A business AI context pack has six sections. Here they are in order of how much money each one saves, with real examples of what good and bad look like — because "add your business details" is advice that helps nobody.
If you're wondering why you'd want one at all, start with why your AI starts from zero every morning. This piece is the anatomy.
1. Canonical facts
The highest-value section, because it's where AI does the most damage when guessing. Legal name, locations, service areas, hours, contact details, and the prices you actually charge right now.
❌ "We offer competitive pricing on spring replacements."
✅ "Garage door spring replacement: $150–$350 depending on door weight and torsion vs. extension. Same-day service available. We do not charge a call-out fee within the OKC metro."
The second version means the assistant can write an accurate quote unsupervised. The first guarantees it invents a number.
One rule that matters more than it sounds: one authoritative version of each fact. If your phone number appears twice with different formatting, you've created ambiguity in the exact place you can least afford it — and inconsistent business details also hurt you with AI search engines, so it's worth fixing once for both purposes.
2. Offers and boundaries
What you sell — and, just as important, what you don't. Boundaries prevent the failure where AI cheerfully promises work you don't do.
✅ "We service residential and light commercial garage doors. We do not do: gate automation, structural framing, or new-construction rough-in. If asked, refer to [partner] rather than declining flatly."
Notice the last clause. Telling the assistant what to do instead is worth more than telling it to stop.
3. Your voice — taught with examples
This is where most attempts fail, because people write adjectives. "Professional yet friendly" transfers almost nothing. Examples transfer everything.
❌ Voice guidance: "Be warm, professional, and approachable."
✅ Voice guidance: "Direct and plain. Lead with the answer, then explain. Contractions fine. Never use 'utilize,' 'leverage,' or 'seamless.' — Instead of: 'We would be delighted to assist you with your inquiry.' — Write: 'Happy to help — here's what that job usually runs.'"
Two or three real before/after pairs beat a paragraph of description. Pull them from emails you've actually sent.
4. Rules and red lines
The things that cost real money when a machine guesses. Discounts you never give, claims you never make, language your industry requires.
✅ "Never promise a specific arrival time — say 'same-day' or a window. Never quote a firm price on an unseen door. Never claim we're 'licensed and bonded' — we're licensed and insured; the distinction matters legally."
Every business has five to ten of these. They're usually the things you'd correct a new employee about in their first week. Write those down.
5. Reusable specifics
The numbers, timelines, and local details that make output credible rather than generic. Lifespans, seasonal patterns, what your climate does to your equipment, typical job durations.
This section does double duty: the same specifics that make internal AI output good are what make your public content citable by AI search engines. Write them once, use them in both places.
6. Working conventions
How you want work delivered — formats, templates, naming, tone by channel. Quote structure, whether estimates include tax, how invoices are numbered, what a good customer reply looks like versus a good internal note.
This is the section that turns an assistant from "writes decent English" into something closer to a trained employee who knows your house style.
The counterintuitive part: write down what doesn't work
Most people write only capabilities. Include the failure modes too — the jobs that go wrong, the customers you're a bad fit for, when the cheaper option genuinely wins.
Two reasons. First, an assistant that knows your limits gives better advice and stops overselling on your behalf. Second, this is the same instinct that makes public content get cited: balanced sources beat promotional ones. Businesses that can articulate their own limits are simply more useful to work with, human or machine.
Practical notes
- Length: a few hundred lines is plenty. Longer isn't better — it dilutes. If it's sprawling, split by topic and let the assistant load what's relevant.
- Format: the open Agent Skills spec — a
SKILL.mdfile, plain Markdown plus a little metadata. Supported by ~40 tools including Claude, ChatGPT, Copilot, Cursor, and Gemini CLI. Same file everywhere. - Maintenance: update it when reality changes — prices, service areas, a new service line. Stale context is worse than none, because it's confidently wrong.
- Where it lives: in your team's AI tools. Not on your website — that does nothing for visibility, and if it contains internal pricing rules you'd rather competitors not read, publishing it is actively a bad idea.
How to actually write one
The fastest honest method: for one week, note every time you re-explain something to an AI assistant. That list is your outline. Most businesses discover they repeat the same eight things constantly.
Then write it plainly, test it on your real recurring tasks — quotes, replies, listings — and compare output with and without. The difference is usually obvious within two tries.
Or have it interviewed out of you. Most of what belongs in a context pack is knowledge you have but have never written down, and a good interview surfaces it faster than a blank page ever will.
Book the free 45-minute consultation. We'll interview your business, show you our own context pack, and scope yours — you leave with a plan whether or not you hire us. Schedule it here, or see how Spotlight Skills works.