The Rule of Five
Push an AI output through five focused improvement rounds
- Difficulty
- Easy
- Time to result
- ~days to results
- Steps
- 6
- Confidence
- 99%
The Rule of Five treats an AI response as the beginning of a managed revision process rather than a finished deliverable. Generate a first attempt, inspect its largest deficiency, and issue a focused correction. Repeat this process up to five times, using each output as evidence for what the model still misunderstands. Instructions should become clearer and more exact as the loop progresses—for example, moving from a generic request for a LinkedIn post to demands for sharper claims and memorable one-liners. The mechanism resembles coaching an enthusiastic intern: supply context, state the standard, inspect the work, and redirect. Quality comes from iterative specificity, not from searching for one magical prompt.
Origin
Extracted from Marketing Against The Grain
Core principles
- 01Never accept the first output as the finished result
- 02Treat AI like an eager intern that needs context and direction
- 03Make each follow-up more specific than the last
- 04Hold the output to explicit quality standards
- 05Use visible deficiencies to choose the next instruction
How to run it
- 1
Generate the baseline
Give the model the task, available context, and desired output format. Preserve the first response so later versions can be compared against it.
Pro tip Start with enough structure to make the first result diagnostically useful.
Watch out Do not mistake fluent prose for a high-quality answer.
- 2
Name the biggest weakness
Identify the single issue doing the most damage, such as generic language, missing reasoning, weak hooks, or failure to follow instructions.
Pro tip Describe the observable problem rather than saying only that the answer is bad.
Watch out Correcting many unrelated issues at once can make the effect of feedback hard to judge.
- 3
Issue a focused correction
Tell the model exactly what must change and what a successful revision should contain. Restate its role or governing criteria when necessary.
Pro tip Ask for concrete structures such as original, revision, and reason.
Watch out Emotional emphasis without additional information rarely supplies the missing standard.
- 4
Compare versions
Check whether the revision improved the targeted weakness without damaging stronger parts of the original. Keep useful lines even when the complete revision is inferior.
Pro tip Evaluate against explicit criteria rather than novelty alone.
Watch out A larger rewrite is not automatically a better rewrite.
- 5
Repeat to five passes
Use the new output to identify the next highest-value improvement and continue the loop. Stop early only if the result clearly meets the standard.
Pro tip Make later prompts increasingly pointed and testable.
Watch out Iteration without changing the feedback can produce superficial variations.
- 6
Perform the human final pass
Select the best elements, check accuracy, and make the final editorial decision yourself. Treat model explanations as proposals rather than proof.
Pro tip Retain a version history so successful instructions can be reused.
Watch out AI output may remain inaccurate or generic after several iterations.
In the wild
An initial AI draft about hands-on leadership is judged too generic. The user asks for a more pointed version with memorable one-liners suitable for a T-shirt. A later pass produces the sharper line, “Leadership is about action, not administration,” which can anchor a stronger post.
→ Focused iteration turns an undifferentiated draft into a more distinctive content angle.
An analyst generates a summary, then runs separate rounds for missing evidence, unclear logic, unnecessary jargon, weak recommendations, and final concision. Each correction targets one visible failure in the previous version.
→ The final memo is more complete and decision-ready than the first response.
Common mistakes
Accepting fluent first drafts
The first result often sounds polished while remaining generic, incomplete, or poorly matched to the real goal.
Repeating the same prompt
Asking again without supplying new diagnosis or context generates variation rather than directed improvement.
Iterating without verification
Five revisions do not guarantee factual accuracy or good judgment; the final output still needs human review.
Is it for you?
Best for
AI-assisted tasks where quality can be improved through concrete feedback and comparison.
Not ideal for
Deterministic tasks requiring one provably correct answer without human verification.
From the transcript
“Never accept its first output. Ever, ever, ever. I call that the rule of five. Push it five times before you get something really good.”
“We give it more context. We hold it to super high standards. Exactly. And look at the difference that we have.”
“But if we did this three more times, the results will get better each and every time.”
From the episode
Generate 10x Views On A LinkedIn Post With These GPT-4o Prompts