Start an AI image edit with one deliberate change, then revise from the first failure.

Answer first: keep the reference subject unchanged, choose one primary transformation, and judge the result against three visible checks before adding more instructions.

1. Lock the reference

Write what must remain: identity, age cues, relationship, silhouette, label, or proportions. These are constraints, not style suggestions. Keep them short enough to check in the output.

2. Choose one primary transformation

Pick one: change the setting, change the clothing, change the lighting, change the crop, or simplify the background. Do not combine all five in the first attempt.

Example: move a father-and-child portrait from indoors to a quiet garden while keeping faces, ages, clothing, and relationship unchanged.

3. Remove instructions that do not serve the job

Delete extra adjectives, unrelated props, and multiple camera directions. A shorter prompt makes the cause of a bad result easier to identify. Add a supporting detail only when it explains the chosen transformation.

4. Review in a fixed order

1. Subject: is the person or product still recognizable?
2. Protected detail: did any locked feature change?
3. Transformation: is the one requested change visible and believable?

5. Revise only the failed step

If the subject changed, strengthen the lock. If the composition is weak, revise the scene or crop. If the requested transformation is missing, clarify that one instruction. Keep the other cards unchanged for the next attempt.

Reusable first-pass card

Must stay: ____________________

One primary change: ____________________

One supporting detail: ____________________

Review order: subject / protected detail / transformation

Next revision: ____________________

Use the identity-and-change audit or read the six-card prompt brief.