AI dropped the cost of building software to a couple of sessions, and the old build vs buy question quietly broke. The story of the helpdesk I nearly built — and the four lines of code that turned out to be the answer.
There are two kinds of AI memory. Remembering your business between sessions is the one everyone works on. The system knowing what it has actually built is the one almost nobody is solving — and the gap is an entire missing layer.
Operators on X keep saying they're four turns into an AI session and the context window's full. The replies miss the actual problem — they've been taught the wrong physics.
Context engineering is the new buzzword. The skill underneath is older and simpler: people burn their AI allowance in two turns and blame the model, when the model is fine — they're paying for synthesis they could have externalised. That's context discipline, and it's the entire skill.
Government departments are funded to collect data, not to present it. The gap between what the state holds and what an ordinary person can use is structural, durable, and almost entirely unclaimed.
Suppliers with APIs get wired into the pipeline. Suppliers still emailing CSV files get checked when someone remembers. Not preference. Physics.
Prompt engineering is the wrong name. Context engineering — what I've been calling ingeniculture for a year — is the practice of building the room the LLM stands in.
Most AI workflows are prompts applied to empty rooms. This is what an operating system looks like instead — document tiers, named characters, a wiki the model can read, and a boot sequence that loads context before the first prompt arrives.
The AI reflects whatever substrate it meets. The dangerous case isn't where it refuses to answer. It's where it answers fluently and nobody in the room can tell it's wrong.
Every profession that charges for opaque expertise is about to discover the trail was always there. AI doesn't invent evidence — it surfaces what's already readable. The age of 'we did the work, trust us' is ending.