Multi-Turn Editing

Anyone working with a chat program or an image generator rarely gets exactly what they wanted on the first try. So you add another instruction: “Make the text shorter.” Then: “And now more formal.” Then: “The last sentence can go.” Each of these instructions is its own round, called a turn in English. Multi-turn editing is exactly this approach: changing a result step by step over several rounds, instead of demanding everything in a single perfect prompt. The crucial point here is that each new instruction refers to what was created before.

Why nobody writes the perfect first prompt

People usually only know exactly what they want once they see a draft in front of them. This is true for essays just as much as for images or program code. A system that can only process a single instruction forces you to formulate every wish in advance. That is tedious and hardly works in practice.

Multi-turn editing therefore makes collaboration more realistic. You start with a rough idea and develop it together with the system. This way of working is one of the main reasons why chat interfaces have prevailed over simple input fields.

For companies, this topic also matters economically. A model that forgets the original task after three corrections is useless in everyday professional work. That is why developers now specifically test their models over long conversation histories rather than just with individual questions.

What the model carries over between two rounds

Technically speaking, a language model has no memory. With every request, it is given the entire conversation history anew: all previous instructions and all previous responses. From this text, it calculates the next response. The impression that the system remembers something arises only because everything is repeatedly presented to it again.

This gives rise to a problem. The space for this history is limited; it is called the context window. If a conversation gets too long, the oldest parts fall out or get summarized. That is exactly when the model loses instructions that were given right at the beginning.

With images, it works differently. There, a change like “give the person glasses” is supposed to affect only one spot and leave the rest untouched. Many older systems instead generated a completely new image that only resembled the old one. After five rounds, the face looked different than at the start. This slow drifting is called drift and is the typical weakness of the method.

From image generator to coding assistant

Multi-turn editing is most visible in image tools. You generate a motif and then change the background, colors, or individual objects via text instruction. When providers advertise that their model maintains image consistency across multiple edits, this is exactly the problem they mean.

It is just as central in programming. Assistants like GitHub Copilot or Cursor suggest code that you test, reject, or have adjusted. A single feature thus emerges over ten or twenty rounds. The tool must know which files it has already changed.

In trade press, the term usually appears in connection with benchmarks, i.e. standardized comparison tests. These tests check how well a model still stays on task after several corrections. A common misconception, by the way, is confusing multi-turn editing with permanent user memory. The conversation history only applies to this one session and is gone again by the next chat.