Every time you open ChatGPT or Claude and start explaining what you do, how you like things done, and what you're working on — you're wasting your best thinking time on context that should already be there.

It's the single biggest productivity gap in how most professionals use AI today — and it has nothing to do with which model you're using.

The fix has a name: context engineering.

Context engineering vs prompt engineering: what actually changed

In June 2025, Shopify CEO Tobi Lütke posted a tweet that crystallised something practitioners had been feeling for months:

"I really like the term 'context engineering' over prompt engineering. It describes the core skill better: the art of providing all the context for the task to be plausibly solvable by the LLM."

That distinction matters. Prompt engineering is about crafting a clever question — a skill that's already going obsolete. Context engineering is about designing the information architecture around the question — so the AI already knows who you are, what you're working on, and how you like things done before you even ask.

Phil Schmid from HuggingFace built on this, framing context engineering as designing systems — not strings — that deliver the right information at the right time. Anthropic's own engineering team now describes it as "finding the smallest possible set of high-signal tokens that maximize the likelihood of some desired outcome."

Translation: the prompt isn't the problem. The problem is your AI knows nothing about you.

Why this matters more than your choice of model

Here's something that might surprise you: a UC Berkeley study analysed over 1,600 multi-agent system failures and found that 36.9% were caused by inter-agent misalignment — AI systems operating on inconsistent views of what they were supposed to be doing. Not because the models were incapable, but because the context flowing between them was wrong.

That's multi-agent systems, but the principle scales down to you and your AI tools. Most AI failures are context failures, not model failures.

Think about it. When ChatGPT gives you something generic and unhelpful, is it because GPT-5 is stupid? Or is it because you gave it nothing to work with beyond a single question?

When you tell your AI assistant to "write me an email to a client" and it produces something that sounds nothing like you — that's not a model problem. It's a context problem. The AI has no idea who you are, what your voice sounds like, or what kind of relationship you have with that client.

The attention budget: why dumping everything in doesn't work

Before you go and paste your entire life story into a system prompt, there's a catch. More context isn't always better context.

Google's own research with Gemini 2.5 found that beyond about 100,000 tokens of context, the AI started "favouring repeating actions from its vast history rather than synthesising novel plans." It got worse with more information, not better.

Drew Breunig identified four ways that context fails:

Poisoning. A hallucination or error gets into your context files, and the AI treats it as truth from then on. It compounds over time.

Distraction. So much context that the AI over-focuses on accumulated history instead of thinking fresh. It becomes a parrot of its own past.

Confusion. Irrelevant information influences responses. The AI doesn't filter — it uses everything you give it, whether it's relevant or not.

Clash. Parts of your context contradict each other. The AI tries to reconcile conflicting instructions, usually badly.

The solution isn't more context. It's the right context, structured so the AI gets what it needs for the specific task at hand.

Five files. Ninety minutes. Done.

Here's the practical part. You can set this up in about 90 minutes with whatever AI tools you already use — Claude Projects, ChatGPT custom instructions, Cursor, or even just a folder of markdown files.

1. Your identity file

Not your CV. Your working identity. The things an AI needs to know to stop being generic and start being useful.

Your role, your industry, how you communicate, what you're opinionated about. The AI should know your positions, not just your job title. If you think most marketing advice is vapid nonsense, say so. If you write in short sentences and hate corporate jargon, put it in writing.

What would you tell a sharp new colleague on their first day? Not your life story — just enough to stop them making dumb assumptions. A few hundred words is plenty.

2. Your active projects file

The file that stops you saying "so I'm working on this project where..." every single session.

What are you working on, what's the status, what are the constraints? Project names, one-line descriptions, who's involved. Update it when things change — weekly is usually enough. It doesn't need to be comprehensive; it just needs to give the AI enough to not ask stupid questions about things you've already decided.

3. Your preferences file

How do you like things done? Everyone has opinions about this, but nobody tells the AI.

Do you want bullet points or prose? Technical detail or executive summary? Do you hate it when AI starts every response with "Great question!"? (You should.) Write it down. Format preferences, tools you use, pet peeves. The AI will stop doing them immediately — but only if you tell it.

4. Your decision log

This is the one most people skip, and it's arguably the most valuable. A running log of key decisions you've made and why.

Why? Because without it, the AI will relitigate settled questions. It'll suggest things you've already considered and rejected. It'll propose approaches you've deliberately moved away from.

A decision log doesn't need to be fancy. Just:
- The decision
- When it was made
- Why (one sentence is fine)
- What was rejected and why

5. Your constraints file

The guardrails. What should the AI never do, always do, or be careful about?

Brand voice rules. Compliance requirements. The "don't ever suggest this" list. If you work in finance, maybe the AI should never speculate about market movements. If you're a lawyer, maybe it should always caveat with jurisdiction. Without explicit constraints, the AI guesses at where the edges are — and it will guess wrong. That's a manageable irritation with a chatbot, and a real risk once the AI can act on its own rather than just answer.

How to use these files with the tools you already have

The beauty of this approach is it works everywhere. The files are just text — markdown, plain text, whatever you prefer.

Claude Projects makes this easy — add your files as project knowledge and every conversation starts with your context loaded. Separate projects for different work streams keeps things focused.

ChatGPT has custom instructions for your identity and preferences, plus the memory feature for building context over time. Custom GPTs can hold project-specific context permanently if you're on a paid plan.

Cursor, Windsurf and the coding agents load rules files automatically into every interaction. Cursor has moved on from the old single .cursorrules file to scoped rules under .cursor/rules/ — and through 2026 the field largely converged on AGENTS.md, a plain-markdown file in your project root that Google, OpenAI, Cursor and the rest have agreed to read. If you code, that's where your decision log and constraints belong.

The tool doesn't matter much. If it supports file uploads or system prompts, you can do this. The files are just text.

This is what practitioners call progressive disclosure — a concept borrowed from UX design. Don't dump all five files into every conversation. Load the identity file always. Load the project file for project work. Load the preferences file when you need output in a specific format. The AI gets what it needs for this task, nothing more.

The difference in practice

Ask any AI tool to "write a blog post about AI adoption in small businesses" cold, and you'll get something generic and lifeless. It could have been written by anyone, for anyone.

Now ask the same question when the AI already knows you run a consultancy, that you prefer a conversational tone, that you've written 15 articles on related topics, that your audience is business owners rather than developers, and that you think most AI advice is too abstract.

Same prompt. Completely different output. One sounds like a Wikipedia article. The other sounds like you.

Where this is heading

Right now, context engineering is mostly manual. You create the files, you maintain them, you decide what to load when.

That's already changing. Muratcan Koylan's "Agent Skills for Context Engineering" — a file-based architecture for AI context that exploded on X — was recently cited in a Peking University research paper on "Meta Context Engineering." The academic world is now studying how these context systems can evolve and improve themselves over time.

The trajectory is clear: from manual curation, to semi-automated, to fully self-maintaining context architectures — the same shift that turns a generic chatbot into a personal AI agent that actually knows you. But you don't need to wait for that future. The manual version works brilliantly today, and it takes less than two hours to set up.

Start here

If you do nothing else after reading this:

  1. Create your identity file. Spend 15 minutes writing down who you are, what you do, and how you work. Load it into whatever AI tool you use most.

  2. Add your active projects. Another 15 minutes. What are you working on, what's the status, what decisions have been made?

  3. Note your non-negotiables. What should the AI always or never do? Write them down.

That's 45 minutes. You'll feel the difference in your very next AI conversation.

Then, over the next week, pay attention to the moments where the AI gets something wrong because it didn't know enough about you. Each of those moments is a signal: that's context you should be providing upfront.

The people getting the most from AI tools aren't writing better prompts. They're engineering better context. The gap between "AI is useless" and "AI is transformative" is almost always a context gap.

Close it.