The Imperial Rise of the Context Engineer
Bernard Lowe is the head of programming in my operating system — a named character with a defined brief: read the architecture as it exists, and surface what the builder is too close to see.
But those seven words didn’t land on a blank canvas. They landed on twelve months of infrastructure: a character definition that shaped how the review was conducted, a wiki page that defined what “audit” means in this specific context, a session protocol that loaded the right documents before the first word was processed, and a command reference that wired the trigger to its context automatically.
The prompt was seven words. The context engineering was everything else.
The Wrong Name
“Prompt engineer” focuses on the wrong unit of work.
The prompt engineering world is built on an assumption: the leverage is in the asking. Better phrasing, better structure, better chain-of-thought reasoning, and better system instructions. As if the right question asked to an empty room will produce a furnished answer.
The leverage isn’t in the asking. It’s in the room the question lands on — document tiers that manage what the LLM can see, named characters that shape the output from the inside, a wiki in plain markdown that the LLM reads natively, a correction loop that turns every mistake into permanent infrastructure, and a boot sequence that loads the full business context before the first prompt arrives.
The person who builds all of this isn’t engineering prompts. They’re engineering context. And the distinction matters — because the prompt is where the conversation starts, and it’s the last thing that matters.
The Magnifying Glass
The LLM is the sun, shining on everyone equally — same API, same capabilities, same training data. You put a question into a fresh conversation and you get diffuse warmth — competent, balanced, and fluent. The same warmth a million other subscribers get from the same LLM on the same day.
A context engineer builds the magnifying glass.
Same sun, same energy, but concentrated through a lens that focuses the light onto a precise point. The LLM doesn’t change. The concentration does. And that’s the distinction the industry keeps missing — everyone is optimising the sun when the leverage is in the glass.
Without the magnifying glass, the sun warms everything equally. With it, diffuse energy becomes focused enough to start a fire.
My seven-word prompt worked because the magnifying glass was already built. Bernard knew how to audit. The wiki defined what mattered. The tier system loaded the right context without loading everything. The correction history had already eliminated the mistakes I kept making over twelve months of daily use. By the time those seven words arrived, the infrastructure had done the work that prompt engineering pretends the prompt does alone.
The Ceiling Everyone Is About to Hit
Alex Rogov — a developer — put it in six words on X in March: “The bottleneck was never the model.”
He named it better than he knew. But the market is arriving at the same realisation now, from multiple directions at once, and each group is hitting the same ceiling without knowing the others exist.
Developers are connecting their note systems to Claude Code and Cursor, getting excited by the results, and about to discover that a growing pile of context creates the same ceiling I hit when I started. Boris Cherny’s thread about his CLAUDE.md (the instruction file that tells the AI how to behave) went viral — over a million views. Two replies were prophetic: “Gonna have 5,000 lines of ‘don’t do this’ in a week” and “betting he has a graduation/weighting system.” They can see the ceiling from the foyer.
Shopify’s CEO used the phrase “context engineering” in a company-wide memo — the way ideas arrive in corporate life: correctly, approximately, and at scale. The term is arriving, and the question is who defines what it means.
Discipline Scales Linearly
The distinction between a good operator and a context engineer is the distinction between discipline and infrastructure.
Garry Tan said the best AI coding happens in the morning when you’re fresh from dreaming about latent space. Dreaming about latent space is not a system. It’s a morning habit with a good name. He’s right about the bottleneck — what’s in your head matters. But he’s carrying the architecture in his head and re-articulating it each morning, while I boot the same architecture in three seconds. His approach works until he has a bad night’s sleep. Mine works regardless.
Morning clarity is discipline — getting up early, thinking hard about your project rules, and loading the mental model fresh. It produces better output, but it depends on the person maintaining it. Miss a night’s sleep, have a difficult week, lose focus for an afternoon — and the discipline fails because the discipline lives in you.
I’ve built the same discipline into infrastructure. The document tiers load whether I’m sharp or exhausted. The characters shape the voice whether I’ve had coffee or not. The boot sequence fires the same way every session. The correction loop holds regardless of whether I remember the specific correction — because I don’t have to remember it. It’s been built into the system permanently.
Discipline scales linearly and breaks when you’re tired. Infrastructure scales permanently and doesn’t care what time it is.
The Gap That Can’t Be Templated
Every “AI workflow” product on the market is selling prompt engineering — templates, chains, agents with pre-built instructions. The packaging improves every month; the agents get names, the templates get case studies, the chains get diagrams with arrows. The underlying problem doesn’t change: the LLM is diffuse without a lens, and the lens requires domain expertise to build.
You can’t template context engineering because context engineering requires having done the work: twenty-six years of knowing which keywords actually matter for a label printer in Norfolk, a correction history that taught the system British English not American, a character I tried and discarded because the LLM gave me television not craft — and the one I found instead that understood the difference — and wiki pages I wrote at ten on a Tuesday night because a real client had a real problem and the methodology evolved to handle it.
A repo with fifty thousand stars can give you the agents, the hooks, and the skills, but it can’t give you the magnifying glass. The magnifying glass is ground from your expertise, your corrections, your clients, and your decisions — that’s not a product limitation, it’s the nature of the thing.
The gap between fluent output and useful output is a context engineer. And the gap is widening — because the LLMs are getting better at fluency every quarter while the context side stays largely unbuilt. Except by the people doing it deliberately.
Ingeniculture
I named it earlier this year: ingeniculture — providing the infrastructure for the intelligence to thrive. The wiki, the tiers, the characters, the correction loops, and the boot sequence — all of it existed before the name did, and the name was just the word that described what had already been built.
The market is arriving at the same place from different directions — some are calling it context engineering, and the name doesn’t matter. What matters is the recognition that the LLM is a commodity — and the differentiation has moved permanently upstream, into the infrastructure the LLM stands on.
A prompt engineer writes “you are an expert copywriter.” A context engineer builds a wiki, a voice guide, a correction history, a cast of characters, and a session protocol — and then types seven words and gets an engineering audit.
The bottleneck was never the LLM.It was always the room it was standing in.
Related: Ingeniculture: Why the Room Matters More Than the Model is the practice this piece names the practitioner of. Who Is Holding the Mirror? explains why infrastructure built by the wrong person produces the wrong reflection — context engineering only works when the substrate is real. The Correction Loop shows how the infrastructure compounds, one fix at a time. The Magic Trick explains installation versus instruction — the mechanism a context engineer uses instead of telling the model what to do. Fluent Nonsense names the problem context engineering solves.