Automation

Checkpoints and Do steps

The two steps that use AI — how to write them, and what they cost.

Checkpoints

A checkpoint is where you say what should be true. The judge sees the screenshot, the text extracted from the page, timing, and any file the step before downloaded — then answers with a verdict, a confidence, its reasoning, and the evidence it used.

Write the expectation as an observable fact.

Good: "A draft invoice appears in the list, with the status Awaiting approval." Poor: "The invoice works."

Two rules the judge follows, which are worth knowing when you write one:

  • The page's own claim ("Saved!") is never enough on its own.
  • If the evidence cannot confirm your expectation, it fails and says what was missing.

Reading a downloaded file. A checkpoint can be told to read what a particular step downloaded — "the invoice PDF shows 120.00". It reads that step's file, so it is never ambiguous which one.

Cost: 1 credit, or 2 when it reads a PDF.

Do steps

A Do step is an instruction in plain language: "download the invoice PDF". The AI looks at the screen, picks from the elements actually on it, and acts — up to five tries, and you see each one with the screenshot it was looking at.

Cost: 3 credits for the whole step, however many tries it takes. Its expectation is judged within that price.

What a Do step will never do

Whatever the instruction says, the runner refuses to click anything that deletes, deactivates, closes an account, unsubscribes, pays, purchases, buys, checks out, places an order, transfers or withdraws — including a "Confirm" inside a "Delete project?" dialog. Send and Cancel are allowed. It also stays on the site under test.

If you need one of those actions, write it as an ordinary Click step, where a person chose it.

Pin it when it works

After a good run, Pin these actions turns what the AI did into ordinary steps. Later runs are then deterministic and cost nothing but the checkpoint.