AGENCYSCRIPT
CoursesEnterpriseBlog
đź‘‘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

Step 1: Define the Outcome in Plain WordsDescribe the Finished ThingNote Any Hard ConstraintsStep 2: Ask the Model to Draft the PromptMake the Request ExplicitRead the Draft CriticallyStep 3: Run a Small Test BatchUse Three to Five Real CasesCompare Against Your BaselineStep 4: Critique and RefineFeed the Failures BackIterate Until It StabilizesStep 5: Lock and Store the PromptRecord the Final VersionNote Its BoundariesStep 6: Reuse and Periodically RevisitRun It as a TemplateRe-tune When Inputs DriftA Worked WalkthroughThe Task and OutcomeDrafting and InspectingTesting and RefiningAdapting the Process to Harder TasksSplit Broad Tasks FirstKeep the Outcome HonestTroubleshooting a Stuck ProcessThe Prompt Keeps DriftingThe Output Looks Right but Feels OffFrequently Asked QuestionsHow many refinement rounds should I expect?What if the first generated prompt is already good?Can I do all of this in one long message?Do I need to keep the baseline comparison forever?Key Takeaways
Home/Blog/Build Prompts That Generate Better Prompts, Step by Step
General

Build Prompts That Generate Better Prompts, Step by Step

A

Agency Script Editorial

Editorial Team

·November 23, 2022·6 min read
meta-promptingmeta-prompting how tometa-prompting guideprompt engineering

Reading about meta-prompting is useful, but the technique only sticks once you run it on real work. This article is built as a procedure: do this, then this, then check for that. By the end you will have produced a tested, reusable prompt for a task you actually care about.

The process below works in any standard chat interface. You do not need plugins, code, or a paid plan. What you do need is one genuine task with a clear sense of what a good result looks like, because meta-prompting without a target just spins.

Follow the steps in order the first few times. Once the rhythm is familiar, you will compress them naturally, but the sequence is the scaffolding that keeps you from skipping the parts that matter. If you want the conceptual model behind these steps, The Draft, Critique, Refine Loop for Prompt Generation names the stages this procedure walks through.

Step 1: Define the Outcome in Plain Words

Before you touch the model, write down what success looks like. Not the instructions, the result.

Describe the Finished Thing

Write a sentence or two describing the ideal output as if it already existed: "a 120-word LinkedIn post that sounds confident but not salesy and ends with a question." Outcome descriptions are easier to get right than instructions.

Note Any Hard Constraints

List non-negotiables separately: length limits, banned words, required facts. These are the rails the generated prompt must respect, and naming them now prevents the model from inventing its own.

Step 2: Ask the Model to Draft the Prompt

Now bring in the model, but only as a prompt designer, not yet as the doer.

Make the Request Explicit

Send something like: "Write a detailed prompt I could use to produce [outcome]. Include the constraints I list below. Do not perform the task yet, just give me the prompt." Pasting your outcome and constraints underneath gives it the raw material.

Read the Draft Critically

When the prompt comes back, read every line. The goal is to catch assumptions the model made that you did not intend. This inspection step is non-negotiable and is the most common one beginners skip.

Step 3: Run a Small Test Batch

A prompt is a hypothesis until you run it on real inputs.

Use Three to Five Real Cases

Apply the generated prompt to several genuine examples, not just one. A single success can be luck; consistency across a few cases is signal.

Compare Against Your Baseline

Run your original rough request on the same cases. Seeing them side by side tells you whether the meta-prompt actually helped or just added words. If there is no clear improvement, the prompt needs work.

Step 4: Critique and Refine

This is where most of the value lives, and where people quit too early.

Feed the Failures Back

Tell the model exactly what went wrong: "On example two, the tone was too stiff and it ignored the length limit. Revise the prompt to fix both." Specific feedback produces specific fixes; vague feedback produces vague ones.

Iterate Until It Stabilizes

Repeat the test-and-refine loop until outputs stop improving meaningfully. When two rounds in a row produce roughly equal quality, stop. Chasing further almost always adds complexity without benefit. The reasoning behind this stopping rule is expanded in Habits That Separate Sloppy From Sharp Prompt Generation.

Step 5: Lock and Store the Prompt

A working prompt is an asset, so treat it like one.

Record the Final Version

Paste the finished prompt somewhere durable with a short note on what it is for. Future-you will not remember the context a month from now.

Note Its Boundaries

Write down where the prompt works and where it does not. A prompt tuned for short social posts will not transfer cleanly to long reports, and noting that saves a future failure.

Step 6: Reuse and Periodically Revisit

The payoff arrives on the second, tenth, and fiftieth use.

Run It as a Template

Each new run becomes a fill-in-the-blank: drop in the new specifics and execute. The hard design work is already done.

Re-tune When Inputs Drift

If your task changes or model behavior shifts, run a quick test batch again. A prompt that worked six months ago may need a small refresh. Concrete before-and-after runs of this whole cycle appear in Watching One Prompt Rewrite Another in Real Work.

A Worked Walkthrough

To ground the steps, here is the procedure applied to a single realistic task from start to finish.

The Task and Outcome

Suppose you write short internal announcements and want them consistent. Your outcome description: "a 90-word announcement that is friendly, leads with the change, and ends with where to ask questions." Your constraints: under 90 words, no jargon, one clear call to action.

Drafting and Inspecting

You ask the model to write a prompt for that outcome without executing it. The draft comes back well structured but includes an instruction to "open with a greeting," which you never wanted because your announcements skip greetings. You delete that line. This is the inspection step catching an invented constraint before it shapes every future announcement.

Testing and Refining

You run the cleaned prompt on three past announcements. Two are great; one runs long at 110 words. You tell the model: "On the third, it exceeded the word limit. Tighten the prompt to enforce the cap." The revision holds the limit. A second test on two more cases shows equal quality, so you stop. The prompt goes into your library with a note: "for internal change announcements, under 90 words." The whole cycle took one focused session and now saves time on every announcement.

Adapting the Process to Harder Tasks

The six steps work as written for most tasks, but bigger jobs need a small adjustment.

Split Broad Tasks First

If you find yourself refining past four or five rounds, the task is probably too broad. A prompt trying to handle both short posts and long reports will satisfy neither. Split it into two narrower prompts, each with its own outcome, and the loop converges quickly again.

Keep the Outcome Honest

As tasks grow, there is a temptation to pile every wish into the outcome description. Resist it. A focused outcome produces a focused prompt; an overloaded one produces a vague compromise. When in doubt, write a separate prompt rather than overloading one.

Troubleshooting a Stuck Process

Even a clear procedure hits snags. Two recurring ones are worth knowing how to clear.

The Prompt Keeps Drifting

If each refinement fixes one problem while introducing another, the prompt is carrying too many competing demands. Strip it back to the core outcome, drop the secondary wishes into a separate prompt, and the drift usually resolves. Competing constraints, not the model, are the cause.

The Output Looks Right but Feels Off

Sometimes outputs pass every explicit constraint yet still miss the mark. This usually means your outcome description captured the rules but not the spirit. Rewrite the outcome to describe the feeling you want, not just the measurable traits, and regenerate. The fix lives in the target, not the instructions.

Frequently Asked Questions

How many refinement rounds should I expect?

Usually two to four. If you are past five and still making changes, the task is probably too broad and should be split into smaller, separate prompts.

What if the first generated prompt is already good?

Still run the test batch. A prompt that reads well can still fail on real inputs, and skipping the test is how a flawed prompt slips into regular use.

Can I do all of this in one long message?

You can, but separating the steps into distinct turns keeps the model from half-executing a draft while still revising it. Distinct turns produce cleaner results.

Do I need to keep the baseline comparison forever?

No. The comparison matters during design to prove the prompt helps. Once it is locked in and reused, the baseline has served its purpose.

Key Takeaways

  • Start by describing the outcome and constraints before involving the model.
  • Ask the model to draft the prompt without executing it, then inspect every line.
  • Test on three to five real cases and compare against your rough baseline.
  • Refine with specific feedback and stop when two rounds give equal quality.
  • Store the final prompt with its boundaries and reuse it as a template.

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