Learn

Architecture & pattern guides

Read the idea, build it on a Grid, prove it in a challenge. Short, practical guides — illustrated with RodGrid's own nodes — on system design, design patterns, flow modeling, and how to get the most out of the app.

All articles

Writing a PRD from the diagram

The Grid says what is being built. Tell AI PRD what it is for, and it drafts the requirements, instructed to list the gaps as questions instead of guessing them.

Four boxes describe what a URL shortener is made of: a browser, an API, a cache and a database. They say nothing about who creates the links, how long a link should work, or how anyone will know the service is doing its job. A PRD (product requirements document) is mostly those answers, so before AI PRD writes one, it asks you for them.

A small shortener

All the PRD can learn from the drawing: four components and what flows between them.

Four questions, one required

On a Diagram Grid, the Analysis Model button in the top bar brings up the Analysis panel, and AI PRD sits further down it, after Architecture Analysis. You need to be signed in with edit access to the Grid; a Challenge Grid does not offer it. The form asks:

  • What problem does this solve? Required: the situation today and why it has to change. For the shortener, "Campaign links are 300 characters long and break when an SMS splits them." While it is empty the button stays disabled, and the panel says why: "Describe the problem first — without it the PRD would be guessed from the diagram."
  • Who is it for?, What does success look like? and What is out of scope? Optional. Anything you leave blank should come back as an open question instead of a guess.

For a first run, fill in only the problem and press Generate PRD - 15 credits. On a very large Grid the button shows a little more; the rest of how credits are spent is in AI credits: what they buy and how they are spent.

Reading the draft

A note that an AI wrote it comes first, then the sections in a fixed order: Summary, Problem, Target users, Goals, Non-goals, User stories (each "As …, I want …", with a "So that" line and acceptance criteria), Success metrics, Open questions and Confidence, the model's own Low, Medium or High rating: it is told to say High only when you gave the problem, the users and what success looks like, and Low when it had little beyond the problem.

With only a problem to go on, the gaps should land in Open questions, because the model is told to ask rather than infer the users, goals or metrics from the architecture: here, perhaps who may create a link, or what would count as success. That section always appears; when it is empty it says "Nothing had to be guessed — every section traces to what you wrote or to the diagram." That is the model's report, not a check: read the draft for anything you did not say.

Answer, then run it again

The fields keep what you typed after the draft appears. Write the answers to two open questions into the fields they belong to, here Who is it for? and What does success look like?, and press the button again. A decision none of the four names, such as how long a link should work, has no field of its own: write it into the problem, which the model also takes as fact. The new run replaces the draft on screen and is charged again.

If a run succeeds after you leave, the latest PRD is waiting when you reopen the Grid. The four answers are not: they are cleared on reopening, and when you switch Grids, so keep a copy if you mean to run it again. To take the draft out, Copy PRD puts it on the clipboard as Markdown and Download markdown saves a .prd.md file; both start with the note that an AI drafted it.

What it reads, and what it won't do

  • It sees part of the Grid: the title, each node's title and description, and the edges between nodes with their labels. Sticky notes and shapes are not sent, so a requirement written on a note never reaches it. Put it in a field.
  • It takes your answers as fact. A success metric nobody has agreed to becomes a goal the document asserts. Write what is decided and leave the rest blank, so it comes back as a question.
  • It does not design. It is told to stay out of components, technology, APIs and deployment. A review of the design is AI Architecture Analysis: a review of your diagram, in the same panel.
  • Deleting the open questions hides the gaps. Each is a decision the diagram could not make for you. Answer it in the fields and run it again.

No Grid yet? Starting from a brief: the diagram generators draws a first version from your answers.

Practice on URL Shortener

The PRD does not run on a Challenge Grid, so draw your design for the URL Shortener Challenge on a new Diagram Grid and generate a PRD with only the problem filled in. Answer two of its open questions and run it once more. Then open the Challenge on its guided route: its first stage asks for the requirements and the scale, and your two answers are a start on it.

Practice this