You will be able to choose the right way to recover when an answer or a whole conversation goes wrong.
Marcus had been working on a proposal for a new client for most of the afternoon, all in one chat. By message thirty, the assistant kept reintroducing a three-year contract option he had rejected an hour earlier. He told it again. Two replies later, the option was back. He tried capital letters. It apologised and did it again.
Feedback is the right tool most of the time, as lesson 4.2, Feedback that gets a better second draft, showed. But some problems are not about the draft. They are about the conversation the draft is sitting in, and for those, there are three other moves: regenerate, edit your earlier message, or start a fresh chat. Each fits a different situation.
Most assistants have a button to get a new attempt at the last answer, often labelled regenerate, retry or try again. Use it when your brief was sound but the answer went in a poor direction. Maybe it took an angle you did not want, or picked an example that does not suit your reader, or simply came out flat.
Regenerating works because of how answers are produced. AI fundamentals lesson 3.3, Why the same question gets a different answer, explained that assistants sample their words rather than always choosing the single most likely one. The same request run again gives a genuinely different attempt, and sometimes the second or third is much better.
What regenerating cannot do is fix a missing instruction. If the brief left out the reader or the length, every attempt is guessing at the same gap, and you are rolling dice. Two or three regenerations that all miss in the same way are a sign the problem is in your request.
When you realise the brief itself was wrong or incomplete, the cleanest fix is often to edit your earlier message rather than reply with a correction. Many assistants let you edit a message you already sent, and when you do, the conversation branches from that point. The old answer, and everything after it, drops out of the version the model is now working from.
This matters because a correction added at the bottom of a chat leaves the bad answer in place. The model can still see it, and as AI fundamentals lesson 3.2, It writes by predicting the next token, again and again, explained, everything in view shapes what comes next. A wrong draft sitting in the conversation keeps pulling later answers towards it. Editing the original request removes it from view entirely.
So if Marcus had noticed early that his brief never said "one-year contract only", the best move would have been to edit the first message to add it, not to argue with the answer in message three.
Some conversations are past saving. The chat is long, the same mistake keeps coming back, or you and the assistant have gone round in circles. AI fundamentals lesson 5.2, Why long chats drift and big files get skimmed, explained part of the reason: early instructions fade or drop out as a conversation grows. The other part is that earlier mistakes are still in the conversation. Every rejected draft, every wrong option and every correction is text the model can see, and it keeps pulling answers back towards things you have already rejected. Marcus's three-year option appeared in five earlier messages and his corrections in only three, so more of the text in view pointed towards the option than away from it.
When that happens, stop correcting and open a new chat. The signs to watch for are a mistake coming back after you fixed it, constraints from the start being ignored, and answers that feel like they are arguing with you.
The new chat needs to know what you agreed in the old one. It does not need the whole old conversation. Pasting thirty messages into a fresh chat brings every rejected idea with it, which defeats the purpose.
Write a short summary of the decisions instead. Five lines is usually enough: the goal, the reader, what has been decided, what has been ruled out, and the current state of the work. You can ask the old chat to draft this summary, but check it yourself, because the old chat is the one that kept forgetting.
Marcus's summary read: "Proposal for managed IT support for a 40-person logistics firm. Reader is their operations director, who is not technical. Decided: one-year contract only, monthly billing, on-site visits twice a month. Ruled out: three-year option, any hardware sales. Current draft of sections 1 to 3 below." He pasted the three sections he was happy with under it. The new chat never mentioned the three-year option once.
Once you have the summary, use it. Paste it at the top of the new chat, follow it with your next request, and carry on from where the useful work stopped. You likely have a chat right now that has grown long and started repeating itself, and it is the ideal place to try this.
Write a five-line summary of a long chat you have, start a fresh chat with it, and note whether the answers improve.
Junxiong-WFG Organisation is an authorised representative of AIA Financial Advisers Private Limited (Reg. No. 201715016G).