Files Not Chats: The 6-File Context System for a One-Person Business

elisabeth hitz · july 19, 2026 · 7 min read

The fastest way to make AI stop producing generic output is to stop talking to it and start giving it files. Six of them. Written once, maintained monthly, attached selectively. This is the entire system, and it costs nothing but an afternoon.

Most AI advice tells you what to type. This is about what to build so that what you type stops mattering so much.

Why files instead of chats?

Because a chat is a room you have to furnish every time you walk in, and a file is furniture that stays.

The technical version comes from how agentic systems handle this at scale. Anthropic's engineering team describes a "just in time" approach: instead of pre-loading everything into context, an agent holds lightweight references, file paths, links, stored queries, and loads what it needs at runtime. They compare it directly to human cognition: we do not memorize entire corpuses, we build filing systems and retrieve on demand.

There is a second reason, which matters more commercially. Anthropic notes that file structure itself carries signal, a file named test_utils.py in a tests folder implies a different purpose from the same file in src/core_logic/. Naming conventions, hierarchy, and timestamps tell a model how and when to use information, before it reads a word of the contents.

Meaning: the organization of your context is itself context. A well-named folder is doing work that no prompt can do.

What are the six files?

This is the minimum viable context layer for a business with one person in it. Plain markdown. No app required.

context/
├── 01-identity.md      Who I am, what I actually do, what I refuse to do
├── 02-audience.md      Who I sell to, in their words, with real quotes
├── 03-voice.md         How I write, with three real examples, not adjectives
├── 04-offers.md        What I sell, prices, what each is for, what it is not
├── 05-proof.md         Receipts. Numbers, testimonials, case outcomes, dates
└── 06-decisions.md     What I have already decided and will not relitigate

1. 01-identity.md, who you are and what you refuse to do Not a bio. An operating description. What you build, who for, what your thesis is, and critically: your banned moves. "Never suggest live calls." "Never say 15 years, it is 10." The refusals are more valuable than the affirmations, because they stop the AI from producing plausible output you would never ship.

2. 02-audience.md, your customer in their own language Not a persona document with a stock photo. Actual sentences your customers have written or said. Their objections, verbatim. What they call the problem before they know the industry term for it. This file is what makes AI copy stop sounding like AI copy.

3. 03-voice.md, examples, not adjectives "Direct, punchy, no fluff" is meaningless to a model, because every brand describes itself that way. Paste three things you actually wrote and liked, and one thing you wrote and hated with a note on why. Anthropic's guidance on this is explicit: curate a set of diverse, canonical examples rather than exhaustively listing rules, because for a model, examples are the pictures worth a thousand words.

4. 04-offers.md, what you sell and what each thing is for Every offer, price, who it is for, and, the part people skip, who it is not for. Without the exclusions, AI will recommend your highest-priced offer to everyone, which is how you get pitches that read like a stranger wrote them.

5. 05-proof.md, dated receipts only Real numbers, real testimonials, real outcomes, with names and dates. This is the anti-hallucination file. If it is not in here, the AI does not get to claim it. Maintaining this file is also, not coincidentally, the single best thing you can do for AI search visibility, because specific dated claims are what models cite rather than paraphrase.

6. 06-decisions.md, the settled questions Every decision you have made and are done making. Pricing is locked. No new landing pages. This platform, not that one. Without this file, you re-argue the same strategic question with a machine every three weeks, and it always argues well, because it does not remember losing.

How do I use them without breaking the context window?

This is where most people fail. They build the files, then attach all six to everything, and wonder why quality dropped.

Attach only what the task needs. Anthropic's guidance on context is that good curation means finding the smallest possible set of high-signal tokens that maximize the likelihood of your desired outcome. More context is not better context, as tokens accumulate, recall degrades. That is context rot, and you can cause it yourself with your own files.

If your instinct is "but everything is relevant", it isn't, and that instinct is exactly what produces mediocre output.

How do I keep them from going stale?

Stale context is worse than no context, because it produces confidently outdated output and you cannot see it happening.

Three habits:

1. 1. Note back out after good sessions. At the end of a session that went well, ask for a summary of decisions made and open questions, and paste it into 06-decisions.md. Anthropic calls this pattern structured note-taking, or agentic memory, the agent persists notes outside the context window and pulls them back later. It is the highest-return ninety seconds in this entire system.
2. 2. Date every edit. Updated: 2026-07-19 at the top of each file. When you see a date six months old, you know to check it.
3. 3. One monthly pass. Fifteen minutes, first of the month, all six files. Prune more than you add. The goal is not a bigger context layer, it is a currenter one.

Why does this beat buying a tool?

Because you can see it.

Memory apps, second-brain platforms, and knowledge graphs all solve this problem at a scale you do not have. Their failure mode is invisible: they remember things that stopped being true, and you find out when the output is subtly wrong in a way you cannot trace.

With six files you can open, read, and edit, you always know exactly what the AI is working from. For a one-person business that is not a compromise, it is a superior architecture. Buy tooling when you have a genuine volume problem: many agents, many people, or data too large to read. Before that, files win.

The discipline behind all of this is in context engineering for operators, and the reason it matters for how customers find you at all is in the assistant is the new homepage.

The part that makes this an asset, not a hack

A prompt is free and copyable. A context layer is neither.

Once these six files exist and are current, they are the thing that makes any model, this year's, next year's, behave like it knows your business. They survive model updates. They survive platform shutdowns. They get more valuable as models improve, because a smarter model does more with the same well-organized information.

That is what "files not chats" actually means. Not a preference for a file format. A bet that the durable asset is the structured information, and the interface on top of it is disposable, a bet OpenAI just confirmed by deleting its own browser and keeping the capability.

or just follow along. new field notes most weeks on x, instagram, and tiktok.