A comment under a 100-round "make it better" experiment describes where those builds go wrong. On the "Make It Better Reel" from the Nate Herk channel, one commenter (6 likes) ran the same prompt on a to-do list and ended up with charts, graphs, and games built around it: "When you ask AI to build a brand new feature every time, rather than focusing on building upon the foundation of what already exists to improve it, AI gets sloppy and ignores good architecture."
The commenter named the mechanism: each round asked for something brand new, so the app underneath got worse. Models rebuild by default when a request doesn't give them the current version to work from. The habits that fix it apply to writing, spreadsheets, and code alike.
Why the model starts over
AI can only edit the material it can access in the current interaction: text in the conversation, an attached file, or a project a connected tool can open. If the current version isn't available, the model may generate a new one, ask for the missing material, or give general advice instead of a precise edit. "Make it better" with no version attached tends to produce something new.
"Better" without a direction invites additions. A response that changes two lines can look less substantial than one that adds a chart or a game, so an underspecified request may produce features nobody asked for.
Models also tend to return complete artifacts: the whole file, the whole page, the whole email. One likely reason: a complete artifact still works when some context is missing. The side effect is what the commenter saw: a request for one small change can come back as a rebuild of everything, with the change inside it somewhere.
The three moves that keep the foundation
Each move gives the model something the vague request left out.
Show the current version. Paste the document, attach the file, or open the project in a tool that can read it. This is context, the same force behind most disappointing AI results: the model improves the version it can see, and invents one when it can't see yours.
Name what to keep. "Keep the existing layout and features" does work that "make it better" never does. Constraints the model can't infer, it may guess wrong.
Change one thing at a time. "Make overdue items red" is checkable when the response comes back; "make it better" is not. One change per round keeps the response small enough to review, and makes each round a question with a checkable answer: did the change land, and did anything else change?
Two supporting habits: iterate in the same conversation so the earlier rounds' decisions stay available, and when the chat runs long, start a new one and attach the latest working version again. Save a copy of the last version that worked, too, so you can restore it if a round goes wrong.
The to-do list, rerun
The same to-do list shows the difference between an open-ended request and a constrained edit. The open-ended version: "add a game to my to-do list." The model has never seen the app, so the game arrives attached to a structure it improvised.
The constrained version starts with the app itself:
Here's my to-do list app (paste the code, or describe it). Keep the existing structure and features. Add one thing: overdue items turn red. Show only what changed.
The next round continues on the same base:
Now add a filter for completed items. Same rules: keep everything else, show only what changed.
Each round ends with the same check: does the app still do what it did before the change? A commenter on the main experiment (5 likes) wanted a critique pass after every round: "If AI made it, it needs your input, or another AI's input, otherwise it just assumes it's like 80% good." The one-change-per-round habit builds the critique into the request: name the change, read what came back, and fix a wrong change before the next round builds on it.
When starting over is the right call
A rebuild is sometimes the cheaper option. The pattern is colliding fixes: each round of small changes breaks something the last one repaired. At that point, ask the model to evaluate the base: "Is this structure worth keeping, or would a restart save time?" If you ask for another small fix, the model may keep modifying the existing structure even when the structure needs a redesign. A deliberate restart replaces the structure once; a request for one more fix may keep a weak structure in place for many rounds.
FAQ
Why does ChatGPT rewrite everything instead of just editing?
Two reasons: the current version often isn't in the conversation, so the model produces a new one, and models tend to return complete artifacts rather than minimal edits. Attach the file and add "show only what changed" to push responses toward edits.
How do I get AI to edit my existing work instead of starting over?
Show the current version, name what to keep, and describe the one change you want next. In many coding tools, the response appears as a diff, the proposed changes shown against the current file, which turns the review into a scan instead of a re-read.
Why does AI get sloppy after a long conversation?
The conversation fills up. OpenAI's FAQ describes a full chat as one where your messages and the model's replies both take up the space; Anthropic's docs recommend starting a new conversation in a long chat, since longer conversations use more of your usage allowance. When a chat gets unwieldy: start a new one, attach the latest working version, and restate the constraints.
When should I let AI start from scratch?
When the foundation is the problem: fixes keep colliding, the requirements changed past what the structure was built for, or the base was a throwaway draft that stopped being one. Ask the model whether to keep making changes or restart, then decide with its answer in view.
Want to go further?
What Is Context in AI? Why It Makes or Breaks Your Results covers what the model can see and how it changes the output, and What Is Prompt Chaining? Building Multi-Step AI Workflows covers the checkpoint habit for multi-step builds. For the next change you were about to request: attach the current version first, and name what to keep.