Use GPT-6 Astra to Test Your Product, Not to Build It.

The single best piece of operating guidance from launch week did not come from OpenAI. It came from Claire Vo, who used Astra on real production work and landed on a rule.

The single best piece of operating guidance from launch week came from Claire Vo, who used Astra on real production work and landed on a rule.

The logic holds up on inspection. Building through a UI is the slowest possible way to build, because the API underneath is faster, cheaper and more reliable. But QA through a UI is the only honest way to QA, because the UI is what your customer touches, and a model clicking through it at 3am finds the things your test suite was never pointed at.

So the highest-value first deployment is not "let it build the feature." It is "let it use the product the way a customer would, every night, and tell you what broke." That work has ground truth, a clear owner, and a baseline you already measure.

Vo's other observation is the one with strategic teeth, and it cuts against the direction most AI architecture writing has been running: UI is back. If agents operate software through its interface, then the interface is an agent surface, not a legacy human affordance to be deleted. That is worth sitting with if you had concluded that the API is the only thing that matters. Both can be true: the record layer matters more, and the screens just became machine-readable.

Start here rather than with a build task, because a nightly QA sweep is a falsifiable pilot. You know by morning whether it found something your tests missed, and you did not have to argue about model quality to find out.

What she actually got out of it

The rule is worth more because of what sits behind it. Vo did not arrive at "QA, not building" from theory; she arrived after a run of one-shot wins that previous models could not produce:

A ChatPRD product-intelligence feature she had failed at for six months, across GPT-5.5 and everything after, landing at 90%. An AIM-style Mac desktop app in a single attempt. A Divoom hardware CLI she had been chasing since GPT-5.5. Blender 3D assets, generated as code.

So this is not a sceptic's rule. It is what someone concluded after the model cleared a backlog she had been carrying for half a year. The building got dramatically better through the API. The interface is still the wrong door to walk through when a door is not the point.

The argument with strategic teeth: UI is back

Vo's second observation is the one that cuts against the direction most AI architecture writing has been running. If agents operate software through its interface, the interface is an agent surface, not a legacy human affordance waiting to be deleted.

That is worth sitting with if you had concluded that the API is the only thing that matters, or that the right move is to strip the UI and expose the data. Both can be true at once: the record layer matters more than ever, and the screens just became machine-readable. A product with an unusable interface is now failing two audiences instead of one.

It also reframes accessibility work and interface consistency as infrastructure rather than polish. A UI that a screen reader can navigate is, not coincidentally, a UI an agent can navigate. Semantic markup, stable selectors, honest labels and predictable state are what let both get through.

How to run the first sweep

Point it at the journey, not the page. "Sign up, add a payment method, invite a colleague, cancel" finds more than any single-screen check, because the failures live in the transitions.

Run it on Batch or Flex. Nothing is waiting on the result overnight, and those tiers are half price. This is the workload that should never run at standard rates.

Define what counts as a finding before you start, or you will get a report describing your own product back to you. A finding is something a customer would have hit and your test suite did not cover.

Give it a throwaway account. It is driving a real session with real permissions, which is the subject of Part 6 and the reason not to hand it a staff login.

Part 4 of Working with GPT-6 Astra: The Operator Guide.