The Prompt Was Never the Bottleneck.
Two people. Same model. Same task.
Marc opens Claude and types: "Write a proposal for a healthcare client, budget ~$50K, they want AI automation."
Sofia opens Claude. But before she types anything, she drops in a folder. Inside: her brand voice guide, her three best past proposals, her pricing tiers as a table, her positioning doc, her qualification framework, the client's website scraped into markdown, and a one-page brief from the discovery call.
Then she types the same prompt.
Marc gets something generic enough to apply to any company on earth. He'll spend the next 90 minutes rewriting it. Sofia gets something she sends with light edits.
Same tool. Same day. Same $20 subscription.
One of them is 4x more productive than the other, and it has nothing to do with their prompting skills.
It's the context.
The word nobody was using 12 months ago
If you were scrolling X in late June 2025, you watched a term go from zero to canon in about a week.
June 19. Tobi Lutke, CEO of Shopify, posted that he preferred "context engineering" over "prompt engineering" because it better described the actual skill: "the art of providing all the context for the task to be plausibly solvable by the LLM."
Four days later, Harrison Chase (co-founder of LangChain) dropped the blog post that became the canonical reference. He called context engineering "the most important skill an AI engineer can develop."
Two days after that, Andrej Karpathy — co-founder of OpenAI, the guy who coined "vibe coding" — blessed it: "Context engineering is the delicate art and science of filling the context window with just the right information for the next step."
Then Simon Willison, one of the most respected developer voices on the internet, admitted he'd been defending "prompt engineering" — but the term had a "poor inferred definition." People dismissed it as laughably pretentious jargon for chatbot typing. He conceded: "context engineering" describes what practitioners actually do far more honestly.
Then Gartner, in July, put a stake in the ground: "Context engineering is in, and prompt engineering is out."
By September, Anthropic published the definitive engineering guide. By November, MIT Technology Review framed 2025's defining arc as "From Vibe Coding to Context Engineering." By February, Harvard Business Review ran a cover piece on context as competitive advantage.
That was eight months ago. The term now has nine academic papers, a Gartner briefing, and a World Economic Forum report.
And 90% of the people I talk to still think AI is about writing better prompts.
OK but what is it, in plain English?
Skip the jargon.
When you type a prompt into ChatGPT or Claude, that prompt is maybe 5% of what the model actually "sees" when it answers. The rest is context — the conversation so far, any attached documents, any system instructions, any tool outputs, any retrieved knowledge.
Prompt engineering asks: "How should I phrase this question?"
Context engineering asks: "What does the model need access to right now — in what order, in what format, with what authority?"
Prompt engineering was a 2023 skill. You got clever about wording. You learned to say "step by step" and "you are an expert in X" and "return JSON."
Context engineering is a 2026 skill. You build a system around the model. You decide what gets loaded when. You version it. You maintain it. You treat the context window like a precious, leaky workspace that needs a janitor, not a poet.
Philipp Schmid broke it into seven layers. I find the breakdown genuinely useful:
- System prompt (who the AI is, what the rules are)
- User prompt (the actual ask)
- Short-term memory (what just happened in this conversation)
- Long-term memory (what happened across conversations)
- Retrieved information (the right document pulled at the right moment — RAG)
- Available tools (what the AI can do, not just say)
- Structured output (how the answer is formatted)
Notice something? The user prompt is one slot out of seven.
Prompt engineers have been optimizing 5% of the equation. Context engineers build the other 95%.
And the infrastructure for those layers is growing fast. Anthropic's Model Context Protocol (MCP) — an open standard that lets AI agents connect to external data and tools — hit 10,000 active public servers and 97 million monthly SDK downloads by March 2026. OpenAI adopted it. Microsoft integrated it into Copilot. AWS added support.
The pipes are being built. The question is whether you're putting anything valuable through them.
The data is, frankly, violent
A paper dropped this month by Calboreanu et al. that I think every AI "expert" on LinkedIn should be forced to read before they post again.
200 real human-AI interactions. Measured first-pass acceptance rates by context quality:
- Minimal context (under 500 tokens): 21% first-pass acceptance. Average 3.4 iterations before the output was usable.
- Moderate context (500–2,000 tokens): 72% first-pass acceptance. 2.1 iterations.
- Comprehensive context (over 2,000 tokens): 89% first-pass acceptance. 1.3 iterations.
From 21 to 89. Same humans. Same models. Only variable: how much relevant context was in the window.
But the finding that actually broke my brain was this one.
The same study measured "authority context" — how the instructions are delivered:
- Documented in a file: 89%
- Told verbally (or in a chat turn, not a persistent doc): 64%
- Absent entirely: 29%
A 60-point quality gap between "nothing" and "a file."
Read that again. The difference between your AI output being garbage and being production-ready is whether or not you wrote a document.
Not a better prompt. A document.
This is the piece the prompt engineering industry missed. They were optimizing a 14-word string. Meanwhile, the teams that wrote a context file were winning by 60 percentage points.
And incomplete context was associated with 72% of all iteration cycles. Those frustrating loops where you keep going back and forth — "no, not like that," "closer but wrong," "you're not understanding" — 72% of the time, the problem wasn't the AI being stupid. It was the AI being brilliant with nothing to work with.
Like handing a surgeon a scalpel and not telling them which patient is on the table.
Receipts from the real world
In case you think this is theory.
LinkedIn built something called CAPT — Contextual Agent Playbooks & Tools — on top of Anthropic's Model Context Protocol. They gave their AI coding agents organizational context across thousands of microservices.
Results: 70% reduction in issue triage time. 3x faster data analysis. Debugging time cut by more than half. Over 1,000 engineers using it internally.
Five Sigma (insurance): 80% reduction in claim processing errors. 25% increase in adjustor productivity. 95%+ accuracy after deployment feedback cycles.
OpenAI ran a five-month internal experiment where a small team shipped a beta product with roughly a million lines of code. The key lesson they published? "Give Codex a map, not a 1,000-page instruction manual."
Nobody is getting these numbers from a better prompt. They're getting them from a system that feeds the model the right context at the right moment.
"Can I just have the AI write the context for me?"
I knew you were going to ask this. Everybody asks this.
The answer, as of 2026, is: no, and the attempt actively makes things worse.
Gloaguen et al. ran the experiment on 138 real-world code repositories. They had LLMs generate the context files — the AGENTS.md, CLAUDE.md style docs — for each repo. Then they measured agent task success with AI-generated context versus human-authored context.
Result: AI-generated context files reduced agent task success rates. And increased inference costs by over 20%.
Let me translate that.
You can't automate your way out of context work. The thing the machine needs from you is the one thing the machine can't produce for you.
Your judgment about what matters. Your voice. Your edge cases. The hard-won scar tissue of "we tried this in 2023, it failed, here's why."
This is the taste gap from Letter 70, showing up with teeth. The Gloaguen finding is the empirical proof. Context about you can only be written by you.
You cannot automate the context that makes automation work.
The moat nobody can copy
In February, Harvard Business Review ran a piece titled "When Every Company Can Use the Same AI Models, Context Becomes a Competitive Advantage."
Their argument is something I've been saying in this newsletter for a year:
The model is a commodity. OpenAI, Anthropic, Google, Meta — within 90 days of each other, they all release roughly equivalent capability. Your competitor has the same GPT-5. Your competitor has the same Claude.
What your competitor does not have:
- Your 11 years of customer support transcripts, tagged by resolution type
- The way your team writes a discovery document
- The 14 edge cases your pricing model handles
- Your post-mortem on the client that went sideways in Q3 2024
- The exact phrasing your founder uses when saying no to scope creep
That is your context. And it's the only thing the AI can't download from Hugging Face.
The World Economic Forum called this the "work graph" — an anonymized map of how your high performers actually execute. How they think through a problem. What they check first. What they'd never do. Which shortcuts work and which ones burn you three months later.
And Adam Monago, writing about Knowledge Management, delivered what might be my favorite line of the year:
"Generation is now free. Curation is now everything."
Everyone can generate. The models are the same. The intelligence is the same. The cost is nearly the same. The only variable left is what you feed the machine.
The Knowledge Management people saw this coming 30 years ago. Nobody listened. Now it's the most important skill in AI and we're calling it "context engineering" because we're embarrassed to admit KM was right the whole time.
What this actually looks like (the CM OS confession)
I've been doing this for 18 months without calling it context engineering.
In Letters 74 and 75, I walked you through my personal setup — what I called CM OS. A folder. Text files. An AI agent sitting on top of it all. In Letter 78, I told you about a team that wrote a 22-page process document and watched their AI go from chaos to near-zero errors. Same agents. Same models. Different context.
At the time, I thought I was just being organized. Turns out I was building a context engineering system before the term existed.
Here's the rough shape of it:
Layer 1 — The Identity Layer. One markdown file. Who I am, what I do, what I believe, how I write, what I refuse to write. Updated every 6–8 weeks. About 1,200 tokens.
Layer 2 — The Current State Layer. This quarter's focus, this month's priorities, this week's commitments. Short, dated, ruthlessly pruned. Anything older than 90 days either graduates to long-term memory or gets deleted.
Layer 3 — The Knowledge Layer. Past newsletters (indexed), client patterns, pricing logic, the two dozen frameworks I use most. Not loaded by default — retrieved when relevant.
Layer 4 — The Task Layer. For each specific job, a brief: what this is, who it's for, what success looks like, what "done" looks like.
Layer 5 — The Guardrail Layer. What I don't do. What I don't say. Topics I decline. Mistakes I always catch.
When I start a serious session with Claude, layers 1, 2, and 5 are always loaded. Layers 3 and 4 are pulled on demand.
The whole thing took me about three weekends to set up and I maintain it for maybe 30 minutes a week. In return, every single AI interaction starts from a baseline that matches who I am — not a generic "helpful assistant" baseline.
And the industry is converging on this exact pattern. Claude Code has CLAUDE.md. Cursor has .cursorrules. OpenAI Codex has AGENTS.md. Google's Jules uses the same format. The emerging standard, literally called AGENTS.md, describes itself as "a README for agents."
It's all text files. It's all context. The document is the AI OS layer. I said it in Letter 78. Turns out the whole industry agrees.
The trap: more context ≠ better context
Now for the part that gets conveniently ignored by people who think bigger context windows magically solve everything.
They do not.
Chroma Research tested 18 LLMs and found something called context rot. Performance follows a U-shaped curve: models are sharp at the beginning of the context window, sharp at the end, and 30% less accurate in the middle — a "dumb zone" where recall degrades and reasoning falters.
This is the "Lost in the Middle" problem (Liu et al.). It's well-documented. Stuffing more context in doesn't help. It can actively make the AI stupider.
And the Calboreanu study found that context has a 47-day half-life. After 47 days, enough has changed in your world that half your context is degrading output quality.
So no — the answer is not "dump the whole company wiki into Claude and let destiny decide." That is not engineering. That is hoarding :)
Anthropic's own guidance is blunt: "Find the smallest set of high-signal tokens that maximize the likelihood of some desired outcome."
Not the most tokens but the RIGHT tokens. Not more hay. Better needles.
Martin Fowler, at Thoughtworks, takes it further. He argues context engineering is one component of something bigger — "harness engineering": the entire system you build around the model, including architectural constraints and entropy management. Agent = Model + Harness. The model is commodity. The harness is the craft.
"So I'll just write everything down, got it."
Slow down.
This is where most people hit the wall I hit in month two. And the wall has a name:
You can't write down what you don't know you know.
The decision your best ops person makes in 0.4 seconds when a vendor quote looks weird — that's context you've never written down. The reason your company always adds a 15% buffer to timeline estimates — context nobody wrote down. The three questions your best salesperson asks that unlock the real objection — context nobody wrote down.
This is what Knowledge Management calls tacit knowledge, and capturing it is genuinely hard work. It requires sitting with your best operators and asking "why did you do that?" until they get annoyed, and then keeping on asking.
The organizations winning with AI right now are the ones doing this UNSEXY work. Writing down what they know. Versioning it. Reviewing it monthly. Treating their context files as infrastructure, not documentation.
The ones losing are still hoping a better prompt will save them.
The AI OS principle
Let me tie this to the central thesis of this newsletter, because context engineering is not a tactic. It's an architecture.
An AI Operating System has four layers:
- Context layer — who you are, what you do, what's true, what's current.
- Process layer — the rules, checks, and stages that wrap the AI.
- Trust layer — how you verify the output before it ships.
- Accountability layer — who owns what, who reviews what, who is responsible when it breaks.
Context engineering is layer 1. Necessary, not sufficient.
If you build a gorgeous context system and have no process, no trust, no accountability around it, you still get chaos — just more eloquent chaos. Letter 78 covered that: you cannot automate a mess. Letter 77: oversight without design is cosplay. Letter 72: the AI that gets it right and the AI that gets it wrong sound exactly the same. Letter 76: the only skill that matters is knowing what to check.
The Calboreanu numbers prove it from the context angle. 29% vs 89% is the same story I've been telling for nine months, told with new math.
The AI Chaser buys a subscription, types a prompt, and wonders why the output isn't magic.
The AI Owner builds a context layer, wraps it in a process, installs trust gates, and assigns accountability. Then the output is magic — except it's not magic. It's engineering.
What to do this week
If you read this far, here's what I'd actually do, in order:
1. Write one file. Not a 40-page manifesto. One file. Call it me.md or context.md or whatever. 500–1,500 tokens. Who you are. What you do. How you talk. What you won't do. Save it somewhere you can paste from in 10 seconds.
2. Paste it at the start of every serious AI conversation for two weeks. Just copy-paste at the top. Notice how much the output quality shifts. Notice how much less you rewrite. Notice how much faster you get to "yes, this is right."
3. After two weeks, add a second file. Whatever bugged you most during those two weeks — a recurring correction you kept making, a format you kept asking for, a fact the AI kept getting wrong — write it down. Add it to your stack.
4. Stop there. Don't build the castle yet. A two-file context stack you actually use beats a twelve-file system you built once and abandoned. The Calboreanu data says it clearly: structure matters more than volume.
5. After a month, ask yourself one question: If I got hit by a bus (just a metaphor :) ), could someone else run my AI with these files and get 80% of my output?
If yes, you have a real context system.
If no, keep going.
That's the loop. Write. Use. Refine. Don't automate the writing — remember Gloaguen.
Back to Marc and Sofia
Same model. Same task. Same $20 subscription.
Marc rewrote his proposal for 90 minutes. Sofia sent hers with light edits.
The gap between them wasn't intelligence. It wasn't prompting skill. It wasn't "AI talent." It was the folder. The voice guide. The past proposals. The pricing table. The client brief. The positioning doc.
The context.
My friend, keep this in mind: That gap is going to get wider, not narrower. Because Sofia's context system learns. Every proposal she writes gets added to the stack. Every correction sharpens the guardrails. Every new client brief makes the next one better. The system compounds.
Marc is starting from zero every single time.
Prompt engineering was never the skill. It was the training wheels.
The real skill is quieter and harder: deciding what matters, writing it down, maintaining it, and knowing when to update it. The boring, human, deeply unsexy work of curating your own reality into something a machine can use.
Everyone has access to the same models. Almost nobody has access to your context.
"Generation is now free. Curation is now everything."
AI is only as good as the human operating it. And the human operates it through context.
Build it. Maintain it. Own it.
That's the new engineering.
Have a great weekend.
— Charafeddine (CM)