The tell

There is now a substantial economy devoted to teaching people how to talk to software. Courses, newsletters, prompt libraries, templates, consultants, a genre of video whose entire premise is that you are holding the tool wrong.

Some of it is genuinely good work. All of it is evidence of the same thing: the product is unfinished, and the gap between what it can do and what it does for you has been handed to the user as homework. Every other category of software treats that gap as a defect to be closed. This one turned it into a curriculum.

We think prompt craft is a symptom, not a skill. The rest of this note is why we think the evidence supports that, and what we are building instead.

The thirty-nine point gap

The most useful study on this is uncomfortable for everyone, including us.

In 2025, METR ran a randomised controlled trial with sixteen experienced open-source developers working on repositories they had contributed to for years — large, mature codebases they knew intimately. Before starting, they forecast that AI tools would make them 24% faster. Afterwards, they reported believing they had been about 20% faster.

They were 19% slower.

Sit with the shape of that. Not the slowdown — that result is context-specific, and the researchers are careful to say so. The gap. Roughly thirty-nine points between what these developers felt and what the stopwatch recorded, in the population best equipped to notice the difference: expert practitioners, on their own code, measured directly.

If the people most likely to be right about their own productivity are that wrong, then “I’ve gotten good at prompting” is not a claim any of us can verify from the inside. Skill you cannot measure is indistinguishable from a habit you have grown comfortable with. And a product whose value depends on users acquiring an unmeasurable skill has quietly made itself unfalsifiable.

Which intelligence, not which leaderboard

The second piece of evidence is that the two most widely used assistants are used for almost opposite things.

As of June 2025, computer programming accounted for 4.2% of ChatGPT messages, and non-work usage had grown from 53% to more than 70% of all traffic between June 2024 and June 2025. Meanwhile, in Anthropic’s own index, tasks in the Computer and Mathematical category account for about 35% of Claude.ai conversations.

Neither number is a verdict on quality. Both are widely used, capable products. What the contrast shows is that “which AI is best” is a malformed question, because the honest answer depends on who is asking and what they are trying to do that afternoon. One tool’s centre of gravity is personal and practical; the other’s leans technical. A leaderboard cannot express that, and a person should not have to research it.

This is the same reason you never pick a model here. Matching a task to the right capability is a real technical problem with a real answer, and it is exactly the kind of problem software should absorb.

The day the prompt has to survive

The last piece is about when this work actually happens, and it is the one that makes prompt craft feel like a fantasy about a calmer person than any of us.

Microsoft’s Work Trend Index surveyed 31,000 knowledge workers across 31 markets. During core hours, they are interrupted every two minutes — roughly 275 times a day. They receive 117 emails and 153 chat messages daily. 57% of meetings are called ad hoc, with no invitation and no agenda. And editing activity on presentation decks rises 122% in the final ten minutes before a meeting starts.

That last figure is the whole argument in one data point. The moment of highest need is nine minutes before the meeting, in a document that is not ready, with two other conversations open. Nobody in that moment is composing a well-structured prompt with clear role definition, explicit constraints, and a worked example. They are going to type what they want in the words they have, and whatever the software does with that is the product.

Designing for the calm, well-briefed user is designing for a person who does not exist during working hours.

What the system should be doing instead

If a request needs decomposition into ordered steps, the system should decompose it. If it needs a plan before it needs prose, the system should plan. If the draft needs checking, the system should check it rather than the user being asked to notice.

That is what Aura and Echo OS are for. You state an intent — in the words you already have, at the moment you actually have the problem — and the decomposition, the routing, the sequencing, and the checking happen underneath. Not because the words do not matter, but because turning a rough intent into a well-formed unit of work is the software’s job. It is the most valuable thing the software does. Handing it to the user and calling the handoff empowerment gets the economics exactly backwards.

There is a version of this argument that flatters users, and we want to be careful not to make it. The point is not that people are incapable of learning to prompt. Plenty have, some are genuinely skilled at it, and that skill will keep paying off in expert tools built for experts. The point is that it should not be the price of entry for someone who wants to book a table, chase an invoice, or get a briefing before a meeting they did not know they had until eleven minutes ago.

The measure of this kind of product is not how much you can extract from it once you have learned its grammar. It is how little you had to learn.