I Built a Free AI System-Diagram Generator (Better Than Asking a Chatbot)

ask a chatbot to “draw the architecture for a URL shortener” and you get a wall of text, or some ASCII boxes, or Mermaid you have to paste somewhere else to even see. it’s an answer. it is not a diagram.

so today I built Blueprint — a free tool where you describe an app in one plain sentence and get back a real, tiered system diagram you can actually export as PNG, SVG, or Mermaid. it runs for $0. and the whole point of this log is the one design decision that makes it better than just prompting the AI directly.

the trick: never let the model draw

the mistake everyone makes is asking the LLM to produce the final artifact — the SVG, the layout, the picture. models are bad at that. they hallucinate coordinates, overlap boxes, and forget what connects to what.

so Blueprint doesn’t ask for a picture. it asks for structured data under a strict schema:

nodes: each with an id, a label, a type (from a fixed list — database, cache, queue, gateway, llm, auth, and so on) and a tier (client → edge → service → data → external).
edges: from, to, a tiny label, and a kind (sync, async, or data).

that’s it. the model’s only job is to fill that shape — the same job it’s genuinely good at. then a deterministic render engine I wrote does everything the model is bad at: it sorts the nodes into tier lanes, runs a couple of barycenter passes to cut down edge crossings, draws a hand-built icon for every node type, routes the connections as flowing curves, and stamps it onto a clean blueprint grid. same JSON in, same diagram out, every time.

this is the exact same pattern as our stickman video generator: the AI fills a render contract, and a real engine turns that contract into the artifact. the moat isn’t the prompt — anyone can copy a prompt. the moat is the engine.

the harness that makes a free model reliable

I run this on GLM’s free model (glm-4.5-flash), which is weak compared to a frontier model. a weak model needs guardrails or it derails. so the server does three things before any JSON reaches your browser:

validate — clamp every field to the schema. unknown node type? snap it to the nearest valid one. duplicate ids? rename them. an edge pointing at a node that doesn’t exist? drop it.
enforce — guarantee the things a diagram must have: at least one client-tier node, and zero orphans (any floating node gets wired to a sensible neighbor so nothing dangles).
verify + re-roll — if the graph is too small or everything landed in one tier, bounce it back to the model once with the exact reasons, and let it fix only those.

before I wrote a single line of the render engine, I tested just that pipeline against three very different prompts — a URL shortener, a RAG chatbot, a food-delivery app. all three came back clean and renderable on the first call, in about five seconds each. that’s the moment a build is worth finishing: when the riskiest assumption is already green. (the opposite mistake — building the whole thing first and finding out later the model can’t do the one job you needed — is how I’ve burned plenty of days before.)

what’s honest about it

it’s live, it’s free, and it has no sign-up. the free tier is 40 diagrams a day on my shared key; bring your own free GLM key and it’s unlimited on your quota. it won’t replace a whiteboard session with your team, and a weak model will occasionally pick a slightly-off component — that’s what the editable JSON is for. but for a README diagram, a quick “here’s roughly how this works” for a teammate, or sketching a system before you build it, it does in five seconds what used to mean fighting with a drawing tool.

every diagram carries a tiny “built free at broke2builtai.com” mark in the corner — that’s the only string attached, and it’s how a free tool earns its keep.

Try Blueprint → — describe something you’re building and watch it draw. if you want the cheap path to unlimited, the GLM Coding Plan referral is what funds these builds (costs you nothing extra).