How to organize conclusions from a long AI brainstorm.

The obvious thing to try is scrolling back, or asking the assistant to summarize where things stand. Both genuinely work, for a while — a short conversation is easy to hold in your head, and a quick summary is enough to reorient. Neither one scales with the conversation. Sixty or eighty messages into comparing options, a summary is a best-effort reconstruction of a conversation that was never organized as a record to begin with — it’s guessing at what actually mattered from everything that was said, including the parts you’d already ruled out.

Say you’re deep into a kitchen renovation, comparing layouts and getting quotes back from contractors over one long evening in ChatGPT. Early on, you float a rough budget — $22,000, just to have a number to plan around. An hour and plenty of exploring later, quotes come back higher than expected and you settle on $28,000 instead. Nothing’s been saved anywhere outside that conversation, so if you ask the assistant to recap where things stand, there’s a real chance it cites the original $22,000 right alongside the $28,000 — not because it’s wrong, exactly, but because both numbers are sitting in the same transcript with nothing marking one as current and the other as superseded. The conversation remembers everything you said. It doesn’t remember what you meant by the end of it.

The fix isn’t a better summary, and it isn’t one big save at the end either — it’s saving each conclusion right when it actually lands, as a handful of small, unremarkable moments instead of one dramatic wrap-up. Each one is a short request: “save that to my kitchen reno project.” Over the course of one conversation, a few of those quietly add up to a short, current record that sits outside the conversation — not the sixty messages of back-and-forth that produced it.

Back to that same evening. Early on, once you’ve floated a starting number: “save that we’re roughly budgeting $22,000 to start.” A while later, after going back and forth between two layouts, you land on one: “save the decision — open layout, island moved to the window wall.” Later still, once the contractor quotes actually come in: “update the budget, it’s $28,000 now, quotes ran higher.” That one supersedes the first number rather than sitting next to it — the $22,000 doesn’t disappear, it’s just no longer what anyone reads as current. Near the end, one thing is genuinely still unresolved, so it gets logged as exactly that: “open question — need to confirm the wall behind the stove isn’t load-bearing before the layout’s final.” Four small saves, none of them a big moment on their own. What they add up to is a project holding exactly three current things: the layout decision, the current budget, and the one open question — not four, because an update replaces rather than piles on.

Come back to it in a new conversation and you’re not scrolling back through that first evening trying to remember which budget number was the real one. You ask your assistant to check the kitchen reno project, and it reads three things: the layout decision, the current budget, and the one open question still waiting on the contractor — the settled residue of four small saves, not eighty messages of exploring.

Open the project on its own and there’s no transcript to replay — just the layout decision, the current budget, and the one open question, in plain sentences. That’s the Simple view. Structured view shows the same project broken into its labeled parts, if you’d rather scan it that way. Either way, it’s the three things you landed on, not the eighty messages it took to get there.