Author editable SVG technical infographics — architecture, flow, comparison, hierarchy — by computing the layout numerically before writing markup, then rendering a 2x PNG via headless Chromium. Carries a CJK-first font stack, a deterministic source lint, and mermaid-vs-SVG selection rules.
Author editable SVG technical infographics — architecture, flow, comparison, hierarchy — by computing the layout numerically before writing markup, then rendering a 2x PNG via headless Chromium. Carries a CJK-first font stack, a deterministic source lint, and mermaid-vs-SVG selection rules.
when_to_use
Use for a static diagram image bound for slides, email, social, or offline use, or a freeform architecture infographic needing pixel control or precise Korean line wrapping. Markdown-embedded diagrams that change often or stay locale-synced remain mermaid.
Produce a hand-editable SVG diagram whose geometry was decided by arithmetic
rather than by eye, plus a 2x PNG raster of it. The output is one static image:
no animation, no scripting, no external asset at view time.
Step 0 — Decide whether this is an SVG job
This skill is additive to the mermaid pipeline, never a replacement for it.
Nothing here migrates, rewrites, or deprecates an existing mermaid diagram, and
no diagram should ever exist in both forms — that is dual maintenance, and it is
the one failure this section exists to prevent.
Route the request before drawing anything:
Signal
Route to
The diagram lives inside a markdown document
mermaid
It changes often, alongside the prose around it
mermaid
It is a standard type: flow, sequence, ER, state, class, gantt
mermaid
Its text labels are kept in sync across locales
mermaid
The deliverable is an image file for slides, email, social, or offline reading
this skill
It is a freeform architecture or concept infographic with no standard shape
this skill
Pixel-level control of position, spacing, or layering is required
this skill
Korean or other CJK labels must wrap at exact, verified widths
this skill
When several signals point both ways, mermaid wins: a mermaid block is cheaper
to keep correct than an image. Choose this skill only when the routing table
gives it an unopposed reason.
One diagram, one home. If a mermaid version already exists, either replace it
outright (and delete the mermaid block in the same change) or leave it alone.
Never ship both.
Runtime prerequisites and degradation
Node 18 or later and a headless Chromium-family browser are needed only to
lint and to render. Neither is needed to install this skill, to discover it, or
to author the editable SVG — authoring is always available.
Node 18+
Headless Chromium
What is delivered
present
present
Editable SVG, machine lint report, 2x PNG with the browser executable and version disclosed and PNG header dimensions verified
present
absent
Editable SVG plus machine lint report. State plainly that no headless browser was found and no PNG was produced
absent
either
Editable SVG plus the manual checklist result from references/authoring.md. Do not attach a machine-lint label, and do not claim a render
Never fabricate a PNG, a pixel dimension, or a lint verdict for a tool that did
not run. Say which step was skipped and why.
The workflow
Six steps, in order. Steps 1 through 3 finish before a single SVG element is
written; that ordering is the whole method.
Frame. Write down the message the diagram must land, the target medium
(slide, README hero, email, print), the canvas size, and the label language.
Pick an archetype. Architecture stack, left-to-right flow, side-by-side
comparison, or hierarchy tree. Skeletons are in references/archetypes.md.
Run the numeric layout pass. Produce the box table and pass every
containment and text-budget check below. Do not proceed on a failing row.
Author the SVG from the table. Every coordinate is either a table value
or a formula over table values.
Lint the source with scripts/check-svg.mjs. Clear every error; triage
every warning.
Render and verify with scripts/render.mjs. Confirm the reported PNG
dimensions match the requested 2x target, then look at the PNG.
The numeric layout pass
Build one table before authoring. Five owned columns per box — id, x, y,
w, h — and nothing else is typed by hand. Every other number in the file is
derived from those.
Grid. For canvas width W, n columns, outer margin M, gutter G:
colW = (W - 2*M - (n-1)*G) / n
colX(i) = M + i * (colW + G)
If colW falls below the archetype's minimum card width, reduce n or widen
W. Do not shrink the margin to rescue a column count.
Containment. Check every row, and stop if any fails:
M <= x and x + w <= W - M
M <= y and y + h <= H - M
parent.x + pad <= child.x and child.x + child.w <= parent.x + parent.w - pad
Derived geometry. Centers and anchors come from the box, never from a
per-language nudge:
cx = x + w/2
cy = y + h/2
iconCenter = (x + pad + iconR, y + h/2)
titleBaseline = y + pad + titleSize
lineBaseline(k) = titleBaseline + titleGap + k*lineHeight
If you find yourself moving an icon down three units "because the Korean text
sits low", the formula is wrong. Fix the formula, not the instance. Hand-tuned
per-language offsets are exactly the render-fix loop this method removes.
Connectors. Endpoints are derived too; the arrowhead length is subtracted
from the terminal end so the marker tip lands on the border, not inside it:
Full formula set, including radial and multi-lane fan-out, is in
references/authoring.md.
Text budget — CJK first
Set a CJK-first font stack on the root so Hangul, Kana, and Han glyphs resolve
before any Latin fallback is consulted. A Latin-first stack makes CJK glyphs fall
through to an arbitrary system font and silently changes every measured width:
font-family="Pretendard, 'Noto Sans KR', 'Noto Sans JP', 'Noto Sans SC',
'Apple SD Gothic Neo', 'Hiragino Sans', 'Microsoft YaHei',
system-ui, sans-serif"
Capacity per line, for usable width u = w - 2*pad at font size s:
Latin: capacity = u / (0.60 * s) average Latin advance is about 0.60em
CJK: capacity = u / (1.00 * s) full-width advance is 1.00em
The ratio between them is the working rule: a Korean, Japanese, or Chinese
line holds roughly 60% of the character count a Latin line holds in the same
box at the same size. Budget the copy against that number and edit the wording
to fit before authoring. A line that mixes scripts is budgeted at the CJK rate
for its whole length.
Two things are forbidden here because both hide the problem instead of solving
it: truncating a label after the fact, and shrinking the font size for one
language only. Rewrite the label.
Every diagnostic carries file:line:column, a stable code, and a message. The
two tiers are not interchangeable:
Errors — deterministic, always fix. Unbalanced tags; missing or malformed
viewBox; a width/height pair whose aspect ratio contradicts the viewBox;
duplicate id; a local reference (url(#id), href="#id") with no matching
id; a <marker> missing required geometry; a <marker> that leans on the
implicit markerUnits default, which rescales arrowheads with stroke width and
is the usual cause of arrowheads that look right in one diagram and wrong in the
next.
Warnings — heuristic, triage individually. Estimated text overflow of its
container rect; a pill too narrow for its label once the round-cap inset is
applied; an element extending past the viewBox. These use character-advance
estimation, so they are advisory: confirm in the rendered PNG rather than
reflowing the layout on the warning alone. A warning that survives visual
inspection is a real defect; one that does not is noise.
Exit status is 0 when no error was found, 1 on any error (or on any warning
under --strict), 2 on a usage or read failure.
Without Node, walk the manual checklist in references/authoring.md instead and
report it as a manual check — never as a lint result.
The renderer resolves a Chromium-family executable from CHROME_PATH, then from
the well-known install locations for the platform, then from PATH. It reports
the exact executable it used and that browser's version string — always
include both in the deliverable, because a diagram rendered by a different
browser build is a different artifact.
It computes the target as round(viewBox_w * scale) x round(viewBox_h * scale),
screenshots at that window size, then reads the PNG's own IHDR header and
compares the stored dimensions against the target. A mismatch is a failure, not
a rounding note.
Exit status: 0 verified, 1 render or verification failed, 2 no headless
browser found, 3 usage error. Exit 2 is the degradation signal — deliver
the SVG alone and state the limitation.
Bundled references
File
Contents
references/archetypes.md
The four archetype skeletons with their canvas presets, grid parameters, and per-archetype containment rules
references/authoring.md
Full geometry and connector formula set, the icon set, palette and type scale, and the manual no-Node checklist
references/sketch.md
Opt-in hand-drawn preset layered over the same computed layout
Script
Purpose
scripts/check-svg.mjs
Deterministic source lint, errors and warnings, file:line:column diagnostics
scripts/render.mjs
Headless-Chromium 2x PNG render with browser disclosure and PNG header verification
Both scripts run on the Node 18 standard library alone. There is no package to
install and no browser bundled.
Relationship to the report renderer
moai-domain-html-report renders a markdown report into one self-contained HTML
file and may embed mermaid inside it. That skill owns reports; this one owns
standalone diagram images. They compose — a report may link or embed a PNG this
skill produced — and neither replaces the other.
Common Rationalizations
Rationalization
Reality
"I will sketch the SVG first and fix the coordinates once I see it"
That is the render-fix loop. Each visual fix invalidates a neighbour and the diagram never converges. Compute the table first.
"The Korean label is only slightly too long, it will fit"
It will not: CJK glyphs are full-width, so the line holds about 60% of the Latin count. Rewrite the label before authoring.
"I nudged the icon down 3 units and it looks right now"
A per-instance nudge means the center formula is wrong. Derive from box geometry and the nudge disappears everywhere.
"No browser here, but the PNG would have been 2400x1600"
An unrendered size is a guess. Deliver the SVG and state that no PNG was produced.
"The lint only reported warnings, so the file is clean"
Warnings are heuristic, not absent. Triage each against the rendered PNG before dismissing it.
"This flowchart would look nicer as an SVG"
A markdown-embedded, frequently-changing standard diagram stays mermaid. Nicer is not a routing reason.
"I will keep the mermaid block and add the SVG for slides"
Two sources for one diagram drift apart. Pick one home.
Red Flags
SVG elements were written before the box table existed.
A coordinate in the file cannot be traced to a table value or a formula.
The same diagram exists as both a mermaid block and an SVG.
A font stack lists a Latin family before any CJK family.
A label was truncated, or a font size was reduced for one language only.
A PNG dimension, browser version, or lint verdict is reported for a command
that was never run.
Lint errors were downgraded to warnings to get to a render.
A <marker> has no explicit markerUnits.
Verification
Routing table consulted and the SVG choice has an unopposed reason.
No mermaid version of this diagram remains alongside the SVG.
Box table complete before authoring; every coordinate traces to it.
All containment checks pass for boxes and for children inside boxes.
Centers, baselines, and connector endpoints are derived, not hand-tuned.
Font stack is CJK-first; every line fits its computed capacity.
check-svg.mjs reports zero errors; each warning triaged and recorded.
render.mjs verified the PNG header against the 2x target.
Browser executable and version disclosed with the PNG.
Any skipped step named explicitly, with no substitute claim.
