Prompt engineering, the process of designing good prompts for generative AI models, is dead, say some. Long gone are the days when you needed to prefix your prompt with ‘Act as an expert…’, provide few-shot examples or spell out each step.

Lately, I’ve spent more time writing prompts – less time in the terminal with Claude Code, more time with Claude desktop. While I used to vibe code systems for automating work, I often find it more efficient prompting the chatbot directly and having it orchestrate the workflow itself. Chances are it will deliver a folder with supplementary code anyway – the code I would’ve vibe coded, only better.

For example, I recently built a pipeline for generating interesting mathematical research questions and answering them, mostly as a way to stay up to date with AI. I wanted to see for myself what a basic scaffolding could accomplish. I supplied my favourite papers within my area, much as a seed, and made agents formulate research questions, coming up with solutions and typeset the findings. At each step, I made other agents adversarially screen the output, asking the agent to redo a step if it wasn’t up to standards. More specifically, I had markdown files with rubrics for research taste and correctness. With this scaffolding, an everyday model like Sonnet could produce decent papers. However, the whole system was too token-intensive¹, and running the pipeline end to end took a few hours².

Instead, I had an AI produce a prompt to make GPT 5.6 Sol run the full pipeline at inference time, articulating questions, solving problems and presenting solutions, with quality checks at every step. Bitter lesson: it works. Giving a chatbot a carefully written prompt proved more time- and cost-effective.

The success of Sol shouldn’t have come as a surprise – some AI labs are quite explicit about their goal to automate intellectual work. The first paragraph in OpenAI’s charter reads:

OpenAI’s mission is to ensure that artificial general intelligence (AGI)—by which we mean highly autonomous systems that outperform humans at most economically valuable work—benefits all of humanity. We will attempt to directly build safe and beneficial AGI, but will also consider our mission fulfilled if our work aids others to achieve this outcome.

Until now, I hadn’t fully internalised this was their goal. Really? I’ll interpret a bit, if I may: automating intellectual work means reducing some of the most rewarding aspects of my work to writing one prompt and having the model one-shot the task in the desktop app.

The frontier models accessible through desktop apps are already very good, integrating seamlessly with many tools and services, and it’s safe to assume they’ll only get better. Even if AI progress stalls, we could, in theory, replace much economically valuable work with agents supplied with the right prompts.

Among other things, this made me question the utility (and viability) of many AI scaffolding products. There are plenty of specialised AI tools for generating and reviewing written work, whether legal reports, research articles or op-eds. Bitter question³: if I generated a prompt asking a frontier model in thinking mode to execute exactly the same steps as my elaborate system, would I achieve comparable results? Unless the system draws on closed data sources, I think the answer is yes. Moreover, it’s likely cheaper and faster – some of the world’s sharpest are racing to produce more efficient models within AI labs. There is a sense in which many AI startups are just million-dollar prompt factories⁴.

I’d expect most companies to rely heavily on AI well within 10-20 years. All companies might become AI companies, in some sense. Do we want to build an economy based on prompts?