Here’s the pattern: someone asks an AI to “write a cold email for my business,” gets something that sounds like a LinkedIn bot wearing a suit, and concludes the tools are overhyped. Then they watch a colleague get output that’s specific, on-voice, and nearly shippable — from the same model, on the same day.
The difference is almost never prompting talent. It’s that the second person stopped making the model guess. Every AI session starts from zero: the model doesn’t know what you sell, who buys it, how you talk, or what “good” means to you. Most people fix this by re-typing fragments of that context every session — badly, from memory, forever. The fix is to stop treating context as something you type and start treating it as something you own.
A context pack is five short files, written once, attached or pasted whenever you work with any AI — ChatGPT, Claude, Gemini, whatever comes next. It’s portable, it compounds, and it turns every future prompt from a cold open into a briefing of a colleague who already knows the operation. Total build time: about an hour. Here’s the whole system.
The five files
Keep each file under a page. The pack works because it’s dense, not because it’s complete — a model drowning in twelve pages of backstory performs worse than one holding five tight ones.
1. brief.md — what you do
What I do: one paragraph, plain language, no mission-statement cosplay.
Who pays me: the 2–3 real customer types, with a sentence on what each actually buys.
The numbers that matter: rough price points, typical deal or project size, current volume.
This year’s goal: one line. Explicitly not doing: one line — this kills half of AI’s bad suggestions on contact.
2. voice.md — how you sound
Ten rules and two samples beat any adjective list. “Professional but friendly” means nothing; “short sentences, no exclamation points, never say leverage, always name a specific detail from the recipient’s work” is executable. End the file with one short before/after pair: a paragraph in generic-AI voice, then the same paragraph the way you’d write it. Models imitate examples far better than they follow adjectives.
3. audience.md — who’s reading
For each audience you regularly write to (clients, subscribers, your boss, procurement), three lines: what they already know, what they care about, what makes them stop reading. This is the file that stops AI from explaining basic concepts to experts and jargon to beginners.
4. standards.md — what “done” looks like
Your definition of acceptable work: formats you want by default (“emails under 150 words,” “always give options as a table”), tools in your stack, hard constraints (compliance rules, words you legally can’t use), and how you want uncertainty handled (“flag guesses instead of smoothing over them”). This file replaces the corrections you currently make on every third output.
5. now.md — what’s true this month
Current priorities, active projects, what changed recently. The other four files are stable; this is the only one you maintain. It’s also the one that makes answers feel current instead of generically plausible.
Build it in one hour — by being interviewed
Don’t write these from a blank page. Make the AI extract them from you. Open your usual assistant and run:
“I’m building a reusable context pack: brief.md, voice.md, audience.md, standards.md, now.md. Interview me one question at a time — short, concrete questions, no more than 12 total. Push back when my answers are vague. Then draft all five files in tight plain language, each under 300 words.”
Two moves make this work. First, paste in three samples of your real writing before the interview starts — the voice file drafts itself from evidence instead of self-image. Second, treat the draft as 70% done and spend twenty minutes correcting it, because the corrections are the product: everywhere the draft says something slightly off is exactly the guessing you’re trying to eliminate.
Deploy it everywhere
Store the pack as five plain-text files anywhere you can copy from fast. Then load it by default: paste into custom instructions or project knowledge in tools that support persistent context, or start ad-hoc sessions by attaching the relevant files with one header line: “Use these files as ground truth about me. If my request conflicts with them, ask.” You rarely need all five — a pricing question wants brief + standards; a newsletter draft wants voice + audience + now.
The payoff shows up immediately in output quality. A freelance designer asking for a cold email gets “I hope this finds you well! I’d love to explore synergies…” without a pack. With one, the same request returns: “Saw the rebrand you shipped for Hollis Coffee — the menu redesign didn’t make it to your packaging. That gap is what I fix, usually in two weeks at $3–5k.” Same model. The second one knows the niche, the proof style, and the price floor — because they were written down once.
Maintenance: ten minutes a week
Update now.md during your weekly review — new priorities in, dead projects out. Re-run the full interview quarterly; businesses drift and packs must follow. If you catch yourself making the same correction to AI output twice, that correction is a new line in standards.md. That’s the whole loop: the pack gets smarter every time the output annoys you.
Failure modes
- The novel. A 15-page pack is worse than a 5-page one — buried context dilutes attention and answers get mushy. If a file passes one page, cut it back.
- The stale now.md. Confidently outdated context beats no context at being wrong. If you won’t maintain it weekly, date-stamp the first line so you can see the rot.
- Secrets in the pack. No passwords, API keys, client-confidential terms, or anything you couldn’t paste into a third-party tool. The pack describes your operation; it doesn’t contain the vault.
- The aspirational brief. Writing “we serve enterprise clients” when you have three freelance gigs makes every answer wrong in the same direction. The pack must describe the business you have, not the one you’re manifesting.
- One mega-file. Merge everything into a single doc and you lose the ability to attach selectively — and selective attachment is half the power.
Our take: Most people’s AI usage is pure consumption — type, receive, discard, repeat. A context pack is the first move that turns usage into an asset: an hour of work that upgrades every session after it, on every model, including ones that don’t exist yet. Models will keep changing; the description of your business, your voice, and your standards transfers. In a world where everyone has access to the same intelligence, owned context is the moat.