Describe the process in words.
Get a first diagram back.

Oval for start and end: Every chart begins with one terminator and ends with at least one. If you cannot name where the process starts, the chart is not ready to be drawn. If a path has no ending, you have found a bug in the process, not in the diagram.

Rectangle for a process step: One action, one box, phrased as a verb: check stock, send the confirmation, log the ticket. If a rectangle needs the word and to describe it, it is two steps that got compressed to save space.

Diamond for a decision: A question with a fixed set of answers. Label every outgoing arrow with its outcome, not just the yes one. An unlabelled branch is the single most common reason two readers walk away with different understandings of the same chart.

Parallelogram for input and output: Data going in or coming out: the submitted form, the generated invoice, the exported report. Distinguishing it from a process step tells a reader instantly where information enters and leaves, which is usually where things break.

Resolve every branch: Each decision diamond needs all of its outcomes going somewhere real. The no path is where processes actually fail, and it is the path people forget to draw, because the happy route is the one everyone has in their head.

Keep one level of detail per chart: This is the most common flowchart mistake there is. One box says handle onboarding and the next says click the blue Save button. Pick an altitude and hold it. If a step needs its own detail, give it its own chart and link to it.

Watch for the step nobody owns: When you review the chart out loud, listen for the pause after who does that. A step with no owner is not a diagramming problem, and drawing it is how it becomes visible enough to fix.

Cut what does not change the outcome: If removing a box changes nothing about how a reader would act, remove it. A chart people can hold in their head gets followed. A chart with sixty boxes gets screenshotted once and never opened again.

One lane per role or team: Lay the chart out in horizontal bands, one per role, and place each step in the lane of whoever performs it. On a canvas a lane is just a labelled region you drag cards into, which is enough to make ownership unambiguous.

Every lane crossing is a handoff: The arrows that jump between lanes are where delay, dropped work, and finger-pointing live. Counting them is one of the fastest process diagnostics available, and it takes about ten seconds once the chart is drawn.

It changes the review conversation: People argue about a generic process forever and get very specific very fast when they see their own lane. Putting names on bands is what turns a review from polite nodding into the corrections you needed.

A first version from a written description: Paste the process as prose and get steps, decisions, and branches laid out across the canvas. It reads the sentences for conditions and sequence rather than making you place every shape yourself.

Ask what is missing: Because the AI reads the whole canvas as context, you can ask which decisions have unhandled branches or where the chart changes altitude, and get an answer about your diagram instead of a generic checklist.

Re-prompt without losing your corrections: Cards you have rewritten and placed stay exactly where they are while the AI builds around them, so the corrections gathered in a review survive the next round of generation.