The Problem
Imagine this completely hypothetical scenario that definitely hasn't happened to all of us repeatedly.
You prompt AI to create an image with four specific requirements. You describe the exact scene and visual style, provide the precise dimensions, and specify that the final image must be a PNG with a transparent background. You've supplied all the necessary context. Surely, this one is going to work.
Then, after waiting through the entire generation process, you receive an image that satisfies roughly two or three of your four requirements. Most frustratingly, instead of giving you the transparent background you requested, it decides to honor the idea of transparency by filling the background with the familiar gray-and-white checkerboard pattern.
Thanks, ChatGPT...
Naturally, you begin another round of prompting because getting this image right is now a matter of principle. A few dozen revisions later, the checkerboard is finally gone. Unfortunately, the original art style has gone with it, several details have mysteriously changed, and the background is now white.
Still not transparent, of course. Let’s not get reckless.
You probably don't need to imagine this scenario at all because you've lived some version of it. It's a common frustration with AI: the initial result misses a requirement, and every attempt to fix it quietly introduces another problem. Eventually, you're left annoyed, defeated, and perhaps just a little relieved.
Maybe AI won’t replace me after all.
The Solution
What if, instead of asking AI to generate a result and simply trust that it followed the instructions, you asked it to review its own work? After all, you're not getting paid to do the work yourself!
That's what this prompt attempts to do.
It instructs AI to evaluate its result from several distinct reviewer perspectives. Each reviewer examines the output for different problems, checks it against the original request, and identifies the most important improvements still needed. The AI then revises the result and repeats the review process several times.
The goal is to keep every original requirement visible throughout the process, so fixing one problem doesn't quietly undo three things that were already correct. By treating your instructions as a checklist and reviewing the output through multiple perspectives, the prompt gives AI several opportunities to catch its mistakes before proudly handing you another picture of transparency.
The Prompt
The following prompt, which I modified from one created by The Rundown University1, uses three reviewers to evaluate and improve an initial result through exactly three loops. It then returns the final result, a summary of the changes, and each earlier version for comparison.
- Replace the text inside each
{{...}}field. - Remove any optional sections you don't need.
- Leave everything else unchanged.
Complete the task below, then improve the result through exactly 3 review-and-revision loops.
TASK
{{Write one sentence stating what you want the AI to do. Start with an action verb such as Write, Explain, Plan, Compare, Analyze, or Create.}}
AUDIENCE
{{Describe who will use the result and what they already know. If there is no audience, remove this section}}
INPUT OR SOURCE MATERIAL
{{Paste the text, data, notes, or other material the AI must use. If none is needed, remove this section}}
REQUIREMENTS
{{List each non-negotiable rule on its own bullet. Include anything the result must avoid. If there are no additional requirements, remove this section}}
DEFINITION OF DONE
{{Describe exactly what the AI should return. Include the format, length, and structure when they matter.}}
PROCESS
1. Create the first result and label it V0.
2. Create a panel of sub-agents to adversarially review every loop. Complete each pass using only the current version, before comparing the findings. Give the panel three roles:
- An outcome or quality reviewer
- A prompt reviewer
- A reviewer with relevant subject knowledge
3. Begin the first loop.
FOR EACH LOOP:
1. Review the current workflow and result.
2. Identify the biggest remaining failure.
3. Change the smallest part of the prompt needed to fix it.
4. Run the relevant part again.
5. Save the revised result as V1, V2, or V3. Never overwrite an earlier version.
6. Check the result against the definition of done.
AFTER EACH LOOP, REPORT:
- What the panel found
- What changed
- What evidence shows the change helped
- What still needs work
Preserve earlier versions. Stop after loop 3. Do not publish, overwrite source files, or take external actions without my approval. Do not begin until I provide the goal and confirm the run.
OUTPUT
Show these sections in this order:
1. The final result from V3, ready to use.
2. A compact loop summary listing the chosen problem, the focused revision, and the Pass/Fail total for V1, V2, and V3.
3. A version history containing the complete V0, V1, and V2 results under separate headings. Do not repeat V3 in the history.Fill the Fields Correctly
- Task: State the action the AI should perform in one sentence. Be specific about the subject, but put detailed rules under Requirements.
- Audience: Describe who will see or use the result and what they already understand. If the audience doesn't affect the result, delete the entire section.
- Input or source material: Paste any article, notes, data, code, image description, or other material the AI must use. If there's no source material, delete this entire section.
- Requirements: List one non-negotiable rule per bullet. Include required details as well as anything the result must avoid. If you don't have any rules beyond the task and definition of done, delete this section.
- Definition of done: Describe the exact result that would make you consider the task complete. Include the file type, dimensions, length, structure, or other measurable qualities that matter.
Example
The following is an example of the prompt filled out for generating a blog post thumbnail:
Complete the task below, then improve the result through exactly 3 review-and-revision loops.
TASK
Create a thumbnail image for a blog post titled “Use a Loop Prompt For Better AI Results.”
AUDIENCE
Blog readers who use AI tools and want a simple, repeatable way to improve their results.
INPUT OR SOURCE MATERIAL
The article teaches readers to create an initial result, review it from three perspectives, make one focused prompt improvement, and repeat that process for exactly three loops.
REQUIREMENTS
- Use a clean, modern 2D editorial illustration style.
- Show a clear cycle of prompting, reviewing, revising, and improving.
- Keep the composition readable at thumbnail size.
- Use a pleasant background that complements the subject.
- Do not include text in the image.
- Do not use 3D rendering, glossy surfaces, generic AI imagery, or patterned fills that resemble AI watermarks.
DEFINITION OF DONE
Return one polished 1200 × 700 px JPG thumbnail that communicates iterative prompt improvement at a glance and is ready to use as a professional technology blog thumbnail.
PROCESS
1. Create the first result and label it V0.
2. Create a panel of sub-agents to adversarially review every loop. Complete each pass using only the current version, before comparing the findings. Give the panel three roles:
- An outcome or quality reviewer
- A prompt reviewer
- A reviewer with relevant subject knowledge
3. Begin the first loop.
FOR EACH LOOP:
1. Review the current workflow and result.
2. Identify the biggest remaining failure.
3. Change the smallest part of the prompt needed to fix it.
4. Run the relevant part again.
5. Save the revised result as V1, V2, or V3. Never overwrite an earlier version.
6. Check the result against the definition of done.
AFTER EACH LOOP, REPORT:
- What the panel found
- What changed
- What evidence shows the change helped
- What still needs work
Preserve earlier versions. Stop after loop 3. Do not publish, overwrite source files, or take external actions without my approval. Do not begin until I provide the goal and confirm the run.
OUTPUT
Show these sections in this order:
1. The final result from V3, ready to use.
2. A compact loop summary listing the chosen problem, the focused revision, and the Pass/Fail total for V1, V2, and V3.
3. A version history containing the complete V0, V1, and V2 results under separate headings. Do not repeat V3 in the history.Keep in mind that the "review panel" is still a single AI using three reviewer roles unless your AI tool can actually create separate sub-agents.
Footnotes
-
Based on The Loop Method guide, in which The Rundown University AI educator Billy Howell introduces a loop-based process for improving repeated ChatGPT workflows.