Use the Loop Method to Get Better Results From ChatGPT

Choose a workflow that works but needs the same cleanup each time. For example:

• a writing skill with an inconsistent voice;
• an editing process that misses obvious cuts;
• an image workflow that produces an uneven series.

Run it once without changing anything. Save the instructions and result as your baseline.

Use this quick brief:

Workflow: [NAME OR FOLDER]
Test input: [INPUT]
Current result: [OUTPUT OR FILE]
Problems I noticed: [LIST]
Files or instructions to preserve: [LIST]

The example in this guide uses a location-based image generator skill. It takes a place name and creates three related illustrations. The first Nags Head result became the baseline for testing research, style consistency, and output format.

*The first location image gave ChatGPT a concrete baseline to review.*

Paste the template below with your workflow attached or open in the project:

Improve [WORKFLOW] through exactly 3 loops.

Create a panel of sub-agents to adversarially review every loop. Give the panel three roles:
- an outcome or quality reviewer
- a workflow and prompt reviewer
- a reviewer with relevant subject knowledge

For each loop:
1. Review the current workflow and result.
2. Identify the biggest remaining failure.
3. Change the smallest part of the prompt, script, checklist, or process needed to fix it.
4. Run the relevant part again.
5. 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 Do not publish, overwrite source files, or take external actions without my approval.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.

Change the three roles to fit the job. An editing workflow might use an editor, pacing reviewer, and visual QA reviewer. An image workflow might use an art director, prompt reviewer, and location critic.

Define what the improved workflow must do before you start the loops:

Goal: [CAPABILITY TO IMPROVE]

Definition of done:
- [CHECK 1]
- [CHECK 2]
- [CHECK 3]

Confirm the goal and checks with me before starting loop 1.

Use checks you can verify in the output. For the location-image skill, the result needed three separate images with consistent dimensions, a shared style, and recognizable details from the named place.

Pro tip: If the finish line is fuzzy, ask ChatGPT Work to propose a goal and definition of done. Edit it until you can tell from the next result whether the workflow passed.

After you approve the goal, start the run:

The goal and definition of done are approved. Begin loop 1.

Run all 3 loops. After each loop, show the panel findings, the workflow change, the evidence, and the biggest remaining issue before continuing.

Check each report for four things:

• Did the sub-agent panel find a specific failure?
• Did ChatGPT change the workflow instead of only repairing the output?
• Did the next result improve against the same checks?
• Is each change saved separately so you can compare versions?

In the location-image example, the panel found generic place details and weak consistency rules. A later two-panel result exposed an ambiguous instruction, which was corrected before the next loop.

*The reviewed workflow produced a more deliberate three-image set with shared visual rules.*

If a result reveals an unclear instruction, correct the instruction before the next loop:

The last result exposed this ambiguity: [AMBIGUOUS INSTRUCTION].

Update the workflow so it means: [PRECISE REQUIREMENT]. Apply that correction in the next loop and verify it in the output.

Pro tip: When a result goes sideways, name the ambiguity before you run another loop. A specific correction—such as “three independent artworks, not a two-panel set”—gives the next pass something it can actually test.

Change the input without rebuilding the workflow:

Use the updated [WORKFLOW] on [SECOND INPUT].

Keep the reviewed instructions and definition of done. Save the result separately from the first example, then report:
- whether it passed each check
- what still needs my review
- which manual judgments should become scripts, rules, or checks

*A second location showed whether the revised research and series rules worked beyond the first example.*

If the second test fails, fix the workflow and repeat the test with another input. Do not keep patching only the original output.