The Notebook That Holds the Truth About Your Business

Why a small business needs one up-to-date written record of how it works before it hands work to AI, and how to start one that an AI assistant keeps for you

A small-business guide to a single source of truth for the AI era: one notebook of prices, policies and procedures, kept current by an AI assistant using Andrej Karpathy’s LLM wiki pattern and the Open Knowledge Format’s provenance fields, with a worked example you can copy.
Author

Behzad Samadi

Published

October 30, 2026

A paper-craft desk: an open notebook with tabbed sections, loose papers, receipts and a price list feeding into it, and a small glowing assistant writing in its margin. Title: The Notebook That Holds the Truth About Your Business.

Ask three people in a small business what the minimum order is, and you may get three answers. The website’s FAQ says one thing, the price list another, and last month’s team meeting changed it again. Most of the time a person catches it: someone remembers the meeting.

Now put an AI assistant in front of customers, or let it draft your quotes. It will answer from whatever it was given, fluently and with confidence. If it was given the old FAQ, it quotes the old minimum. Nobody remembers the meeting for it.

This article is about the fix that comes before any clever AI: one written record of how your business works today, where each fact lives in one place, says where it came from, and gets checked. Call it a notebook. The new part is that an AI assistant can now do most of the work of keeping it. We built one for an example business and let an assistant maintain it; the results, and the folder, are below for you to copy.

NoteIn short
  1. AI answers from what you give it. Every major assistant can be pointed at your own documents, and none of them can tell that a document is out of date. The quality of its answers is capped by the quality of your records.
  2. Write each fact once, in one place. Prices, policies, procedures, suppliers, decisions: one page each, and everything else (the website, quotes, the assistant’s answers) reads from there.
  3. Let the assistant do the upkeep. In April 2026 Andrej Karpathy described an “LLM wiki”: you drop raw material in a folder, and an AI keeps a set of plain-text pages up to date from it. People abandon wikis because upkeep is boring; an assistant doesn’t get bored.
  4. Keep the receipts. Every fact should name its source, every page should say who wrote it and whether a person has checked it, and each should carry a date to check it again. An open format, the Open Knowledge Format, defines fields for exactly this.
  5. In our test it worked, and found problems we hadn’t planted. For about C$3 of AI time, the assistant built the notebook from five documents, drafted a customer reply without sending it, and listed fourteen questions only the owner could answer.

What we mean by “the truth”

In data management this idea has a name: a single source of truth. Each piece of information is edited in one place only. There can be copies (the website, a printed menu, an answer an assistant gives), but they are read from the master and never edited on their own.

You’ll also hear “digital twin” used for this. Strictly, a digital twin is a model of a machine or a building that stays in sync with it through live data; using the term for any model of a business is, as Wikipedia’s article puts it, a “largely marketing-oriented interpretation”. What a small business needs is humbler and more useful: a notebook that is right.

What goes in it:

  • What you sell and what it costs: services, products, prices, minimums, what’s included.
  • Your policies: deposits, cancellations, delivery, guarantees, hours.
  • How the work is done: the steps for a quote, a booking, a job, a complaint; who does what.
  • Your suppliers and partners: terms, contacts by role, recent changes.
  • Decisions: what was decided, when, and from what date it applies.
  • Questions customers actually ask, with your current answers.

What stays out: anything you’d keep in your accounting system or CRM already (invoices, the customer list itself) and personal details that the work doesn’t need. The notebook can point to those systems; it shouldn’t copy them.

Why it matters more now

Canadian businesses are adopting AI fast. Statistics Canada found that 12.2% of businesses used AI to produce goods or deliver services in the year to spring 2025, twice the 6.1% of a year earlier. Among them, the most common uses were working with text, data analysis and chatbots.

All of those depend on what the AI is given to read. The vendors’ own documentation is plain about this:

  • Claude Projects let you give an assistant a set of your own documents and instructions to work from; OpenAI’s file search does the same for apps built on its platform, and cites the files each answer used.
  • Microsoft’s Copilot Studio “grounds only on the content of the files and pages you register”; switch off web search and general knowledge, and an agent with no answer in your files says it has no response, rather than guessing.

That last behaviour is the right one, and it shows the limit. An assistant grounded in your files can be no better than your files. If two documents disagree, it may pick either. If a document is out of date, nothing in it says so. None of these tools fixes a wrong source; they faithfully repeat it.

We made the same point from the other direction in Train It Before You Use It: don’t automate chaos. The notebook is how you stop the chaos first.

The idea: let the assistant keep the notebook

Businesses have tried internal wikis for decades, and most go stale. The reason isn’t the software. As Andrej Karpathy put it in April 2026, people abandon wikis “because the maintenance burden grows faster than the value”. Every new decision should update five pages; nobody does it.

His proposal, which he calls an LLM wiki, is to give that job to an AI assistant. It has three layers:

  1. Sources. A folder of raw material: emails, meeting notes, price lists, letters. The assistant reads them and never changes them; they are the record of what was actually said.
  2. The wiki. Short plain-text pages, one subject each, that the assistant writes and keeps current. You read them.
  3. The rulebook. One file of instructions (the CLAUDE.md or AGENTS.md that assistants such as Claude Code and Codex read automatically) setting out how pages are written and kept.

And three jobs you ask it to do:

  • Ingest: “Here’s a new source; update every page it affects.” One meeting note might change the price list, the policy page and the FAQ.
  • Answer: “What do we tell this customer?” It answers from the pages, and a good answer can become a page itself.
  • Lint: “Check the notebook”: contradictions, pages past their check-by date, facts without a source, rules whose start date has passed.

Two more files hold it together: index.md, a one-line-per-page catalogue the assistant reads first, and log.md, a dated history of every change. (Our example’s index and log show what they look like.) Karpathy reports that at around a hundred sources the index alone is enough for the assistant to find what it needs, with no special search software.

The idea has sceptics, and their objection is the right one. Commenters on Karpathy’s post pointed out that once a wrong fact is written into the wiki, “errors compound over time”; when a source changes, a page built from it can be wrong without knowing it. The answer isn’t to give up on the idea. It’s to make every page show its working.

Show your working: where it came from, who checked it

The Open Knowledge Format (OKF) is a small open specification for exactly this: a folder of markdown pages, each with a short header saying what it is and how far to trust it. It is at version 0.2 and published in Google Cloud’s GitHub organisation as “a universal, vendor-neutral format”; the project notes it isn’t an official Google product. Collegica publishes its own articles and research in this format so that assistants can read them with their sources attached.

You don’t need the whole specification. Four of its ideas do most of the work in a business notebook:

  • Sources, cited per fact. Every statement carries a footnote naming the document it came from, so “minimum 20 guests” points at the meeting notes where that was decided.
  • Who wrote it, and who checked it, kept separate. The assistant writes; only a person marks a page as checked. When the assistant changes a page, the check is cleared, because the person checked the old text, not the new one.
  • Three levels of trust that follow from that: unverified, checked by a process (a test, a script), or reviewed by a person.
  • A check-again date. Prices and policies go stale. A page that says when it should be reviewed can be flagged automatically once that date passes.

Here is the top of one page from our example, as the assistant wrote it:

---
type: Policy
title: Orders and deposits
sources:
  - id: meeting-2026-10-14
    resource: sources/2026-10-14-team-meeting.md
  - id: price-list-2026-09
    resource: sources/2026-09-price-list.csv
generated:
  by: claude-opus-5-5
  at: 2026-10-30T11:00:00-04:00
verified: []        # nobody has checked this version yet
stale_after: 2027-04-30T00:00:00-04:00
---

We tried it

We wrote five documents for a fictional six-person catering company: an old FAQ copied from its 2025 website, a September price list, a supplier’s letter announcing a price rise, notes from an October team meeting, and an email from a customer. Two contradictions were planted on purpose: the old FAQ’s 15-guest minimum and 25% deposit, both since changed.

We added a rulebook (AGENTS.md, about a page long) and gave the folder to Claude Code, with permission to read and write files in that folder and nothing else. Then we asked three things, one at a time:

Request Time AI cost
“Ingest every file in sources/, oldest first.” 2 min C$1.04
“Answer the customer email.” 1 min C$0.88
“Lint the wiki.” 1½ min C$1.01

Costs are pay-as-you-go API prices (US$2.09 in all, at the Bank of Canada rate of 28 October 2026); on a Claude or ChatGPT subscription the same work is part of the monthly fee.

What it built

Seven pages: orders and deposits, the price list, delivery, dietary options, how quotes are done, the dairy supplier, and a list of open questions; plus the index and the log. Prices were copied exactly. Rules starting on 1 November were written as “from 1 November”, not applied early.

What it found

Both planted contradictions, and three we hadn’t noticed when we wrote the documents:

  • “Minimum order 20 guests for every menu” (the meeting) would, read literally, lower the hot buffet’s minimum from 25 and the cocktail reception’s from 30. Was that intended? It didn’t apply it until a person confirms.
  • The meeting notes say butter and cream go up 8%; the supplier’s letter also names cheese.
  • “Existing bookings are honoured” doesn’t say whether a booking exists from the enquiry, the confirmation or the deposit. For the customer who emailed on 21 October about a lunch for 16 in November, that decides whether the order can be taken at all.

That last one is the kind of gap that costs a small business a customer, or a margin, and it was invisible until something tried to answer from the written rules.

What it did with the customer

It drafted a reply, in a file, for a person to check and send: the menu and price, the gluten-free option with three days’ notice, and a clear warning that the kitchen can’t promise a gluten-free environment for a guest with coeliac disease. Where the rules were unclear, it wrote both versions of the deposit paragraph and asked a person to choose. It left the total to the quote template rather than doing the arithmetic itself, as the meeting had decided.

Its one slip

The draft opened “we’d be glad to do a working lunch for your team of 16”, which is wrong if the 20-guest rule applies. Its own lint caught it and added it to the open questions. That’s the point of the design: the assistant’s work is visible, dated and unverified until a person says otherwise.

The full folder (the five sources, the rulebook, every page the assistant wrote and a note on the run) is on GitHub.

Start yours this week

  1. Make a folder with two subfolders, sources and wiki. A shared drive is fine; so is a folder on one computer, to begin with.
  2. Copy the rulebook. Take our AGENTS.md, change the first paragraph to describe your business, and adjust the page types to yours.
  3. Gather ten documents that define how you work today: the price list, the policies on your website, your quote template, the last few meeting notes, the emails where something was decided. Put them in sources, each file name starting with its date.
  4. Open the folder in an assistant that can work on files: Claude Code, Claude Desktop, ChatGPT’s desktop app or Codex. Our guide to desktop AI explains the options. Ask it to ingest the sources, oldest first.
  5. Read the open questions and answer them. The assistant writes them to open-questions.md in your wiki folder (here is the one from our example). This is the valuable part: each answer is a decision you’ve now made once, in writing. Put your answers in a new dated note in sources and ask it to ingest that.
  6. Check the pages that matter most (prices, deposits, anything a customer sees), then mark them verified, with your name and the date.
  7. From then on, every decision goes into sources as a note and gets ingested. Ask for a lint once a month, and whenever a dated rule kicks in.

When the notebook is solid, point your assistant at it for customer questions and quote drafts, and have a person send what goes out.

Things to watch

  • Personal information. Canada’s private-sector privacy law expects you to collect and keep only what the purpose needs. Keep customers’ names, contact details and anything about their health in the systems built for them, and out of the notebook. Our rulebook says so; check that the assistant follows it. If you use an online assistant, its provider sees what you give it; our privacy notes for small businesses cover what to check.
  • Errors can compound. A wrong fact copied into three pages is harder to spot than one. The footnotes are your defence: every fact can be traced back to a dated source in seconds.
  • Size. The simple index works for a notebook of hundreds of pages, which is plenty for most small businesses. Beyond that you’ll want search tools, and probably help.
  • It’s not your accounting system. Invoices, payroll and the customer list stay where they are. The notebook explains how the business works; it doesn’t replace the systems that record transactions.
  • A person stays responsible. The assistant drafts; nothing it writes goes to a customer, or counts as checked, until a person says so.

Later: letting other tools read it

Once the notebook exists, it can serve more than one assistant. Two open conventions are worth knowing about, though neither is needed to start:

  • llms.txt is a short file at the root of a website that tells AI assistants where its clean, current information is. Publishing your public pages (prices, policies, FAQ) that way lets an assistant a customer uses find the right facts. Ours is an example.
  • The Model Context Protocol (MCP) is an open standard that lets an assistant ask a system a question or run a defined action (“look up this booking”) and get a structured answer back. It’s how a notebook, a booking system and an assistant can eventually work together.

Next: a construction project

Construction has worked on this problem longer than most industries. On a building project, drawings, specifications, RFIs and change orders pass between dozens of companies, and the cost of someone working from an old revision is measured in rework. The international standard for managing that information, ISO 19650, is built around a common data environment: one place where each document lives, with a status that moves from work in progress to shared to published. That’s the notebook idea at the scale of a building. A follow-up article will look at what it means for a small contractor, and how the same assistant-kept notebook fits alongside it.

Sources