Five-Pass AI Iteration Loop
Make AI critique and improve its output five times before human editing
- Difficulty
- Easy
- Time to result
- ~days to results
- Steps
- 6
- Confidence
- 98%
The Five-Pass AI Iteration Loop embeds revision directly into a prompt rather than asking a human to accept or manually repair the first response. The AI creates an initial artifact, identifies specific ways to make it more compelling or useful, and applies those edits in a new version. This critique-and-revision cycle continues until five versions have been delivered, with the fifth pass serving as an explicit stopping condition. Each round should build on the prior output rather than merely restating it. The approach exploits AI's willingness to perform repetitive work and gives the user a stronger starting point for human judgment. Because models may ignore long or complex instructions, the workflow must verify that every iteration was actually produced rather than trusting a claim of completion.
Origin
Extracted from Marketing Against The Grain after Kieran Flanagan demonstrated a campaign prompt designed to run five self-improvement cycles before returning its final output.
Core principles
- 01Treat the first AI response as a draft
- 02Make critique part of the generation process
- 03Apply feedback in the next iteration
- 04Define an explicit stopping condition
- 05Begin human editing only after AI completes the loop
How to run it
- 1
Define the artifact
State exactly what the AI must produce, who it serves, and what a strong result should accomplish.
Pro tip Include concrete quality dimensions such as specificity, novelty, persuasion, or usability.
Watch out A vague objective will produce vague critiques and revisions.
- 2
Generate the first version
Have the AI create a complete initial artifact rather than an outline of what the user should do.
Pro tip Explicitly say that the AI, not the user, must execute every requested step.
Watch out Some models may return instructions instead of performing the work.
- 3
Critique the output
Require the AI to identify focused edits that would make the current version better, more compelling, or more useful.
Pro tip Ask for weaknesses tied directly to the stated quality criteria.
Watch out Generic feedback such as 'make it more engaging' is not actionable enough.
- 4
Apply the critique
Tell the AI to produce a revised version that implements the identified improvements rather than stopping after listing them.
Pro tip Require a full replacement artifact on every pass.
Watch out Models may describe the next iteration without actually generating it.
- 5
Repeat to five passes
Run the critique-and-revision sequence until five complete versions have been delivered, with each one building on the last.
Pro tip Declare that the loop is only complete when the fifth artifact exists.
Watch out Long prompts and tool-specific limits can cause the model to truncate or abandon the loop.
- 6
Perform human review
Evaluate the fifth version for strategic fit, accuracy, originality, and unintended errors before using it.
Pro tip Compare it with the first version to confirm that iteration created meaningful improvement.
Watch out Repeated self-critique does not replace fact-checking or independent judgment.
In the wild
Flanagan asked AI to create an Airbnb campaign using Robert Cialdini's persuasion principles. The intended loop produced a campaign, proposed edits, and then rebuilt it repeatedly. Claude completed five iterations and evolved the concept into an Airbnb Explorers Club and ultimately a broader travel-lifestyle subscription.
→ The fifth concept built on earlier campaigns instead of relying on a generic first response.
A marketer asks AI for a webinar campaign, then requires five complete critique-and-revision passes. Each pass must improve audience specificity, proof, offer clarity, and calls to action. The marketer reviews only the fifth version and selectively restores any strong ideas lost during iteration.
→ The human begins with a more developed campaign while retaining final editorial control.
Common mistakes
Stopping after the critique
Listing suggested edits is not an iteration. Require the model to apply its own feedback and generate the next complete version.
Trusting the loop without checking
A model may claim to iterate while returning only one artifact or delegating the work back to the user. Count the completed versions.
Confusing repetition with improvement
Five variations are not useful if they do not build on explicit criticism of the previous output.
Is it for you?
Best for
Marketers and creators generating campaigns, outlines, scripts, prompts, or other artifacts that benefit from rapid revision.
Not ideal for
High-stakes outputs that require independent factual verification, expert review, or genuinely diverse external perspectives.
From the transcript
“most people they just take the first output from Ai and it's really generic and it's not that useful”
“this here is a loop and that Loop should be done five times and at the end the each time it goes around it should…”
“my tip here is have the AI iterate it on its own work four or five times before you take the output to iterate it…”
From the episode
I Created An Ai Email Course Using ChatGPT