You ask an AI to draft a post about shipping and it cheerfully promises “free delivery nationwide in 24 hours”. You charge $8 and ship to two countries. The next draft quotes last spring’s price; the one after that mentions a store you closed in March. You keep fixing the same facts, and at some point writing it yourself starts to look faster.
The model isn’t the problem. It knows nothing about your business beyond whatever happened to land on the public web, and it fills the gaps with something plausible. A tone of voice document controls how the AI writes, not what it says. The what is the job of a brand knowledge base: a set of documents the AI pulls facts from on every draft, not only when someone remembers to paste them into the prompt.
Below: what goes into the base, how to write documents a model reads correctly, how to build the whole thing in one evening and test it with ten questions. And when it won’t help.
Prompt, chat memory, knowledge base: what’s the difference
“Training AI on your brand” sounds expensive and technical. In practice almost nobody in social media fine-tunes a model: it takes time, costs money and goes stale the first time your prices change. What works is handing the model the right facts at the moment it writes.
| Method | How it works | Where it breaks |
|---|---|---|
| Facts in the prompt | You paste prices and terms into every request | Tedious by day three; everyone on the team has their own version |
| Chat memory | The model remembers what you told it earlier | Lost in a new chat, invisible to colleagues, hard to correct |
| Uploaded file | A PDF deck inside a project or custom GPT | Outdated the moment it’s uploaded; nobody re-uploads it |
| Live knowledge base | The tool reads your documents on every draft | Works as long as someone keeps the documents current |
The gap between rows three and four looks trivial until a price changes. The uploaded PDF will quote the old number for months; the live document stops the moment you fix the cell in the spreadsheet.
What goes in: eight documents
You don’t need to dump the whole shared drive. The model needs the facts that show up in posts and customer replies:
- Company card. Who you are, who you serve, locations, opening hours, contact channels, payment methods. One page.
- Price list. A table, not a paragraph: product, price, currency, what’s included. Sell in both USD and EUR? Separate columns, never “approx.”. Temporary prices get a “valid until” column.
- Products and services. What’s included and — crucially — what isn’t. A line like “In-home measuring is free within city limits, $40 outside” answers half your DMs before they arrive.
- FAQ from real questions. Not invented “why customers choose us”, but 20–30 actual questions from DMs, comments and the support inbox over the last three months, with your answers.
- One-page tone of voice. Sliders, vocabulary, bans. How to build one in an evening is covered in our guide to brand tone of voice.
- Reference posts. 10–15 posts you’re happy with, each with a note on why.
- Stop-list. Topics you don’t touch; promises you don’t make (“guaranteed results”, “fixes it in one session”); competitors you don’t name; regulated claims in health or finance.
- Current. Promotions with end dates, new locations, holiday hours. The shortest document and the one that changes most often.
Add case studies with client numbers only if you have permission to publish them. Internal material — margins, contracts, customer personal data — stays out entirely: anything in the base can end up in a draft.
How to write documents a model reads correctly
A model reads a document literally. Marketing language comes back out as marketing language, and anything vague gets filled in by the model itself.
Compare:
Bad: “Flexible pricing, fast shipping worldwide, personalized approach.”
Good: “US shipping: $8 flat, free over $75, 2–4 business days. Canada: $15, 5–8 business days, duties paid by the customer. We don’t ship anywhere else. Local pickup: Austin store, Mon–Sat 10:00–19:00.”
The first version produces “we ship anywhere!”, the second a reply you don’t have to edit. A few rules:
- One fact, one place. If a price lives in both the price list and the FAQ, within a month they’ll disagree and the model will pick the wrong one. In the FAQ, write “prices — see price list”.
- Spell out the negatives. “No shipping to Mexico”, “no payment plans”, “closed on Sundays”. Without those lines, the model assumes yes.
- Date everything temporary. “15% off annual memberships — until October 31, 2026.” An undated promotion lives in your drafts forever.
- Units and currency on every line. “From 300” — dollars, euros, minutes, units?
- Short headings. “Shipping”, “Payment”, “Warranty” — both the model and your colleagues find things faster.
Build it in one evening
For a small business the realistic effort is 3–4 hours. In this order:
- 30 minutes: customer questions. Scroll through three months of DMs and comments and write down the questions that repeat. That’s your FAQ draft and the list of things the model must know.
- 40 minutes: price list and company card. Put prices straight into Google Sheets — a table is easier to keep current than prose.
- 40 minutes: services with “included / not included”.
- 60 minutes: tone of voice and reference posts. If you already have a tone of voice doc, just pick the posts.
- 20 minutes: stop-list and “Current”.
- 30 minutes: the test. Ask the AI ten questions you know the answers to: “How much is shipping to Toronto?”, “Are you open on New Year’s Day?”, “Is there a warranty?”, “Can I pay with Apple Pay?”. Add two questions the base doesn’t cover. The right response to those is “I’ll check with the team”, not an invention.
When an answer is wrong, fix the document, not the prompt. A prompt lives in one chat; a document shapes every draft that follows.
A live base: who updates it and when
An outdated base is worse than none. With no base, the model writes generic filler and you correct it. With an outdated one, it confidently quotes last year’s price, and that kind of mistake slips through.
So the base has to live where you already work. In Repost21 it connects live to Google Docs, Sheets or Notion: change a price in the sheet and Luna’s next draft uses the new one, with nothing to re-upload. Supported sources and setup are in the documentation.
Beyond the connection, you need:
- An owner. One person responsible for keeping it current. Not “the team” — a name.
- Triggers. Price change, new location, promotion ended — updated the same day.
- 15 minutes a month. Rerun the same ten test questions. A wrong answer means two documents contradict each other somewhere.
When a knowledge base won’t help
- You post rarely and write it yourself. At 4–6 posts a month the facts are in your head, and a base is extra paperwork.
- The facts change daily. Market prices, exchange rates, stock levels. That needs an integration with your inventory system or a manual check of every number before publishing.
- The content has no facts in it. Memes, trend reactions, the founder’s personal stories — a base won’t make them funnier or more alive.
- A dump instead of a base. Two hundred pages of decks from the last three years: the model will find the contradictions faster than you and pick the wrong version.
- Nobody proofreads anymore. Even with a good base, a model can get a number wrong. Prices, dates and addresses get a human check before publishing.
A knowledge base is the dullest and fastest-paying step in working with AI: one evening, and drafts start arriving with the right prices and addresses. It’s also what the next step stands on — AI agents that publish on their own: without facts, an agent just makes things up with confidence. You can try the knowledge base plus Luna on your own documents on the free plan.