Reading Time: 5 minutes

Hey Prompt Lover,

Module 5 is done. Module 6 starts today.

Before we get into it let me tell you where we are in the research because we're at a point in The Prompt Report that genuinely surprised me when I first read it.

Everything in Modules 1 through 5 was about making you a better prompt writer. Better structure. Better examples. Better reasoning instructions. Better sequencing. Better quality control. Fifteen newsletters of techniques that assume you are the one doing the work of improving the prompt.

Module 6 is different.

Module 6 is about what happens when you stop trying to fix the prompt yourself and let the AI fix it for you.

I want to be honest with you about my reaction when I read this section of the paper. My first thought was that it sounded like the kind of thing that works in a research lab and falls apart in real workflows. My second thought was that I should actually test it before having an opinion about it.

I tested it for three weeks before writing this newsletter.

It works. Not always perfectly. Not as a replacement for knowing what you're doing. But as a tool for diagnosing broken prompts and generating improved versions, it works consistently enough that I've built it into my process for any prompt that matters and isn't performing.

Here's what changed my mind.

Hiring in 8 countries shouldn't require 8 different processes

This guide from Deel breaks down how to build one global hiring system. You’ll learn about assessment frameworks that scale, how to do headcount planning across regions, and even intake processes that work everywhere. As HR pros know, hiring in one country is hard enough. So let this free global hiring guide give you the tools you need to avoid global hiring headaches.

Download the free guide today

Here's Why This Matters

Here's the thing about prompt problems that makes them hard to fix from inside the process.

When you write a prompt, you know what you mean. You have context the AI doesn't have. You're reading the prompt through that context and it makes sense to you because you're filling in the gaps with your own understanding.

The gaps that confuse the AI are invisible to you because you're not experiencing them as gaps.

When you ask the AI to evaluate the prompt, it reads it the way it will actually run it. Without your context. Without your assumptions. Without your intent filling in the missing pieces. It experiences the gaps as gaps. And it can tell you where they are.

That's the core value of meta-prompting. Not that the AI is smarter than you about prompts. It's that the AI reads your prompt the way the AI reads your prompt. Which is different from how you read it. And that difference is exactly the information you need when the output isn't right.

The research documented this formally through a technique called APE — Automatic Prompt Engineer.

The researchers ran automated experiments generating multiple prompt variations, scoring them, and iterating toward the best performing version. The automated approach found prompt improvements that human engineers missed.

Not because the algorithm was clever. Because it was testing variations the humans didn't think to try.

You don't need the full APE system to get the benefit. The meta-prompt below gives you the same diagnostic insight in four minutes.

What You'll Learn In This Newsletter

By the end of this issue, you'll have:

• A clear explanation of why you can't always see what's wrong with your own prompt

• The exact meta-prompt that diagnoses broken prompts and generates improved versions

• A three-iteration process for using AI to systematically improve any underperforming prompt

• The specific questions to ask that produce useful diagnosis rather than vague suggestions

Let's get started.

What Most People Do Wrong

When a prompt produces bad output, most people do one of three things.

They rewrite the prompt from scratch, which throws away everything that was working along with everything that wasn't. They add more instructions to the existing prompt, which usually makes things worse because the prompt was already telling the AI too many things. Or they switch to a different AI tool, which changes the model but not the prompt problem.

None of these approaches diagnose the actual issue. They're all guesses at a solution without a clear understanding of the cause.

The result is prompt iteration that looks like progress but isn't. You're not improving the prompt systematically. You're changing it randomly and hoping the next version works. Sometimes it does. For the wrong reasons. So when it breaks again you're back to guessing.

Meta-prompting breaks this cycle by making the diagnosis explicit before the rewrite happens. You find out what's actually wrong before you change anything. Then you fix the things that are wrong instead of changing everything and hoping.

The Prompt That Works

▼ COPY THIS PROMPT:

How To Use This Prompt

Step 1: Paste your broken prompt into the designated section exactly as written. Don't edit it first. Don't clean it up. The meta-prompt needs to see the actual problem, not a tidied version of it.

Step 2: Be specific in the "supposed to produce" section. This is where most people are too vague. "Better output" tells the AI nothing. "A 200-word product description in a conversational tone for a technically skeptical B2B buyer" tells it everything.

Step 3: Be equally specific in the "actually producing" section. "It's not good" is not useful. "The output is written in formal corporate language when I need it to sound like a direct email from a founder" is useful. The gap between intended and actual is the diagnosis. The more clearly you describe it, the better the diagnosis.

Step 4: Read Step 1 and Step 2 of the output before you look at the rewritten prompt. The diagnosis and explanation are the most valuable part. If you understand why the original prompt was failing, you'll be able to spot the same problem in future prompts before it costs you five revision rounds.

Step 5: Test the rewritten prompt. If it's better but not quite right, run the meta-prompt again on the improved version. Three iterations covers most prompt problems completely. The research found that automated prompt improvement converges toward a stable high-performing version within three to five iterations. The manual version follows the same pattern.

Why This Prompt Works

Meta-prompting works because it changes who is reading the prompt.

When you read your own prompt, you're reading it with your intent. You know what you meant. The gaps don't register as gaps because you're filling them in. When the AI reads your prompt, it has no intent to fill in the gaps with. It reads exactly what's there. Which means it experiences the prompt the way it will execute the prompt.

Asking the AI to diagnose the prompt is asking it to report on its own experience of reading it. That report is useful in a way that your own reading never can be. Not because the AI is smarter. Because it's on the other side of the gap you can't see from your side.

The four-step structure matters too. Diagnosis before rewrite prevents the common failure mode of jumping to a solution before the problem is understood. Prediction after rewrite gives you a way to evaluate whether the improved prompt actually addresses the diagnosis or just sounds better.

The APE Method Worth Knowing

The research documents a more formal version of this through Automatic Prompt Engineer.

The full APE system works like this: generate ten to twenty variations of your prompt automatically, score each one against your target output, identify the highest-scoring versions, generate variations of those, score again, repeat until performance stabilizes.

You don't need the automation to use the principle.

Manually: run your meta-prompt, get an improved version, test it, run the meta-prompt again on the improved version, test again. Three cycles. You're doing by hand what APE does algorithmically. Slower. But accessible without any technical setup and useful for any prompt in any workflow.

The research found APE-improved prompts outperformed human-engineered prompts on multiple benchmarks. Not always. But consistently enough that the method is worth building into your process for any prompt you're going to use repeatedly.

Quick Reality Check

The research paper includes a finding I've thought about a lot since I read it. In several experiments, the automatically generated prompt improvements that performed best were phrasings the human engineers said they never would have tried. Not because the algorithm was creative.

Because the algorithm had no assumptions about what should work. It just tested everything. The best human prompt engineers I know have a version of this quality — they test things that seem wrong and let the results decide. Meta-prompting builds that into a process anyone can use.

What Most People Write Instead

Typical approach to a broken prompt: Delete it. Start over. Write a new prompt based on a slightly different interpretation of what you need. Run it. Hope.

Why it fails: You're changing everything when only one or two things are wrong. You throw away the parts that were working. You introduce new problems while solving old ones. And because you're guessing at the cause, you might fix the symptom without fixing the problem.

The meta-prompt approach: Keep the original. Diagnose specifically. Fix precisely. Test the improved version. Run one more diagnosis cycle if needed.

Why it wins: You know what changed and why. The improvements are targeted. The parts that worked stay in. And you build a clear understanding of what was wrong that makes you better at writing the next prompt from scratch.

The Bigger Lesson Here

The best prompt engineers I've met share one quality that separates them from everyone else.

They don't assume they know what's wrong. They find out.

Meta-prompting is a tool for finding out. Not guessing. Not rewriting from instinct. Actually diagnosing. Actually understanding. Then fixing what the diagnosis identified and nothing else.

That discipline — diagnose before rewrite, test after, iterate deliberately — is worth more than any individual technique in this series. It's how you get better at prompting systematically instead of accidentally.

What Changes After Using This

The first time you run the meta-prompt on something broken, the diagnosis will probably identify at least one problem you hadn't noticed. That's the immediate value.

After a few weeks of using it consistently on underperforming prompts, something more useful happens. You start seeing the problems in your prompts before you run them. The diagnostic questions the meta-prompt asks become questions you ask yourself while writing. Role section clear enough? Context complete? Format instruction contradicting tone instruction?

The tool teaches you to read your own prompts the way the AI reads them. That skill doesn't go away when you close the meta-prompt. It stays.

Try This Right Now

Find one prompt in your current workflow that's been producing output you're settling for rather than output you're happy with.

Paste it into the meta-prompt above. Fill in what it's supposed to produce and what it's actually producing.

Read the diagnosis. Before you read the rewritten version, decide whether you agree with the diagnosis. Then look at what the rewrite changed and see if those changes match what the diagnosis identified.

That alignment between diagnosis and fix is what systematic prompt improvement looks like. It's different from rewriting and hoping. Once you see the difference you won't go back.

What's Coming Next

Next newsletter we go deeper into Module 6 with ProTeGi — textual gradient descent applied to prompts. The research team treated prompt improvement like a machine learning optimization problem. The results outperformed human engineering on multiple benchmarks.

The concept sounds technical. The application is straightforward. And the finding it produced changed how I think about why prompts work — because the prompts that performed best in the automated tests were often ones that shouldn't have worked according to conventional prompting logic.

That finding alone is worth the next newsletter.

See you then.

Reply With Your Results

Run the meta-prompt on something broken this week and reply with what the diagnosis found.

Tell me what problem it identified that you hadn't seen. Tell me if the rewritten version performed better. Tell me if the diagnosis taught you something about how you write prompts that you'll carry forward.

I read every reply.

— Prompt Guy