I set out to build the Ferrari of prompts. I ended up with a VW.

I set off with a mature vision, the know-how, and a real appetite for AI topics on board. With the healthy urge to want to understand things and solve problems. And with the slightly unhealthy urge to want to perfect things, I got myself hopelessly lost.

My vision: one prompt that generates prompts.

Not just any run-of-the-mill jobs, but quality master prompts. Prompts you use over and over or that you can build on anew. Prompts worthy of a place in a catalog. User-friendly prompts that get the AI to ask the right questions, prompts that are genuine everyday helpers or specialized problem-solvers. The kind you can snap together like Lego into something even bigger.

I wanted all of that. Not right away — over the course of a months-long evolution, sure. But I wanted all of it in a single prompt.

But “Nach fest kommt ab, und nach ab kommt Arbeit.” applies in IT too.

Gemini gave up, and ever since that undertaking I’ve had a written-down pinky swear with Claude, in all our projects, to deliberately and critically check for Verschlimmbesserung.

But let’s start from the beginning. Good prompting was, is, and remains important if you want to coax the really good results out of an LLM. Which makes sense, really — the prompt is the instruction to the machine. You used to have to actually program and speak in genuinely complicated computer languages. With the arrival of AI, instruction and data are blended together and wrapped in perfectly ordinary everyday language, accessible to everyone.

A mumbled “just fix the mistakes in the text and in Spanish” has thereby become a working programming command.

And it’s precisely in this simplicity that a double edge lies. Because every prompt works and delivers results. There’s no truly wrong prompt or obviously bad prompt. But there are bad prompts all the same — and methods for making better ones. If you want exceptionally, surprisingly good results, there’s no getting around a little optimizing. Which, in the end, means describing more precisely what you want.

To stick with the mini-example from above: Which mistakes? Only grammatical ones, or stylistic ones too? Do you want just the Spanish text as output, or also the corrected German version? And do you want to know where the mistakes were, or what all got changed?

Of course, if you don’t need or want any of that, you don’t have to say so. But usually you do have an implicit set of expectations, and it feels systemically more wrong when they’re let down by a machine. With AI you have to make your expectations explicit — and you don’t even have to spare anyone’s human feelings while doing it.

And the AI result gets better still when you hand over context and the whole picture, instead of the individual instructions “remove mistakes” and “translate into Spanish.”

Let’s flesh out the example concretely, then: Say you had really bad luck on your last trip to Spain. The hotel wrongly charged your credit card after the fact. You’ve pulled together all the information and want to sort it out directly with the hotel. With that information and input, the AI can help you far more than you might imagine. Work out the tone, the psychologically and legally best approach, together with you. Guide and advise you concretely through the entire process.

So the most important step is pure mindset, on your end. Not petty micro-instructions, but also not just generic, half-baked sentence fragments. If you just feed the AI properly, it can answer better than you think.

For that it needs precisely the information that’s often obvious and self-evident to us, or that seems unimportant to you — because all you want is to have the email translated, rather than expecting or demanding more. But whoever wants more result also has to give more information. And that info is — and here the circle closes — best worked out through good prompting, not by hammering it out on the keyboard.

But back to my little crusade in search of the holy prompt.

Because this new way of interacting with a machine is unfamiliar, and the vast majority haven’t yet learned to handle it properly, there are lots of aids. And I tried a lot of them, because I wanted extra-good results out of the AI, knowing — or at least suspecting — what it can do.

The whole crop of prompt generators I find rather poor and cumbersome. You have to fill in pseudo-forms that, more or less thoughtfully, spit out a prompt that looks nice but never really fits and that I always have to re-tune.

The whole crop of guru master prompts or magic prompts — the fully honed, highly optimized ones that do exactly one thing really well — mostly do actually do it really well. But they’re all built differently and scattered across every conceivable channel and website, and they demand near-bookkeeping discipline if you don’t want to be forever hunting for them.

So that leaves good old handiwork. Prompt patterns and best-practice rulebooks. There are heaps of those, so: read up, think your way in, put it into practice. But in the long run that’s baloney too. You’ve got an AI that genuinely understands colloquial speech, and then you have to beat Freytag’s pyramid into your head like a kid in school and recite five acts in the right order. The template doesn’t make the drama good on its own. Structure is good and important, but it also kills a little of the fun and the soul. But complaining doesn’t help. Structure demonstrably helps. The results are worlds better with clear, well-thought-out prompts. More reproducible. More reusable. More extensible. More shareable.

Through cherry-picking I pulled out of all those patterns the elements they pretty much all had in common: role description, context information, special conditions and constraints, task description, and the desired output format. Then gave it an international name and packed it into a catchy acronym format:

Assign-Inform-Modify-Task-Output (AIM-TO for short)

That knocks out 80% of my use cases. And when it needs to be a bit more, I can top it up as needed with a few more building blocks like Priority, Ask back, Chain of thought or Example (PACE for short).

I had my own prompt pattern. Nice. You gain the bulk of the added value anyway just through the act of structuring itself. It’s no masterpiece, but solid craft. The world-class acronym AIM-TO-PACE, which I made my trademark, actually only came along much later, chronologically speaking.

The problem: prompting by a pattern — any pattern — works well, but it’s soooooo booooooring. And mindless. You have to force your thoughts into a shape. With long, convoluted thoughts that’s genuinely laborious, even. And it’s slow. Before I’m done I’ve usually already got new ideas that this anti-brain-dump setup won’t let me convey properly.

If only we had something to take the boring, mindless tasks off our hands. A number-crunching drudge. Ideally one that also understands my original jumble of thoughts.

Sure — I’ll let my house AI do that. The first 5–6 iterations flew by, the first three full versions were all usable and each a little better than the one before.

The meta-prompt: the prompt that generates prompts was born. My meta-prompt.

And since version No. 3 I’ve never again pressed a long train of thought into a prompt pattern, or written or read a long prompt in full. Gone are the days when I would flatteringly prompt an AI just to please it, playing to what it wanted to hear. That, I let it handle.

But I wanted more. Naturally the meta-prompt itself had to — out of pure hubris — obey the AIM-TO pattern. And it was supposed to optimize the output format in the target prompt. Nicely thought around two corners. From the request, the generated prompt should automatically detect whether the user wants dialog mode or just a direct result. Ask for the most important parameters, but always also deliver a copy-ready output based on well-founded, mirrored-back assumptions. And it was supposed to have many more features besides. I have a great many ideas. But a prompt is not a computer program that you can — or should — extend at will.

The meta-prompt grew more powerful and bigger, but in the process also more fragile and no longer sensibly extensible. Every optimization broke something else again. Side effects everywhere. After a 2-hour improvement session going in circles, Gemini reckoned it wasn’t getting anywhere. So I asked Claude instead. He didn’t know the meta-prompt yet and could optimize away with fresh eyes. Which he did. But here it only took a few iterations before he, too, said it was a dead end.

Of course it hurts when you’ve put time and energy in and then it simply won’t budge. It feels wasted. But of course it isn’t. I learned so much over those months. Understood, too, why it had been tilting at windmills from the very start. You can’t “program your way through” a long prompt with the old mindset. There are technically insurmountable problems.

But the principle is and remains good — a prompt that generates quality prompts. Always and everywhere. Just text. No tool. No dozens of sources to save or memorize. Something that works on every common LLM. No vendor lock-in that chains me to one LLM or that only works with the best AI version.

Too good to simply quit. Good enough to start over completely. So: radically start from 0, with tight, simple goals and anti-goals.

• The meta-prompt itself doesn’t have to be AIM-TO — the hardest step for me, but also the most effective
• It has to generate AIM-TO:
  • Stable
  • Robust
  • User-friendly, but without the bells and whistles

That was finished in one session. Just the way it is in its present form. That’s how I distribute it. And it works — for me, for friends, acquaintances, and colleagues. The barrier to entry is and remains there, but it’s low. The skepticism is there too. But whoever clears that first hurdle often gets hooked. And the AIM-TO-PACE framework — pattern and meta-prompt combined — actually has more advantages than you’d see at first glance. The pocket-sized prompt generator. A by-the-book route to a structured prompt. That’s boring too, sure, but less effort. And best of all: I don’t have to tie my virtual tongue in a knot. After a little preparation, I can prompt away with no compromise on quality.

The thoughts are free again.

Wait — I’m sure I had the perfect prompt somewhere in a chat history. Or did I save it locally? Or email it to myself? Ah, no, I saw it on YouTube, that one’s really good. Let me just dig it out.

“Twooo hours later…” SpongeBob French Narrator voice

When the search takes longer than building the prompt … Those days are thankfully over. And ever since, “verschlimmbessern” has been my favorite untranslatable word.