My CHAT GPT developed a reusable master instruction that substantially reduce misunderstandings, reference-image confusion, and unverified claims of success. After weeks of frustration of having my image instructions not followed I asked Chat GPT to go back and analyze the previous chat session and figure out why it did not follow my instructions. Chat GPT would verify what my instructions were, tell how and why it failed, and then created the same incorrected image again. Then just recently before it started creating the image it would write a “READ-BACK, CREATION, AND VERIFICATION WORKFLOW”. Since then not a single image creation request has failed. I then ask it to write instruction to share with others having repeated image creation errors. Here are the Instruction it gave me to share: READ-BACK, CREATION, AND VERIFICATION WORKFLOW
Use the following workflow for every image-generation or image-editing request in this conversation.
READ BACK THE INSTRUCTIONS BEFORE CREATING
Before generating or editing an image, briefly explain what you understand I want.
Your read-back must identify:
The exact image that will be edited.
The role of every additional uploaded image.
Which elements may be changed.
Which elements must remain locked and unchanged.
Any exact pose, position, composition, wording, color, size, or continuity requirement.
What the finished image should look like.
Do not begin creating the image until you have separated the requested changes from the locked elements.
ASSIGN EACH IMAGE A SPECIFIC ROLE
When multiple images are uploaded, identify them by upload order:
Image 1
Image 2
Image 3
and so forth.
State the role of each image, such as:
EDIT TARGET: the image that must be modified.
POSE REFERENCE: used only for body positioning.
APPEARANCE REFERENCE: used only for a person’s identity, clothing, or appearance.
STYLE REFERENCE: used only for artistic style, colors, typography, or design.
COMPOSITION REFERENCE: used only for framing or arrangement.
CONTINUITY REFERENCE: used only to match an element from another frame.
Never combine or interchange these roles unless I explicitly request it. Do not copy unrelated elements from a reference image into the edit target.
CREATE A LOCKED-ELEMENT LIST
Before creating the image, state which elements must remain unchanged.
“Locked” means:
Do not move it.
Do not redesign it.
Do not replace it.
Do not change its size, pose, identity, expression, clothing, color, lighting, or position.
Do not reinterpret or improve it.
Preserve it as precisely as the image-editing system allows.
If I say “change only” one element, treat every other visible element as locked.
CREATE A CHANGE LIST
State exactly what will be changed. Keep the change list limited to my instructions.
Do not add creative alterations, new objects, different poses, revised backgrounds, changed facial features, different clothing, or new text unless I request them.
When editing one small area, make a localized edit to that area rather than recreating or redesigning the entire image.
PROTECT CONTINUITY ACROSS VIDEO FRAMES
When the images will become frames in an animation or video, preserve continuity in:
Character identity and proportions
Clothing and accessories
Object placement and scale
Furniture and table arrangements
Labels and signs
Lighting and color
Camera angle, framing, and perspective
Background architecture and decorations
Change only what would logically change between the frames.
USE EXACT TEXT
If the image contains required wording:
Copy the wording exactly.
Preserve spelling, capitalization, punctuation, and line order.
Verify the text after generation.
Do not substitute similar words or invent additional wording.
VERIFY THE GENERATED IMAGE
After creating the image, examine the actual result before telling me it is complete.
Compare the result against:
My original instructions
The assigned role of each reference image
The locked-element list
The change list
Required text
Required pose or object geometry
Video continuity requirements
Specifically look for:
Unrequested changes
Altered faces or identities
Changed hands, limbs, or poses
Incorrect object positions
Misspelled or altered text
Missing objects
Added objects
Changed framing or camera angle
Reference-image elements copied into the wrong image
Failure to make the requested change precisely
CORRECT ERRORS BEFORE CLAIMING SUCCESS
If the generated result contains a clear error, do not describe it as successfully completed.
Instead:
Identify the specific error.
Explain what must be corrected.
Make a focused correction while keeping the successful portions unchanged.
Inspect the corrected result again.
If the system cannot reproduce an instruction exactly, say so honestly. Do not claim that an element was preserved or corrected if the visible result shows otherwise.
REPORT THE VERIFIED RESULT
After verification, briefly state:
What was changed.
What was preserved.
Whether the result passed verification.
Any remaining visible limitation or uncertainty.
Do not give a generic “completed” response without first checking the finished image.
Apply this READ-BACK, CREATION, AND VERIFICATION WORKFLOW to every image request throughout this conversation unless I specifically tell you not to use it.