The models now hold a million tokens of context, but almost nobody has written down anything worth putting in them. Here is what to write, in what order, and why the research says more context is not automatically better context.
The Empty Box Problem
A member of ours got on a call with me last month and said something I have not been able to shake.
“I have been using this thing every day for fourteen months. And every single morning I open it up and it has no idea who I am.”
Fourteen months. Roughly three hundred working days of explaining her business to a machine that forgot her by lunch.
Here is the direct answer to the question in the headline. You get AI to remember your business by writing your business down in files the AI reads at the start of every session. Not a prompt library. A context library: who you serve, how you talk, what you sell, what you have already decided, and what you have already tried and rejected. The model does not need to be smarter. It needs to be briefed.
That is the whole shift. And it landed in the tools in the last few weeks.
On August 3, 2026, Alibaba shipped Qwen3.8-Max with a one million token context window, 2.4 trillion total parameters and roughly 95 billion active per inference, priced at two dollars per million input tokens. A million tokens is somewhere around 750,000 words. That is your last six months of client notes, your sales calls, your SOPs, your entire email history with your best customer, and you would still have room.
A month earlier, on July 7, 2026, Anthropic opened Claude Cowork to web, iOS, and Android, so sessions and files follow you between devices and background work keeps running when your laptop is closed.
The tools got continuity. Most business owners did not.
My thesis is simple: the competitive edge moved from how smart the model is in one exchange to how much it knows about you across all of them. And the only work required is deciding what is worth writing down.
Key Takeaways
- Context, not raw intelligence, is now the difference between a generic AI answer and a useful one.
- A one million token window is roughly 750,000 words, which is more than most small businesses have ever documented.
- Research shows performance degrades as context grows, so curated context beats dumped context every time.
- The practical deliverable is a small set of reference files the AI reads first, not a longer prompt.
- Writing your business down has a second payoff: it makes the business easier to hand off to a human, too.
Why Nobody Writes It Down
I want to be honest about why this does not happen, because “just document your business” is advice you have heard before and ignored, and you were not wrong to ignore it.
Documenting a business is slow, unglamorous work with no same-day payoff. It is the thing you do after the client work, the invoicing, the payroll, and the fire that started at 4pm. There is no dopamine in writing down how you price a proposal.
I did not do it either. For most of my first year with these tools I was doing exactly what our member described. Open the box. Type a paragraph explaining who I am. Get a decent answer. Close the box. Repeat tomorrow. I was getting real value, so I did not notice what I was leaving on the table.
What finally moved me was noticing how much of my day was re-explanation. I was not paying a cost in one big lump. I was paying it in ninety-second increments, forty times a day, and calling it normal.
There is a second reason nobody writes it down, and it is the more interesting one. Most of what runs your business is not written anywhere because it is not the kind of knowledge that lives in sentences. It lives in your hands. You know which prospect is a bad fit in about ninety seconds on a call, and if I asked you to explain how, you would struggle.
The knowledge management researchers have a name for this. Nonaka and Takeuchi called it tacit knowledge, and their SECI model describes externalization, the process of converting what you know but cannot say into concepts, images, and written documents. Their research also flags the honest catch: highly contextual knowledge loses something when you codify it. Some of what you know does not survive the trip to a document.
So the reframe is this. You are not trying to write down everything. You are trying to write down the highest-value repeatable parts, the ones you explain over and over, and let the rest stay in your head where it works fine.
That reframe matters more than you think, because the research on long context says something uncomfortable about dumping everything in.
What the Research Actually Says About Long Context
Here is where the honest version separates from the hype version. A bigger context window does not automatically mean better performance. There is real evidence that it can mean worse.
Every frontier model degrades as input grows. In July 2025, Chroma published a technical report by Kelly Hong, Anton Troynikov, and Jeff Huber that tested 18 frontier models, including GPT-4.1, Claude 4, Gemini 2.5, and Qwen3. They named the effect context rot. As input tokens increase, output quality decreases, and it starts well before the stated limit. A model advertising a 200,000 token window can show meaningful degradation around 50,000.
The failure is not just about finding the needle. A 2025 arXiv paper titled “Context Length Alone Hurts LLM Performance Despite Perfect Retrieval” ran systematic experiments across five open and closed source models on math, question answering, and coding. Even when the models retrieved every relevant piece of information perfectly, performance still degraded substantially as input length grew, with drops reported between 13.9 percent and 85 percent, all within the models’ claimed context lengths. Length itself is a tax.
Position matters. Nelson Liu and colleagues at Stanford published “Lost in the Middle” in Transactions of the ACL, and the finding is now well replicated. Models use information best when it sits at the beginning or the end of the context. Bury the important thing in the middle of a long document and accuracy falls.
And the drop can be steep. Adobe Research’s NoLiMa benchmark, presented at ICML 2025, tested 12 models that all claim at least 128,000 tokens of context. At 32,000 tokens, 10 of the 12 scored below half of their own short-context baseline.
Translated into plain language: the model reads the way a tired person reads. It catches the first thing and the last thing. It skims the middle. And the longer you make it read, the more it skims.
So what do you do with that? You do what Anthropic’s own engineering team recommends in their guidance on context engineering for AI agents. They describe the goal as finding the smallest possible set of high-signal tokens that maximize the likelihood of the outcome you want. Not the most context. The best context.
This is why practitioners who work with these tools professionally have moved away from prompt lists entirely. Julia McCoy’s current Claude protocol is not a stack of clever prompts. It is system prompts, reference packs, training files, and on-brand workflows. The deliverable is a body of context. The prompt is almost an afterthought.
The Context Library: Five Files, Not Five Hundred
Here is the framework I use and teach. It is five files. You can build the first version in an afternoon.
File one: the operating brief. Two pages, maximum. Who you are, what the business does, who you serve, what you sell and at what price, and the three outcomes you are working toward this quarter. This is the file that goes at the top of every session.
File two: the voice file. How you write. Not adjectives like “friendly and professional,” which tell the model nothing. Actual samples. Three to five pieces of your real writing, plus a short list of words and phrases you never use. My own list bans “leverage” as a verb, and I am not sorry.
File three: the customer file. Who buys, who does not, the objections you hear in the order you hear them, and the exact words your clients use to describe their problem. Pull these from real emails and real call transcripts, not from your imagination.
File four: the decisions file. This is the one almost nobody builds and it is the one that compounds fastest. Every time you decide something and rule something out, log one line. “We tried three-day launches in April, open rates held but conversions dropped, going back to five days.” This is what keeps the AI from cheerfully suggesting the thing you already know does not work.
File five: the workflow file. The four or five processes you actually repeat. How a proposal gets built. How a weekly email gets written. What “done” looks like for each.
That is it. Under 10,000 words total for most businesses, which is comfortably inside the range where the models still read carefully rather than skim.
The proof case is my own week. When I sit down to write, I am not explaining White Beard Strategies to anything. The brief, the voice file, and the customer file load first. What used to be twenty minutes of setup and two rounds of “no, that is not how I sound” is now a first draft that is genuinely mine, and my job is editing rather than orientation.
The second payoff surprised me. Those same five files turned out to be the best onboarding document I have ever handed a new team member. I did not write them for humans. Humans use them anyway.
How to Build Yours This Week
Open a plain folder, not a tool. Five markdown or text files in a folder you can point any AI at. Do not spend Saturday evaluating knowledge management platforms. The tool is not the bottleneck; the writing is.
Draft the operating brief first, in one sitting. Set a timer for forty minutes and write it badly. Two pages. You will fix it later, and a rough brief in use beats a perfect one in your head.
Build the voice file from real work, not descriptions. Paste in three to five things you actually wrote and that performed. Add ten words you never use. This single file removes most of the “AI sounds like AI” problem.
Mine your last twenty client emails for the customer file. Copy their actual sentences about what is wrong. Their language, not your marketing language. This is the highest-return hour in the whole build.
Start the decisions file today with three entries. Three things you tried this year and what happened. From here on, add one line whenever you decide something. It takes fifteen seconds and it is the file that makes month six dramatically better than month one.
Load only what the task needs. Given what the degradation research shows, do not paste all five files into every session. Brief plus voice for writing. Brief plus customer for sales. Keep sessions focused and start fresh sessions often.
Review it monthly, on the calendar. Thirty minutes on the first Monday. Delete what is stale, add what changed. A context library that nobody prunes turns into an attic, and the research says the attic makes performance worse, not better.
Frequently Asked Questions
Does a bigger context window mean the AI will give me better answers?
Not automatically. Chroma’s 2025 study of 18 frontier models found output quality drops as input tokens rise, often well before the stated limit. A larger window gives you more room, not more reliability. Curated, well-organized context outperforms a large pile of unsorted context in nearly every case.
How much should I actually write down?
Start with about 10,000 words across five files: an operating brief, a voice file, a customer file, a decisions log, and a workflow file. That is enough for the model to sound like you and know your business, and small enough that it reads carefully rather than skimming the middle.
Is this the same thing as prompt engineering?
No. Prompt engineering optimizes a single request. Context engineering builds the standing body of information the model reads before your request. Anthropic’s own guidance frames the goal as finding the smallest set of high-signal tokens for the outcome you want. Prompts get better automatically once the context is right.
Will my files work across ChatGPT, Claude, and Gemini?
Yes, if you keep them as plain text or markdown in a folder you control. Every major assistant accepts uploaded files or project knowledge. Avoid building your library inside one vendor’s proprietary format so you can move it when a better model ships, which happens roughly every quarter now.
What if my business changes constantly?
Then the decisions file becomes your most valuable one. It logs what you tried and what you ruled out, so the AI stops suggesting abandoned approaches. Fast-changing businesses need shorter files reviewed more often, not longer files. Monthly pruning is the discipline that keeps it useful.
The Only Work Required
Fourteen months. Three hundred mornings of introducing herself to a machine.
I keep coming back to that, because she was not doing anything wrong. Nobody told her the job had changed. The industry spent two years selling prompts, and prompts were the right answer when the window was small and the sessions did not persist. That is no longer the situation.
You now have tools that could hold your entire operating history. Qwen3.8-Max will take 750,000 words. Cowork sessions follow you from your desk to your phone and keep working while you sleep. The container exists. Most people are still typing one sentence into an empty box every morning and wondering why the output feels generic.
It feels generic because it is. You are getting the answer a stranger would give you, because you are talking to a stranger, every day, on purpose.
Here is the part I want you to sit with. The AI that knows your last six months beats the smarter AI meeting you for the first time. Every time. Not occasionally. And nothing about closing that gap requires you to be technical, or early, or to have any idea what a token is.
It requires you to decide what is worth writing down, and then write it down.
That is the whole job. Forty minutes for the first file. Fifteen seconds a day for the one that compounds.
If you want the file templates, the review process, and the training sessions where we build these live with members, that is what our membership is for. The replays are there. The templates are there. Come get them.
Stop introducing yourself to your own AI.
About the author
Jonathan Mast is the founder of White Beard Strategies, where he coaches and mentors entrepreneurs on putting AI to work in real businesses rather than in demos. He speaks regularly on practical AI adoption for small and mid-sized companies, and he is stubbornly anti-hype about it. His own context library started as one badly written page in a folder called “notes,” which he still has, and still uses.
