`node app “Web app” node app.ui “Interface” node app.api “API” below app.ui
node store “Database” right of app level with app
edge app.api -> store “queries” from: right to: left `
Nothing is nested, so no line depends on another line’s position or indentation.
The draft is in SYNTAX.md, with a worked example in examples/.
The common case is not drawing a diagram, it is changing one. Ask for the auth service to move left and a queue to go behind it. With pixel coordinates, an agent has to rebuild the picture from the numbers before it can work out which numbers to change. With auto-layout there is nothing to read at all, because the arrangement was never written down — it can only reword the source and re-render. With stated placement the arrangement is in the file as sentences, and changing the picture is changing the sentence that says where the thing goes.
Writing has the same shape. An agent emitting Mermaid is guessing at a layout that an algorithm settles later, and its only way to find out is to render and look — a round trip that comes back as a picture rather than as a list of what is wrong.
Intent is confirmable, outcomes are not, and the difference is worth being precise about. An agent can re-read its own file and see that the database is under the API and all four machines hang off the sync hub. It cannot see that two clusters anchored to different things now overlap, that a text overflowed its node, or that an edge crosses four others — those are resolved from the statements rather than stated, so they need the diagnostics in the scope section below.
reladraw is too new to be in any model’s training data, so an agent has to be told the language before it can write it. This repository ships an agent skill that does exactly that — the syntax, when to reach for the language, and what re-reading its own source can and cannot confirm.
npx skills add reladraw/reladraw -g
That installs it for whichever agent you use — Claude Code, Codex, Cursor, Copilot and others — each into its own skills directory. Drop the -g to install it into the current project instead.
To install it for one agent rather than all of them, name it with -a:
npx skills add reladraw/reladraw -g -a claude-code
Re-run whichever command you used after a release that changes the syntax. The skill is a copy taken at install time, not a live link, so nothing refreshes it on its own.
It is plain Markdown with the syntax reference beside it, so it is worth reading whatever you use, and copying the directory by hand works just as well.
Version 0.5.0. Early, but it runs: a parser, resolver and SVG renderer in TypeScript with no runtime dependencies, and a command-line tool that takes a text file and writes a standalone SVG. The comparison at the top of this page is that pipeline run on examples/arch.reladraw. What is still visibly off there is typography, not placement.
npm install -g reladraw reladraw diagram.reladraw -o diagram.svg
Or from a clone, which also gets you the examples:
npm install && npm run build node dist/cli.js examples/arch.reladraw -o out.svg
Not built yet, roughly in the order they are missed:
The language is not stable. Expect the syntax to change.
A gap is a minimum distance, never an exact one. Say two things sit side by side, then say a third goes between them, and the first two are pushed apart by exactly what the third needs; delete the third and they close back up. That is the step an author otherwise does by hand — shove things apart to make room, then drag everything back so the diagram is not full of holes — and no number goes stale when a text grows.
So the resolver solves a system rather than walking a chain. Each axis is a set of minimum distances, and the tightest arrangement satisfying all of them is found by longest paths: one answer, no search, no arrangement ever tried and rejected. The engine works out distances; which side of what a thing sits on came from the file.
That is also what makes non-overlap affordable, so it holds for every pair of nodes without anyone writing it down. On its own “these two must not overlap” is a choice among four directions, which is the search this design refuses — but the file has usually settled it already: if your arrangement lets one node travel away from another and offers no way back, that is the only separation it permits. Where the file orders a pair on neither axis, the tool names them rather than guessing; where it orders them on both, the tie breaks toward the axis of least overlap, which is the smallest movement and the one place the tool decides something nobody wrote.
Nothing is nudged. Each round derives the separations the file already implied, adds them as ordinary minimum distances, and solves the whole thing again from scratch — repairing a solved layout in place is the thing being avoided.
- Parser (done)
- Deterministic resolver: minimum distances in, tightest arrangement out (done)
- Static SVG renderer (done)
- A command-line tool: text file in, SVG out (done)
- A placement grammar that can say what a real diagram needs: several placements on one node, one thing between two others, exact side-to-side alignment (done)
- Nodes that do not overlap by default, with the separation direction derived from the stated arrangement (done)
- Minimal node-avoiding edge routing
- Machine-readable diagnostics from the solved geometry
Diagnostics are a real output rather than a debugging aid. What they cannot do is stand in for the grammar: a check catches only what the language genuinely leaves open, and “these must not overlap” rules arrangements out without naming one, so it can never place anything. Everything the source cannot tell you is computable once the geometry is solved, with no image involved: overlapping nodes, crossed edges, text exceeding its container, anything off-canvas, large dead regions. So the tool reports hub overlaps laptop1 and edge auth->db crosses 4 edges, and the fix is written in the same vocabulary as the source. An agent working this way reads a report about a text file it wrote and edits that text file — no rendering, no vision model, no pixel arithmetic.
A diagnostic never repairs a solved layout in place. That is the line the design holds: a checker allowed to nudge nodes is a layout algorithm with a bad search strategy, fixing one overlap into the next with no view of the whole. Deriving a constraint the file already implied and solving the whole system again is a different thing, and is how non-overlap works. What is left over — anything the source genuinely does not settle — is reported, naming the statement that was broken, and the author edits the source. Open, and it decides how far this goes: may a diagnostic describe a fix in words, or only name the symptom? Describing one means the tool has an opinion about layout, which is the auto-layout instinct coming back in through the side door.
Three design problems decide how much machinery this needs, and the first outranks the other two:
Saying enough. The benchmark contains arrangements the grammar cannot express at all, which is why some nodes land in the wrong place no matter how the file is written. So the work is adding statements, not restricting them. Expressiveness is not the danger; the engine choosing an arrangement is.
What the engine is allowed to decide. Auto-layout is refused, because a picture chosen by an algorithm is not predictable from its source, and that predictability is the entire point. Working out coordinates from an arrangement the author stated is a different thing and is simply the job. The test between them: the engine’s freedom may affect distances and never relationships. If a default can change which side of something a node sits on, the language was short a statement and the tool should say so rather than guess.
Overlap and edge routing. Relative placement with default spacing collides as soon as two clusters grow toward each other. Stating placement and then routing edges afterward with no influence on them reproduces the exact failure this is meant to avoid, so minimal node-avoiding orthogonal routing belongs in the first version. Routing and diagnostics are complements, not substitutes: routing fixes what it can, and the diagnostics report what it could not.