The third time she typed make the hero section green, the entire homepage rearranged itself. The big top banner did turn green, fair enough. But the menu reshuffled. A whole section of glowing customer quotes simply vanished. And a stock photo of a coffee shop, which she had never asked for and did not want, materialized in the footer like an uninvited guest. She had asked for one color. She got a different website.
This was someone using our website builder, the one where you just chat and the site changes. Round one, it looked sharp. Round two, a little off. Round three, unrecognizable. By round five she was starting over from scratch, and you could practically feel the sigh through the screen.
The obvious way to wire up a thing like this is the way every tutorial does it. Take what the person typed. Hand it to the AI. Get a page back. Show it. We had built it exactly that way, and it demos beautifully on the very first prompt, because the first prompt is a from-nothing job. The AI invents a page out of thin air, and a confident assistant invents confidently. Every prompt after that is a completely different kind of job, and we had not noticed the difference.
The assistant never saw what it had built the turn before. It saw only the new instruction, treated it as another invent-from-scratch request, and produced a fresh page that matched the words green hero. Everything else already on the page was, as far as it could tell, an open invitation to redecorate. The drift was not a flaw in how we asked. It was the wrong shape of question entirely.
If you have ever pasted a proposal into a chat box, asked it to rewrite the proposal, and watched it confidently rewrite a proposal it had never actually seen, you know this feeling. Or tighten this email, with the email in one message and the change in the next, and the second turn lands like the first one never happened. Every turn that does not carry the previous state forward is a fresh roll of the dice.
So we changed how the conversation works. Now every turn quietly hands the assistant the page as it currently stands, along with the new request, and asks it to return that same page with only the requested change made. Here is the current thing, here is what they want, give it back modified. The result is small and obvious. The parts nobody touched stay exactly where they were. The assistant is no longer conjuring a new page each time. It is editing the one that already exists. Same assistant, same request, completely different shape of question. Round three now looks like round one with a green banner, because that is genuinely all that changed.
For anyone sizing up an AI tool, this is the question that matters more than any benchmark or feature list. It is not can the AI do the thing. It is does the AI know what state the work is already in. A draft-the-contract assistant is one product. A draft-the-contract assistant that, on the next turn, understands it is editing the contract you just drafted is an entirely different product. The first looks great in a demo, because a demo is one turn. The second looks great in real life, because real life is twenty turns.
This is also why so many AI features tucked inside bigger software feel like a letdown. They were bolted on as one-shot suggestion buttons. Click, get an idea. Click again, get a different idea that has forgotten the last one completely. The actual state of your work, the half-finished invoice, the job post you were partway through, never makes it into the box. The assistant is doing a fresh interview every single turn, with no memory of the one before.
When you are choosing tools, this is a cheap little test. Open the AI feature. Ask it for something. Then ask it to change one small thing. If the second answer ignores or steamrolls the first, the tool was built for demos. If the second answer keeps everything you liked and changes only what you named, the tool was built for work. Most are still built for demos. The ones built for work win quietly, because the customer stops fighting the software and just starts using it.
The lesson is older than any of this technology. Inventing impresses. Editing compounds. If a tool forgets the state of the work between turns, no amount of raw smarts will rescue it. If it remembers, even a modest assistant feels like a careful colleague who was paying attention. Build your tools around the state of the work, not around the turns.