Cross-Model Prompt Refinement Loop
Pass prompts between frontier models, curate their edits, and iterate to convergence.
- Difficulty
- Easy
- Time to result
- ~days to results
- Steps
- 5
- Confidence
- 95%
Create a substantive first prompt, then move it to another capable model and ask for improvements tailored to the target reasoning model or task. Review the resulting revision section by section, retaining useful additions and rejecting verbosity, changed intent, or unsupported assumptions. The curated prompt can then be returned to the original model or passed to a third model for another editing pass. Each model acts as a fresh critic with different tendencies, while the user remains the arbiter of quality. The loop works because models can identify omissions in one another’s specifications and expand useful detail, but it depends on active human curation. The goal is convergence on a clearer and more complete prompt, not endless expansion.
Origin
Extracted from Marketing Against The Grain, where the hosts describe building advanced prompts by alternating between Gemini and ChatGPT.
Core principles
- 01Different models expose different weaknesses and opportunities.
- 02Treat each model’s output as editable material, not authority.
- 03Preserve strong sections while requesting targeted improvements.
- 04Human judgment decides which changes survive each round.
How to run it
- 1
Build a complete first version
Draft enough of the prompt that another model can understand its objective, audience, constraints, and output. Do not outsource the initial intent.
Pro tip Use a structured prompt framework for the starting version.
Watch out A vague seed prompt gives the reviewing model too much freedom to redefine the task.
- 2
Request an external critique
Give the prompt to a different model and ask how it should be improved for the intended task or class of reasoning model. Invite both omissions and structural criticism.
Pro tip Name the target model category, such as an advanced reasoning model, when useful.
Watch out Do not assume every proposed change is an improvement.
- 3
Curate the revision
Compare both versions and explicitly identify the parts you like, dislike, or want changed. Restore any important constraints lost during rewriting.
Pro tip Evaluate each addition by whether it improves the final output, not whether it sounds sophisticated.
Watch out Longer revisions can obscure the original goal.
- 4
Repeat across models
Send the curated prompt to another model for a fresh editing pass when the task warrants it. Continue preserving intent while addressing newly identified gaps.
Pro tip Alternate models with meaningfully different strengths.
Watch out Repeated rewriting can introduce contradictions unless each pass is reviewed.
- 5
Stop at convergence
End the loop when another pass produces cosmetic changes rather than material improvements. Save the stable version as a reusable template if the task recurs.
Pro tip Test the prompt on a representative task before declaring it final.
Watch out Do not confuse perpetual iteration with quality assurance.
In the wild
A marketer drafts an O1-style prompt, gives it to Gemini with a request to improve it for an advanced reasoning model, then keeps the stronger context and verification sections while rejecting irrelevant elaboration. The revised prompt is tested back in the original model.
→ The final prompt incorporates strengths from both models without surrendering human control.
Successful video transcripts are distilled into a style in Gemini, refined in ChatGPT, and iterated between the two. The creator compares changes instead of accepting either model’s first draft wholesale.
→ The resulting style template captures more production and narrative detail than a single-pass draft.
Common mistakes
Accepting every revision
Models may add verbosity, alter scope, or remove important constraints. Each revision needs human selection.
Starting with an empty brief
Cross-model refinement improves an existing specification; it cannot preserve intent that was never articulated.
Iterating without a stopping rule
Stop when changes become cosmetic or fail to improve representative outputs.
Is it for you?
Best for
It is best for high-value prompts where additional review justifies using two or more strong models.
Not ideal for
It is inefficient for simple requests or workflows where latency, privacy, or model-access costs prohibit multiple passes.
From the transcript
“But if you flip-flop back and forth in the models, they can edit on each other's work.”
“Now, what I would do is I would go through the prompt and I would say, cool, I like these parts. I don't like these…”
“It works way better to between two or three models.”
From the episode
Use This AI Trick To Get 10x Better Results Every Time