One-Section-at-a-Time Generation Rule
Break large outputs into reviewable sections and approve each before continuing
- Difficulty
- Starter
- Time to result
- ~days to results
- Steps
- 5
- Confidence
- 97%
The One-Section-at-a-Time Generation Rule replaces a single oversized request with a sequence of bounded assignments. Begin by defining the full deliverable and dividing it into coherent sections, but ask the model to generate only the first one. Review that section for accuracy, tone, completeness, and alignment with the source material. Iterate until it is acceptable, then explicitly instruct the model to continue to the next section. This creates regular quality gates and keeps corrections local rather than forcing a rewrite of an entire document. After all sections are approved, conduct a final consistency pass to reconcile terminology, dates, dependencies, and duplicated content across the assembled artifact.
Origin
Extracted from Marketing Against The Grain during Sully Omar's live demonstration of converting a long analytical output into separate internal team memos.
Core principles
- 01Specific tasks produce more controllable outputs
- 02Review should occur before the next section is generated
- 03Large requests should be decomposed by deliverable
- 04Iteration is faster with responsive models
- 05Approved sections form the final artifact
How to run it
- 1
Map the full deliverable
Describe the final artifact and list its logical sections. Establish any requirements that must remain consistent throughout.
Pro tip Divide by audience, initiative, chapter, or decision.
Watch out Sections that overlap heavily will create duplication.
- 2
Request one bounded section
Ask the model to generate only the first section and state where it should stop. Supply the source context relevant to that unit.
Pro tip Use explicit language such as 'only do' the named section.
Watch out A vague stopping point may cause the model to continue through the whole deliverable.
- 3
Review and iterate
Check the section's facts, reasoning, tone, and usefulness. Request focused revisions before allowing the workflow to proceed.
Pro tip Resolve structural problems in the first section because they may recur later.
Watch out Do not approve polished prose that misstates the source.
- 4
Advance deliberately
Once the section meets the standard, instruct the model to move to the next one. Repeat the same review gate for every section.
Pro tip Keep a short list of approved conventions for later sections.
Watch out Automatic continuation removes the main quality-control benefit.
- 5
Reconcile the assembled artifact
Review the complete document for consistent terminology, sequencing, ownership, dates, and transitions. Remove duplication introduced by separate generation cycles.
Pro tip Perform this review against the original outline and source material.
Watch out Individually strong sections can still conflict when combined.
In the wild
Instead of requesting memos for every recommendation in one prompt, a founder asks the model to write only the churn-reduction memo. After reviewing and revising it, the founder authorizes the next initiative.
→ The resulting memos are more specific and easier to validate than a single bulk-generated set.
An analyst divides a report into methodology, findings, limitations, and recommendations. Each section is generated and checked against the evidence before the model proceeds.
→ Errors are caught early and do not propagate through the entire report.
Common mistakes
Requesting every section at once
The model may compress later sections, overlook requirements, or repeat itself when asked for a large deliverable in one turn.
Advancing without approval
Uncorrected assumptions and style problems can propagate into every later section.
Skipping final reconciliation
Separately generated sections may use inconsistent terms or contain conflicting dates and responsibilities.
Is it for you?
Best for
It is best for long reports, plans, memos, and other outputs that can be divided into independently reviewable sections.
Not ideal for
It is not ideal for tightly coupled outputs that cannot be evaluated until generated as a complete whole.
From the transcript
“What you want to do with any models, you want to like break it down into as many sort of specific tasks as possible.”
“So here I said only do it for the trend reduction initiative.”
“It looks good. I can iterate and I'd be like, okay, okay, you know, looks good to me. Move on to the next.”
From the episode
Which AI Model Should You Use? (Claude vs GPT & O1 Pro Live Prompt Guide)