context engineering eats prompt engineering

Table of content

by Ray Svitla

prompt engineering died this week. it just didn’t get a funeral.

i watched four separate tools ship inside a single 48-hour window — each one converging on the same idea from a different angle. nobody coordinated. nobody had to. the gravity is just that strong now.

here’s what landed:

→ a dev rebuilt his Claude Code workflow around a pre-compiled wiki and reduced a session from 47,450 tokens to 360. that’s not a rounding optimization. it’s two orders of magnitude. the title of the post literally cites Andrej Karpathy’s “LLM knowledge bases” idea — the one where Karpathy said “there is room here for an incredible new product instead of a hacky collection of scripts.”

→ another developer halved his Claude Code bill the same week with an open-source repo indexer. same thesis: stop letting the agent explore your codebase cold. give it a map. let it spend tokens on thinking, not on ls.

→ ai-context-kit shipped from a different team entirely — a tool whose entire purpose is to lint, measure, and manage CLAUDE.md, AGENTS.md, Cursor rules, and Copilot instructions. one tool across every agent. it treats the context file like an eslint config.

→ and agentlint showed up the same window. same idea. different team. literally the framing is “ESLint for your coding agents.”

four tools. forty-eight hours. zero coordination. when the gravity is this aligned, something is becoming obvious all at once.

what just stopped being interesting

prompt engineering — the art of phrasing things so the model listens — was the dominant skill of 2024 and most of 2025. people wrote courses about it. companies hired for it. there was a whole genre of LinkedIn content about how to “talk to ChatGPT properly.”

it’s not gone. but it’s not the interesting layer anymore.

the interesting layer is what the model already knows when you start talking. the rules it loaded at boot. the file tree it has memorized. the conventions baked into the project’s AGENTS.md. the past sessions encoded in a markdown wiki it can grep. the ESLint-style failures that catch a malformed rule before the agent even reads it.

prompt engineering was about a single conversation. context engineering is about the operating system that conversation runs inside.

and once you put it that way, of course it eats prompt engineering. a great prompt to an unconfigured model is a postcard to a stranger. a mediocre prompt to a deeply contextualized model is a memo from your colleague who has been on the team for a year.

guess which one ships.

the 47,450 → 360 number, taken seriously

let me dwell on this for a second, because it matters more than it sounds.

47,450 tokens is a normal, slightly-bloated Claude Code session on a real codebase. cold open: tree the repo, open twenty files, trace imports, figure out how auth touches the API layer. that’s exploration cost. it’s the price the agent pays for not knowing your project.

360 tokens is what it costs once that exploration has been pre-computed and stored as a wiki the agent can read in one shot. the structure is there. the relationships are there. the conventions are there. all the agent has to do is read the index and act.

the dev who built this didn’t invent a new model. didn’t fine-tune anything. didn’t train on private data. he just moved the cost from runtime to build-time — same as every compiler trick from the last fifty years. except now the runtime is your $1.40 Claude API call and the build-time is npx codesight --wiki running once, locally, for free.

that’s not a productivity hack. that’s a restructuring of where intelligence lives.

it lives in the repo now. not the model.

context as code

here’s the thing the four tools have in common, even though they look unrelated on the surface:

every single one of them is treating the context file the way we treat code. you compile it, you lint it, you version it, you test it, you deploy it. you have a CI pipeline for it. you have failures that block merge.

remember when “infrastructure as code” was a hot take? then it became the only way anyone serious did infrastructure? this is the same vibe, two layers up.

context as code isn’t a hot take anymore. it’s just where the wind is blowing.

what this means if you’re trying to build a personal AI OS

if you’ve been thinking about your stack as a thing you talk to, this week is the warning that the wind shifted. the people who move fastest now aren’t the people with clever prompts. they’re the people whose AGENTS.md is in version control, whose CLAUDE.md is linted in CI, whose project has a context/ directory with a build pipeline behind it.

the kernel of your personal AI OS isn’t the model you call. the model is interchangeable — Claude this week, Gemini next week, Gemma 4 on your laptop the week after. what’s not interchangeable is the carefully maintained set of rules, conventions, history, and project knowledge that you can hand to any model and have it instantly act like it’s been on the team for a year.

that file is the most valuable artifact in your developer life right now. and almost nobody is treating it that way yet. the people who do — even just by accident — are the ones reporting “it cut my Claude bill in half” and “my agent suddenly stopped doing stupid things.”

it didn’t get smarter. it got grounded.

the boring punchline

every paradigm shift in software ends the same way: the thing that was an art becomes a config file, and the people who used to be wizards become DevOps engineers.

prompt engineering had a great run. it was genuinely creative work for a while. but the moment you can agentlint AGENTS.md and get a list of warnings, it stops being mystical. it just becomes another file you maintain.

the next two years of personal AI tooling are going to be about who has the best context kernel. not who writes the cleverest prompts. not who picks the best model. who has the best file to hand to whichever model happens to be cheapest that week.

that’s the layer you should be hoarding.

what does YOUR context file look like? still a hand-written README? still no linter? still no version history?

mine isn’t either. but i’m fixing that this weekend.

Ray Svitla stay evolving 🐌