Why Prompt Engineering Was the Wrong Abstraction
Two years of advice about wording prompts better missed the actual lever: what material the model has to reason from.
The claim
For a couple of years, the dominant advice for getting good results out of AI was about wording: phrase the instruction better, assign the model a role, give it a few examples, ask it to think step by step. Some of that genuinely helps. None of it was ever going to be the discipline that made AI reliable inside a real business, because it was solving the wrong problem.
What prompt engineering assumes
The premise underneath most prompting advice is that the model already has what it needs somewhere inside it, and the job is just to ask correctly: find the phrasing that unlocks the right answer. That premise holds for a certain kind of task: self-contained, general-knowledge, the kind of thing a competent stranger could do with no background on your business. Summarise this. Rewrite that. Explain a concept. Phrasing genuinely moves the needle there.
Where it breaks
Most business tasks aren't that. They depend on facts a general-purpose model has no way to know: what already happened with this specific customer, what this company's policy actually says once you account for the exceptions, what "acceptable risk" means for this particular deal given its history. No amount of clever wording supplies information that was never in the prompt to begin with. You can ask a model to "be thorough" as many ways as you like; it still won't know about the conversation that happened last Tuesday unless someone tells it.
What the demos actually show
The three interactive demos on this site's Work page hold the instruction constant on purpose ("reply to the customer," "write the investigation brief," "write the clause review") and vary only the material behind it, across three levels of curation. The quality gap between the least and most curated version in each demo is enormous, and none of it comes from wording. The instruction never changes. What changes is what the model has to reason from.
The actual skill
Once phrasing is taken off the table as the lever, what's left is judgement: deciding what's relevant to a specific situation, what's already been resolved and shouldn't be re-litigated, what to explicitly exclude, and which vague instruction ("be professional," "be careful") needs to become a concrete, checkable rule instead. That's a different kind of expertise than writing a good prompt. It's closer to editing, or to the judgement a good manager exercises when briefing someone new on a case: knowing what actually matters here, not just how to phrase a request.
Where this leaves prompting
None of this makes prompt phrasing useless. It's a real, small skill, worth having. It was just never going to scale to reliable, repeatable AI use inside a business, because the thing that was actually failing wasn't the wording. It was what the model had in front of it. That's a curation problem, not a phrasing problem, and it needed its own name.
Frontmatter — content/thinking/why-prompt-engineering-was-the-wrong-abstraction.md
title: Why Prompt Engineering Was the Wrong Abstraction slug: why-prompt-engineering-was-the-wrong-abstraction summary: 'Two years of advice about wording prompts better missed the actual lever: what material the model has to reason from.' stat: Same instruction stat_label: three curation levels, three completely different outputs
If this way of thinking is relevant to a problem you're facing, I'd be glad to talk it through.
Start a conversation