AGENCYSCRIPT
CoursesEnterpriseBlog
👑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
© 2026 Agency Script, Inc.·
Privacy PolicyTerms of ServiceCertification AgreementSecurity

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

On This Page

Before You Start the LoopDefine the TargetSet the Stopping RuleDuring Each Refinement TurnDiagnose Before You InstructKeep the Context CleanBefore You Call It DoneVerify Against the BarCapture What WorkedThe Recurring Failure Modes This PreventsThe Random WalkThe Perfectionism SpiralSilent Constraint DriftContext Loss in Long ThreadsAdapting the Checklist to the StakesHigh-Stakes WorkLow-Stakes WorkHow to Use This ChecklistAs a Solo OperatorAs a TeamMake It a Living DocumentFrequently Asked QuestionsWhich checklist item matters most?Isn't running a checklist slower than just iterating?Why cap the number of turns?What should I do if the model loses track of my latest draft?Do I really need to save successful prompt sequences?Key Takeaways
Home/Blog/Checks for Before, During, and After a Refinement Loop
General

Checks for Before, During, and After a Refinement Loop

A

Agency Script Editorial

Editorial Team

·September 6, 2020·7 min read
prompting for iterative refinement loopsprompting for iterative refinement loops checklistprompting for iterative refinement loops guideprompt engineering

A refinement loop fails for predictable reasons: no defined target, vague feedback, no stopping rule, lost context, or polishing past the point of diminishing returns. Each of these is preventable with a small upfront discipline. The trouble is that in the moment, you are focused on the content and forget the process.

This checklist exists to be used, not admired. It is organized into three phases—before the loop, during each turn, and before you ship—so you can run it as a literal sequence. Each item carries a one-line justification so you understand why it earns a place, not just what it asks. Skim it once to internalize the logic, then keep it open while you work.

It pairs naturally with the underlying structure in The Draft-Diagnose-Constrain Method for Iterative Refinement Loops; the framework explains the shape, and this checklist enforces it. Think of the framework as the theory and the checklist as the practice—the thing you actually run at the keyboard when a draft disappoints and you are deciding what to do next.

A word on how to read what follows. The items are deliberately short and concrete, because a checklist you have to interpret is a checklist you will abandon. Each is phrased as an action you can take in seconds, and the justification beneath it is there so you trust the item enough to keep doing it under deadline pressure, when the temptation to skip straight to nudging is strongest.

Before You Start the Loop

Define the Target

  • Write down what good looks like in one sentence. A loop without a target is a random walk; the model has nothing to converge on.
  • Find or write a reference example. Showing the model what you want is more reliable than describing it, especially for tone and style.
  • State the hard constraints up front. Length, format, audience, and forbidden words belong in the first prompt, not in turn four.

Set the Stopping Rule

  • Decide your good-enough bar before the first output. Defining "done" in advance prevents the perfectionism spiral.
  • Cap your turns. Agree with yourself that if you are past three or four turns without convergence, you restart rather than nudge again.

During Each Refinement Turn

Diagnose Before You Instruct

  • Name the specific defect, not a vibe. "The second paragraph makes an unsupported claim" beats "make it better"; the model can only fix what you name.
  • Show the discrepancy when possible. Paste the wrong output next to the expected result; demonstration corrects faster than description.
  • Pin the invariant. State explicitly what must stay fixed, or the model will rewrite parts you were happy with.

Keep the Context Clean

  • Check the model is still seeing the latest version. In long threads, the model may anchor on an earlier draft; restate the current text if in doubt.
  • Change one thing at a time on hard problems. Bundling three fixes into one turn makes it impossible to tell which instruction worked.
  • Avoid stacking vague nudges. Two consecutive "tighten it up" turns is a signal to stop and rediagnose, a failure pattern detailed in Six Refinement Loops That Turned Mediocre AI Output Into Shippable Work.

Before You Call It Done

Verify Against the Bar

  • Re-read the output against your one-sentence target. Confirm it actually hits what you defined, not just that it improved.
  • Check the constraints held. Length, format, and forbidden-word rules drift across turns; verify the final version still complies.
  • Confirm no facts were invented. Refinement turns sometimes add unsupported claims; spot-check anything that reads as a statistic or quote.

Capture What Worked

  • Save the prompt sequence that succeeded. A loop you can reproduce is worth more than a one-off result.
  • Note any reusable constraint. If a particular phrasing reliably fixes a recurring defect, it belongs in your starting prompt next time.

The Recurring Failure Modes This Prevents

The Random Walk

The most common loop failure is iterating toward no fixed point. You say "punchier," then "more energy," then "still not quite," and each turn drifts somewhere new. The before-phase items—target and reference example—exist specifically to give the model a destination so it converges instead of wandering.

The Perfectionism Spiral

The second failure is not stopping. Without a stopping rule, every output can always be marginally improved, and writers burn hours on changes no reader will notice. The good-enough bar, set before you start, is the antidote. Defining done is not a nicety; it is the only thing that ends the loop on purpose rather than by exhaustion.

Silent Constraint Drift

The third failure is subtle: across several refinement turns, the output quietly violates a constraint you set early—it creeps over the word limit, or reintroduces a forbidden phrase. The before-you-ship verification catches this. Loops do not preserve early constraints automatically, so you have to re-check them at the end.

Context Loss in Long Threads

In extended loops, the model can anchor on an earlier draft and refine the wrong version. The during-phase reminder to restate the current text resolves this. When in doubt, paste the latest version and instruct the model to refine that exact text.

Adapting the Checklist to the Stakes

High-Stakes Work

For client-facing copy, production code, or analysis that drives a decision, run every item. The few minutes the full checklist costs are trivial against the cost of shipping a flawed output to a client or a codebase.

Low-Stakes Work

For a disposable internal note or a quick brainstorm, compress aggressively. A one-sentence target and a glance at the result may be all you need. The checklist scales down; forcing the full sequence on throwaway work wastes the discipline it is meant to protect. The judgment of when to apply which depth is itself a skill, informed by the decision logic in Iterate, Restart, or Rewrite the Prompt When Output Disappoints.

How to Use This Checklist

As a Solo Operator

Keep the three phases visible while you work. The before-phase takes thirty seconds and saves the most time. If you skip only one phase, never skip defining the target and the stopping rule.

As a Team

Turn the during-phase rules into a shared norm—especially "name the defect, not a vibe." Teams that standardize the diagnose step see faster convergence regardless of who is writing, as covered in Which Numbers Tell You a Refinement Loop Is Actually Healthy.

Make It a Living Document

Treat this checklist as a starting point, not scripture. As you run loops, you will notice recurring defects specific to your work—a tone your clients dislike, a structural habit the model keeps repeating. Add a before-phase constraint that heads off each one, and your personalized checklist will prevent your most common failures before they happen. The best version of this list is the one you have edited to match the work you actually do, which is why capturing what works in the final phase feeds directly back into a stronger before-phase next time.

Frequently Asked Questions

Which checklist item matters most?

Defining the target in one sentence before you start. Nearly every failed loop traces back to a missing or fuzzy target, because the model cannot converge on something you have not specified.

Isn't running a checklist slower than just iterating?

The before-phase costs about thirty seconds and the during-phase rules cost a few extra words per turn. Both reliably cut total turns, so the loop finishes sooner overall even though each step feels more deliberate.

Why cap the number of turns?

Because past three or four turns without convergence, the problem is almost always a missing constraint, not something another nudge will fix. The cap forces you to stop and rediagnose instead of spiraling.

What should I do if the model loses track of my latest draft?

Restate the current version explicitly. In long threads the model can anchor on an earlier draft; pasting the latest text and saying "refine this exact version" resolves most context drift.

Do I really need to save successful prompt sequences?

If you ever repeat the task, yes. A reproducible loop turns a lucky result into a reliable one, and recurring fixes can graduate into your default starting prompt.

Key Takeaways

  • Run the loop in three phases: set up the target before, diagnose precisely during, and verify against the bar before shipping.
  • A one-sentence target and a stopping rule, defined before the first prompt, prevent most failed loops.
  • Name the specific defect and pin the invariant on every turn instead of stacking vague nudges.
  • Verify constraints held and no facts were invented before you call the output done.
  • Save the prompt sequences that work so a one-off success becomes repeatable.

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

Prompt Quality Decides Whether AI Earns Its Keep

Prompt quality is the single biggest variable in whether AI delivers real work or expensive noise. The model matters, the platform matters — but the prompt you write determines whether you get a first

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

Counting the Real Cost of Every Token You Send

Tokens and context windows sit at the intersection of AI capability and operational cost—yet most business cases treat them as technical footnotes. That's a mistake that costs real money. Every time y

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

Rolling Out AI Hallucinations Across a Team

Most teams discover AI hallucinations the hard way — 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, 2026·11 min read

Ready to certify your AI capability?

Join the professionals building governed, repeatable AI delivery systems.

Explore Certification