes connected with arrows). This is the bread and butter; ~80% of D2 you write.
Time-ordered messages between actors → shape: sequence_diagram (alice → bob → carol with messages).
UML class → shape: class (fields + methods with visibility prefixes).
Anything that needs grid layout → grid-rows: N / grid-columns: N on the container.
Tree from connections → shape: hierarchy.
Image-only node (no rectangle around it) → shape: image.
Plain text label / code block / LaTeX → shape: text, shape: code, or a |md ... | / |tex ... | block.
If the user asks for a "Mermaid flowchart" or "Graphviz dot" or "PlantUML sequence diagram", D2 can produce the same diagram — just tell them you'll use D2's equivalent.
Writing a D2 file: the minimum to know
# Comments start with '#'
direction: right # up | down | left | right (default: down)
# Shapes are declared by mentioning them, optionally with a label
api: API Gateway # `api` is the key, "API Gateway" is the label
db: { shape: cylinder } # set shape via attributes
worker.shape: queue # dot syntax also works
# Connections — these IMPLICITLY create the shapes if not already declared
api -> db: writes # directed edge with label
worker <- queue # reverse direction
client <-> server # bidirectional
metric -- log # undirected
# Chained connections
client -> api -> worker -> db
# Containers (nesting)
prod: {
api -> db
worker -> db
}
prod.api -> staging.api: replicate # cross-container reference
# Styling — quote hex colors (unquoted '#' starts a comment!)
api.style.fill: "#4baae5"
api.style.stroke-width: 2
(api -> db)[0].style.stroke: red # style a specific connection
That's enough to draft most diagrams. The rest is refinement.
Identifiers and labels
Spaces are allowed in keys: web server.shape: hexagon — web server is one identifier.
Dots nest: aws.vpc.subnet.ec2 creates a 4-deep hierarchy.
Quote when needed: use double quotes for keys with . in them, single quotes to bypass ${var} substitution and special meanings: 'literal text'. Backticks are not used.
Multiline labels: use \n (api: my\nmulti-line\nlabel) or a |md ... | block.
Labels can be Markdown: desc: |md\n # Title\n - bullet 1\n - bullet 2\n|.
Reserved keywords (cannot be used as plain shape names without being attributes): label, shape, icon, near, top, left, width, height, style, class, classes, vars, layers, scenarios, steps, direction, link, tooltip, constraint, source-arrowhead, target-arrowhead, grid-rows, grid-columns, grid-gap, vertical-gap, horizontal-gap. To use one literally as a shape name, escape with a leading \ or quote it.
The eight shape: values that change layout/parsing rules
These have special handling. The rest (rectangle, square, page, parallelogram, document, cylinder, queue, package, step, callout, stored_data, person, diamond, oval, circle, hexagon, cloud, c4-person) only change appearance.
Shape
What changes
sequence_diagram
Whole container is laid out as actor lanes + time arrows. See references/sequence-diagrams.md.
sql_table
Children become rows (column-name: type with optional {constraint: primary_key}). See references/sql-and-er-diagrams.md.
class
Children become UML fields/methods. Visibility via + - # prefix. See references/class-diagrams.md.
hierarchy
Tree layout from connections. Position keywords forbidden. See references/grid-diagrams.md.
image
Borderless icon-only node. Pair with icon: <url>.
text
Plain text node, no surrounding box.
code
Code block (use style.font: mono and `
grid-rows: N / grid-columns: N
Children flow into a grid in declaration order. See references/grid-diagrams.md.
Authoring workflow
Pick the diagram type (see table above). For ambiguous requests, draw it as a general flowchart — that's what most of D2 is.
Sketch the structure first: which shapes exist, what connects to what. Don't style yet.
Render it AND look at it. This is the most important step — see "Render and review" below. Compilation success ≠ a good diagram. D2 will happily emit a syntactically valid SVG that's visually broken.
Iterate based on what you actually see. Tight feedback loop. If something looks wrong, don't ship it — fix it and re-render.
Set direction: if the auto-layout direction is wrong. Default is top-to-bottom; direction: right is the most common override (good for sequence/process flows).
Pick a layout engine if defaults look bad:
dagre (default) for plain hierarchical flowcharts.
elk for SQL/UML/anything with deep nesting and cross-container edges, and for crisp orthogonal routing.
tala (paid) only when the user explicitly asks or the diagram needs near: <other-shape> or per-container direction:.
Apply theming in one place via vars.d2-config rather than scattered style.* overrides. Use semantic theme color codes (B1–B6, AA2/4/5, AB4/5, N1–N7) so themes still work.
Add icons sparingly from https://icons.terrastruct.com for AWS/GCP/Azure/dev/tech categories — they elevate architecture diagrams.
Use classes: if you find yourself repeating styles more than twice. Don't preemptively define classes.
Render and review (DO NOT SKIP)
D2 errors loudly when syntax is wrong, but stays silent when the diagram is visually broken. The compiler will happily produce a diagram with overlapping text, truncated labels, mis-aligned cards, orphaned containers, illegible contrast, or comic spacing — and report success. So you have to look.
Workflow:
Render to PNG (easiest to view inline): d2 input.d2 input.png. PNG renders through headless Chromium so it requires Playwright/Chromium to be installed; SVG is a fallback that works without.
Open the rendered file and actually look at it. Use the Read tool on the PNG — it returns the image inline. For SVG, render PNG too just for the review step (the production output stays SVG).
Run the visual-quality checklist below.
If anything is wrong, fix the source and re-render. Do not declare done until the rendered image meets the bar. Two or three iterations is normal.
Visual-quality checklist — scan the image for each of these and fix what's broken:
Truncation. Any label cut off mid-word? ("Personalizatio…", "Device mg…"). Usually a card-width problem — widen the card via width: on the class, or shorten the label.
Alignment. In a row of cards, are they all the same height? D2's auto-sizing makes uneven rows by default. Set explicit height: on the class to force uniform.
Overflow. Is text spilling out of its box? |md blocks especially can render text outside the box bounds. Use a fixed width: and height: to constrain, or shrink font-size.
Empty space / awkward gaps. A row with 4 items in a 7-wide grid leaves three blank cells. Either re-center, re-grid, or fill with invisible padding (shape: text with style.opacity: 0).
Connection routing. Are edges crossing through unrelated shapes? Try --layout=elk (orthogonal) instead of dagre. For complex architecture, ELK is almost always better. If edges still cut diagonally across content under ELK, the cause is usually a grid-rows: / grid-columns: container — see pitfall 15 and references/edge-routing.md.
Text readability. At the diagram's natural scale, can you read every label? font-size: 8 is the floor; below that, increase the box size or split content.
Border visibility. Is the stroke visible against the fill? stroke-width: 1 on a light-on-light combo disappears; bump to 2–4.
Color contrast. Is white text on a light theme color invisible? Theme color codes (B1 darkest → B6 lightest) are designed to pair text/fill correctly.
Visual hierarchy. Does the most important thing draw the eye first? If everything is equally bold/colored, nothing stands out.
If you're tempted to skip the render step because "the syntax is correct": don't. That's the failure mode this section exists to prevent. The user will be looking at the image, not at the source.
A clean default vars.d2-config block
When you produce a polished diagram, put this near the top — it sets sensible defaults and makes the output theme-aware:
vars: {
d2-config: {
layout-engine: elk # crisp orthogonal routing, good for most cases
theme-id: 0 # Neutral default; see references/styling-and-themes.md for the full list
pad: 40 # tighter than the 100px default
sketch: false # set true for hand-drawn whiteboard look
}
}
CLI flags (--theme=N, --sketch, --layout=elk) override these. Don't set sketch: true by default — the user usually wants the polished version.
Rendering D2 to a file
If d2 is available on the user's machine (or in this sandbox), render directly:
Or offer to install with brew install d2 (mac) / curl -fsSL https://d2lang.com/install.sh | sh -s -- (linux/mac) / Windows MSI.
When you do produce a rendered file, prefer SVG unless the user explicitly needs PNG/PDF/PPTX. SVG keeps tooltips, links, animations, and infinite zoom; PNG strips all of that and needs a headless Chromium.
Common pitfalls (these will bite you)
The full list is in references/full-grammar-cheatsheet.md. The top hits:
Unquoted # is a comment.style.fill: #4baae5 is broken. Use style.fill: "#4baae5".
-> not --> for connections (no Mermaid-style double-dash).
There's no alt/opt/loop/par keyword in sequence diagrams — those are just nested groups whose key happens to be named alt etc. See references/sequence-diagrams.md.
Class diagrams' # (protected) visibility collides with comments — escape as \# or use the # prefix without alphanumerics on the next char.
fill is not allowed on connections. Use stroke for connection color.
Inside a sequence_diagram, all actors must be predeclared at the top of the container if you want to reference them in nested groups.
Single quotes prevent ${var} substitution. Double quotes / unquoted text expand it. Surprising if you're escaping for shell reasons.
(a -> b)[0] indexes the first edge between a and b. Multiple a -> b lines create distinct edges, not overrides — use the index syntax to restyle one of them.
near: <shape-id> is TALA-only. Cardinal near: top-center works in any engine.
|md blocks ignore card box styling in practice. Even though |md is documented as defaulting to a rectangle, applying style.fill / style.stroke / style.border-radius / style.shadow via a class on a |md shape often does nothing visible — the box is invisible behind the markdown rendering. For boxed content with bullets, use shape: rectangle with a \n-separated label (e.g. iam: "Identity & Access\n\n• Authentication\n• SSO\n• User management" { class: card }). Save |md for situations where you actually want rendered Markdown text without a surrounding box.
Grid rows of unequal-content children render at unequal heights. A row of cards where one has 5 lines of text and another has 2 will produce different-height boxes. Set explicit height: on the class so every card in the row is the same size.
Multiple top-level grids lay out horizontally as siblings. If you want stacked rows, wrap them in a parent container with grid-columns: 1 (or grid-rows: N). Otherwise dagre/elk treats each grid as a peer and lines them up sideways.
Hide intermediate-container labels with label: "" when you're using nested grids structurally. Otherwise the row containers' names ("domains", "concerns") show up as visible labels in the rendered diagram.
grid-rows: / grid-columns: on a connected container breaks ELK's orthogonal routing. When grid children connect to shapes outside the grid, ELK silently falls back to diagonal/polyline routes for those cross-grid edges, producing the spaghetti-line look the engine is supposed to prevent. Drop the grid (let ELK arrange siblings itself) or connect to the parent container instead of individual children. See references/edge-routing.md.
Per-edge style.border-radius is the curve lever, and it scales past the documented max. Docs say 0–20, but values up to ~100 work and progressively soften corners. border-radius: 20 is the architectural-soft sweet spot; border-radius: 50–100 reads as a Bezier curve while keeping ELK's clean orthogonal routing. Per-edge AND per-class. See references/edge-routing.md for the full picture, including why true Bezier curves require switching engines.
Reference files (load on demand)
The body above is enough for ~70% of cases. For the rest, read the relevant reference file:
references/edge-routing.md — connection styling, curves vs. orthogonal, corner softening, multi-bend workarounds, the grid-vs-ELK pitfall in depth.
references/shapes-and-connections.md — every shape value, every connection operator, arrowheads, labels, the full styling menu.
