# Three edit rounds. One change at a time.

A versioned edit trail with an explicit stop rule.

## What you need

An image you may edit and three small, non-conflicting changes. Save the unedited original first. Use the same tool and model throughout so the exercise tests the edit process.

## The steps

1. Write an acceptance list for the original: identity, layout, text and lighting. Choose three changes, such as removing a background prop, shifting wall colour and softening its shadow.

2. Run round one using the prompt. Compare against both the original and acceptance list; save only a passing result as v1. Keep a separate rejected folder.

3. Repeat for rounds two and three using the last accepted image. Name the new change and carry forward the unchanged list. If a round introduces drift, return to its last accepted input.

4. Stop after three accepted rounds or one failed repair. Write which request caused drift and which version you would deliver. Do not hide a failed round by starting a new unrelated image.

## Prompt

```text
Edit the attached last accepted image, version [number].
This round changes only: [one specific edit and its location].
Preserve exactly: [identity cues, layout, text, colours and details not being edited].
Do not restyle, recrop or enhance unrelated areas.
Keep previous accepted changes: [list, or none for round one].
The attached image, not an earlier generated draft, is the source for this round.
```

## Review

- [ ] Each accepted version has one named change and its input saved.
- [ ] Unrequested regions pass a comparison with the previous version and original.
- [ ] The final log includes failures and the chosen delivery version.

## If it fails

Return to the last accepted input and narrow the request to one visible region. If the same failure recurs, use a conventional editor instead of accumulating drift.

## Recall later

Why compare against the original as well as the previous round?

Small errors can accumulate across individually plausible edits. The original remains the identity baseline.

Tool: https://chatgpt.com/
Written by The Leverage Club. Reference edition 2026-09-08.

## Evidence and limits

Original first-look exercise, not a club benchmark. Earlier course images are practice inputs, not Images 2.5 outputs. Check current access and cost.

Launch: https://openai.com/index/introducing-chatgpt-images-2-5/
API guide: https://developers.openai.com/api/docs/guides/image-generation

## Your run log

Date:
App and exact model label:
Input filenames and permission:
Requested change:
Accepted or rejected, and why:
Elapsed time:
Displayed cost (or unknown):
Next action: