AGENCYSCRIPT
CoursesEnterpriseBlog
πŸ‘‘FoundersSign inJoin Waitlist
AGENCYSCRIPT

Governed Certification Framework

The operating system for AI-enabled agency building. Certify judgment under constraint. Standards over scale. Governance over shortcuts.

Stay informed

Governance updates, certification insights, and industry standards.

Products

  • Platform
  • Certification
  • Launch Program
  • Vault
  • The Book

Certification

  • Foundation (AS-F)
  • Operator (AS-O)
  • Architect (AS-A)
  • Principal (AS-P)

Resources

  • Blog
  • Verify Credential
  • Enterprise
  • Partners
  • Pricing

Company

  • About
  • Contact
  • Careers
  • Press
Β© 2026 Agency Script, Inc.Β·
Privacy PolicyTerms of ServiceCertification AgreementSecurity

Standards over scale. Judgment over volume. Governance over shortcuts.

On This Page

Why One-Shot Prompting UnderperformsWhat you cannot know up frontWhat a loop gives youThe Anatomy of a Refinement LoopGenerateCritiqueReviseCheckDefining the Standard Before You LoopA usable standard is concreteWrite it before the first generationCritiquing Without Sabotaging YourselfMake the critique specificUse the model as a critic carefullyChanging One Thing at a TimeWhy isolation mattersThe cost of speedKnowing When to StopStop whenAvoid the polish trapBuilding the Loop Into Your HabitsHabits that make it stickFrequently Asked QuestionsHow many iterations does a good loop usually take?Should I let the model critique its own output?Is iterative refinement just slower one-shot prompting?What if I do not know what good looks like yet?When does a refinement loop stop being worth it?Can I run refinement loops at scale across many pieces?Key Takeaways
Home/Blog/Mastering the Generate-Critique-Revise Cycle With Prompts
General

Mastering the Generate-Critique-Revise Cycle With Prompts

A

Agency Script Editorial

Editorial Team

Β·August 28, 2020Β·6 min read
prompting for iterative refinement loopsprompting for iterative refinement loops guideprompting for iterative refinement loops guideprompt engineering

The single biggest gap between people who get mediocre results from a model and people who get excellent ones is not prompt wording. It is whether they treat generation as one shot or as a loop. The people who win revise. They generate a draft, examine it against a standard, identify the specific gap, and feed that back into the next attempt. They do this deliberately and repeatedly, and the output converges on something good.

This guide is a complete overview of that practice. It covers why one-shot prompting underperforms, how to structure a refinement loop so it converges rather than wanders, what each stage of the loop requires, and how to know when to stop. It is written for someone who wants to move from occasional good outputs to reliably good ones.

The companion pieces drill into specific angles: a step-by-step procedure, the common mistakes, and the hard-won best practices. This piece gives you the whole shape so those drill-downs have a frame to hang on.

Why One-Shot Prompting Underperforms

The instinct to perfect a prompt and fire it once is natural and usually wrong. It assumes you can fully specify what you want in advance, which you almost never can.

What you cannot know up front

  • The exact phrasing that will surface the model's best answer.
  • Which constraints will conflict until you see them collide in output.
  • What "good" looks like in detail before you see a draft to react to.

What a loop gives you

A loop converts the problem from prediction to reaction. Instead of specifying perfectly, you generate something, react to it, and steer. Reacting is far easier and more accurate than predicting, which is why iterative work outperforms one-shot work on any task with real quality stakes.

The Anatomy of a Refinement Loop

A refinement loop has four stages, and skipping any of them breaks it.

Generate

Produce a draft. Early in the loop, this draft is a probe, not a product. Its job is to give you something concrete to react to.

Critique

Compare the draft against an explicit standard. The critique must be specific: not "make it better" but "the second section assumes knowledge the reader does not have."

Revise

Feed the specific critique back as a targeted instruction. Change one thing, not everything, so you can see what the change did.

Check

Confirm the revision addressed the critique without breaking something else. Then decide whether to loop again or stop.

These four stages, run with discipline, are what separate a loop that converges from a conversation that drifts. The procedural version is laid out in A Step-by-Step Approach to Prompting for Iterative Refinement Loops.

Defining the Standard Before You Loop

The loop only converges if you know what you are converging toward. A vague standard produces a wandering loop that never finishes.

A usable standard is concrete

  • It names specific qualities the output must have.
  • Each quality is checkable, not a matter of vibe.
  • It fits the task and the reader, not generic notions of quality.

Write it before the first generation

Defining the standard after you start generating lets each draft redefine what good means, which is how loops go in circles. Set the target first, then aim at it. This discipline overlaps heavily with audience work; see Tailoring Prompts to Readers: Direct Answers to Real Questions for how reader-task checklists work.

Critiquing Without Sabotaging Yourself

The critique stage is where most loops fail, because vague critique produces vague revision.

Make the critique specific

  • Point to a location: which sentence, which section.
  • Name the gap: what standard it misses.
  • State the direction: what would close the gap.

Use the model as a critic carefully

You can ask the model to critique its own output, and it often surfaces real issues. But treat its critique as a candidate, not a verdict. Models can rationalize their own drafts, so the human keeps the standard.

Changing One Thing at a Time

The temptation in revision is to fix everything at once. Resist it.

Why isolation matters

  • When you change five things and the output improves, you do not know which change helped.
  • When a change makes things worse, you cannot cleanly back it out.
  • Compounding changes hide interactions you would want to notice.

The cost of speed

Changing one thing at a time feels slower but converges faster, because every step teaches you something usable. The mistakes piece, 7 Common Mistakes with Prompting for Iterative Refinement Loops, treats this as a central failure mode.

Knowing When to Stop

A loop without a stopping rule runs forever, polishing past the point of value.

Stop when

  • The output meets the standard you set up front.
  • The remaining gaps cost more to close than they are worth.
  • Successive revisions stop producing meaningful improvement.

Avoid the polish trap

Once the output clears the standard, additional loops usually trade real time for marginal gains nobody will notice. A defined standard is also a defined finish line, which is half of why setting it first matters.

Building the Loop Into Your Habits

The difference between knowing about loops and benefiting from them is habit. The loop has to be how you work, not something you remember to do.

Habits that make it stick

  • Always generate a probe draft rather than agonizing over the perfect first prompt.
  • Always write the standard before generating.
  • Always critique to a location and a direction.
  • Always change one thing per revision.

These habits, once internalized, make refinement feel natural rather than laborious. The opinionated version of them is in Prompting for Iterative Refinement Loops: Best Practices That Actually Work.

Frequently Asked Questions

How many iterations does a good loop usually take?

Most well-run loops converge in three to five iterations when the standard is clear and each revision changes one thing. If you are routinely going past seven, the usual cause is a vague standard or compounding changes that hide what is actually improving. Tightening either one shortens the loop.

Should I let the model critique its own output?

You can, and it often surfaces genuine issues, but treat its critique as a candidate rather than a verdict. Models can rationalize their own drafts. Keep the standard in human hands and use the model's self-critique as input, not as the final judgment of fitness.

Is iterative refinement just slower one-shot prompting?

No, it is a different strategy. One-shot prompting bets you can fully specify the output in advance; refinement bets that reacting to a concrete draft is more accurate than predicting. On tasks with real quality stakes, reacting beats predicting, which is why the loop wins despite using more steps.

What if I do not know what good looks like yet?

Then your first loop is partly about discovering the standard, but you should still write down a provisional one and refine it as you go. The danger is letting every draft silently redefine good, which prevents convergence. Make the standard explicit even when it is rough.

When does a refinement loop stop being worth it?

When the output already meets your standard, or when successive revisions stop producing improvements anyone would notice. Polishing past the standard trades real time for invisible gains. A clear standard doubles as a finish line, which is the simplest defense against over-polishing.

Can I run refinement loops at scale across many pieces?

Yes, by standardizing the standard and the critique format so the loop is the same shape every time. The procedure becomes a template: same probe, same checklist, same one-change-per-revision discipline. Standardization is what lets a loop that works once work a hundred times.

Key Takeaways

  • One-shot prompting underperforms because you cannot fully specify good output in advance; loops let you react instead.
  • A refinement loop has four stages: generate, critique, revise, check, and skipping any breaks it.
  • Define a concrete, checkable standard before the first generation so the loop converges instead of wandering.
  • Make critiques specific to a location and a direction, and change one thing per revision.
  • Stop when the output meets the standard; polishing past it trades real time for invisible gains.

Search Articles

Categories

OperationsSalesDeliveryGovernance

Popular Tags

prompt engineeringai fundamentalsai toolsthe difference between AIMLagency operationsagency growthenterprise sales

Share Article

A

Agency Script Editorial

Editorial Team

The Agency Script editorial team delivers operational insights on AI delivery, certification, and governance for modern agency operators.

Related Articles

General

Rolling Out AI Hallucinations Across a Team

Most teams discover AI hallucinations the hard way β€” a confident-sounding wrong answer makes it into a client deliverable, a legal brief, or a published report. The damage isn't just to the output; it

A
Agency Script Editorial
June 1, 2026Β·11 min read
General

Case Study: Large Language Models in Practice

Most teams that fail with large language models don't fail because the technology doesn't work. They fail because they treat deployment as a one-time event rather than a discipline β€” pick a model, wri

A
Agency Script Editorial
June 1, 2026Β·11 min read
General

Thirty-Second Wins Breed False Confidence With LLMs

Working with large language models is deceptively easy to start and surprisingly hard to do well. You can get a useful output in thirty seconds, which creates a false confidence that compounds over ti

A
Agency Script Editorial
June 1, 2026Β·10 min read

Ready to certify your AI capability?

Join the professionals building governed, repeatable AI delivery systems.

Explore Certification