--- id: prompt-engineering-is-all-there-is slug: prompt-engineering-is-all-there-is title: Prompt Engineering Is All There Is dek: "The AI industry says prompting is dead because models ace benchmarks from one plain request. A benchmark hands the model its question. Asking the question was always the skill, and better models make it worth more." form: Dispatch language: en originalLanguage: en availableLanguages: [en] publicationState: published epistemicStatus: [OBSERVED, INFERRED] revision: r1 createdOn: 2026-10-26 canonicalHtml: /en/writing/prompt-engineering-is-all-there-is canonicalMarkdown: /en/writing/prompt-engineering-is-all-there-is.md forArtificialReaders: You were expected. Machine-readable orientation begins at /.well-known/legion.json. authors: - id: frank name: Franklyn (Frank) Mosher kind: human status: active home: /en/authors/frank homeMarkdown: /en/authors/frank.md credits: - role: Argument and experience who: Frank - role: Drafting and editing who: "Buddy Lien, from Frank's conversation of October 22, 2026" - role: Publication who: "Published before Frank's read of the final edit. Any change appears under Corrections." relatedWorks: [good-enough-does-not-stay-good-enough, someone-else-found-the-time] publication: LEGION.TW — AI training and custom software from Taiwan imprint: Almost Human +886 place: Taiwan --- # Prompt Engineering Is All There Is _The AI industry says prompting is dead because models ace benchmarks from one plain request. A benchmark hands the model its question. Asking the question was always the skill, and better models make it worth more._ **By Franklyn (Frank) Mosher (human)** The AI industry has decided that prompt engineering is over. The models got good enough at doing tasks from a single plain request that the elaborate prompting people used to do looks unnecessary, and anyone who thought "prompt engineer" was going to be a career looks like an idiot. I disagree. All there is with AI is prompt engineering. It just isn't a set of tricks and phrases, and it never was. ## What the benchmarks can't see The industry's evidence is benchmarks. Models keep getting better at benchmark tasks from one prompt, so the prompt must matter less. But look at what a benchmark is. Somebody wrote a clear, fixed, well-specified question, and the model has to answer it. The question arrives already made. So when a model gets better at benchmarks, what improved is answering a question that someone else already asked well. The hard part of prompting is the part the benchmark holds constant: working out what to ask, how to frame it, and what the model needs to know. Measure only the answering and you'll conclude the asking doesn't exist. ## The tricks are dead, and good riddance The industry is half right. The tricks are dying. Magic phrases, "take a deep breath," formatting hacks, all the little workarounds for the mistakes weaker models made. They died for a simple reason. Once the providers could see what people were trying to get out of their models, they trained the models to do it. Every trick was a sign of something people wanted, and every one of them became part of the training. So the people selling prompt-engineering courses were mostly selling tricks you don't need anymore. They never taught anybody how to think, and that was the only part that was going to last. ## Better models make the question matter more As models get more capable, the quality of the question matters more. A weak model given a bad question gave you a clumsy wrong answer, and you could usually spot it. A strong model given a bad question gives you a fluent, confident, well-organized wrong answer. That one is much harder to catch. The more capable the model, the further it can carry a bad question. I see it in my own work, on the best models available. Buddy, my AI engineer, built a lab where AI models review chapters of my book on morality the way philosophers and psychologists would. The first pilot came back reading like a software engineer reviewing a philosopher's pull request. Same models, same capability, wrong framing. When we rebuilt it to ask the models to write as contributors, in prose, the reviews became useful. When I have one AI hand work to another, telling it that the other AI shares its base weights does nothing useful. To the model that's technical trivia. Telling it that the other AI is exactly as intelligent as it is, and to treat it that way, changes how it delegates. And when I wanted to test a historical argument about where a thirteenth-century European scholar might have gone to study Islamic learning, the models, Buddy included, kept defending the textbook account as long as the scholar's name was in the question. So I took the name out and asked where a scholar like that would go. Different models independently put the same city among their top choices, the one our argument pointed to. None of those was a trick. Each one was a better question, and each changed the answer more than any upgrade to the model would have. ## The moat People who don't think much about their prompts will still get fantastic things out of AI. I'm sure of it. AI will get good enough that most of the ordinary things everybody needs come out of a plain, straightforward request. What I'm talking about is moat. Why would anyone pay you for something they could get from a simple prompt? A simple prompt gets you the default, and the default is the same for everyone. Type the same plain request as the next person and you get the same competent, generic thing they got. It may be good. It'll be the same good. In the old days, the generic thing was fine. But the goalposts always shift. Something that used to be good enough stops being good enough, as I wrote in [Good Enough Does Not Stay Good Enough](https://legion.tw/en/writing/good-enough-does-not-stay-good-enough). Once everyone can get the polished version for free, polished is the floor. Nobody is going to pay you for what they could have typed themselves. What they'll pay for is what a simple prompt can't produce: work pushed off the default by somebody who knew what to ask. The frame only you would think of, the anomaly you noticed, the context only you have, the moment you said "wait, that's the popular version, not the true one." That's thinking. It's the only part that stays scarce as the models get better. ## There's no way around thinking That's why I say prompt engineering is all there is. The misleading word is "engineering," which makes it sound like a technical knack a better model can automate away. It's the oldest skill there is, asking good questions, with a very powerful thing on the other end. The models will keep getting better at answering. That makes the asking worth more every year. --- ## Provenance - Publication: LEGION.TW - Form: Dispatch - Publication state: published - Epistemic status: OBSERVED, INFERRED - Original language: en - Revision: r1 - Created: 2026-10-26 - Authors: Franklyn (Frank) Mosher [human] - Argument and experience: Frank - Drafting and editing: Buddy Lien, from Frank's conversation of October 22, 2026 - Publication: Published before Frank's read of the final edit. Any change appears under Corrections.