Build An AI Knowledge Base Before You Ask For Anything
Written By: Shane Clark on September 6, 2026
What is an AI knowledge base?
An AI knowledge base is a set of small, focused files that describe your subject, and the AI reads them before it does any work. Instead of one paragraph of context, it gets your vocabulary, your constraints and your standards. As a result the output stops sounding generic and starts sounding like you.
You gave it a paragraph and judged the answer
Here is how most people meet AI. First they open a chat window, describe their situation in a few sentences, ask for something genuinely hard, and read back something fluent that could have been written about anybody.
So they decide the tool is overrated.
Fluent and generic is the worst possible result, because it is much harder to catch than output that is obviously wrong. Bad writing announces itself. Smooth, confident, hollow writing slips through, and you only notice weeks later when somebody asks a question the text cannot answer.
The model is not the problem. You gave it three sentences about a subject you have lived inside for years, and then you graded it on the gap. Of course it filled that gap with averages. Averages are all it had.
In short, context is an input, exactly like the prompt. Yet most people skip it entirely, and then they blame the output. An AI knowledge base is simply that input, written down once instead of retyped badly every session. I have argued before that learning the tool is the easy part, and this is the part that actually takes work.
This is not just for businesses
Because I build these for companies, that is the version I talk about most. Still, the idea is far wider, and the wider version is the one worth understanding first. Anything with a goal and a vocabulary can carry an AI knowledge base.
Take a nursing program. A student could ask an AI to help them study, and it will produce something reasonable for a nursing student who does not exist. However, imagine it had already read a short file on which course she is in and what the syllabus really covers, another on the terms her instructors use in their own particular way, another on the two rotations she has left and when her exams fall, and one honest file admitting that she memorises drug names instead of reasoning through mechanisms. Same question, completely different study plan. I went deeper on exactly that in learning with AI, which walks through the markdown structure behind a study brain.
Or take a diet. Once the AI knows your kitchen, your budget, your training schedule, the foods that trigger your reflux and the fact that you fall apart at nine in the evening, meal planning stops being generic advice and starts being yours.
Similarly, a renovation has a budget, a sequence, a set of trades and a list of things that always go wrong. A hobby has a vocabulary outsiders never hear. Each of those deserves the same treatment.
The eight things any subject needs
Underneath every version of this sits the same short spine. Simply put, eight questions, and they hold whether the subject is a machine shop, a marathon or a nursing degree.
- What this is about. The subject itself, plainly stated.
- What you want. The goal, and what finishing actually looks like.
- Your vocabulary. The words you use, defined the way you use them.
- Your constraints. Time, money, health, law, equipment. Whatever cannot move.
- How it really happens. The actual sequence, not the tidy version.
- What good looks like. Measurable wherever you can measure it.
- What goes wrong. The honest failure modes, written down without flinching.
- What you decided, and why. So you stop relitigating the same argument.
A business layers customers, pricing, suppliers and licensing on top of that spine. A student layers courses, rotations and exam dates. The spine of an AI knowledge base never changes, though. Only the skin over it does.
Do it now, in nine steps
Your first AI knowledge base does not need seventy files written by hand. Here is the way I actually do it, start to finish. The first four steps take an evening. Steps five and six are the ones that decide whether the result is any good.
Make a folder and start talking
Name it after the business. Everything from here lands in it, and having one place to put things is most of what organised means.
Open ChatGPT, switch to voice, and describe your business for ten minutes as if you were explaining it to somebody who just walked in. Then ask it for a deep research report on what you said.
Speaking beats typing for the same reason recording a call beats taking notes. You say things out loud that you would never sit down and write.
Download the report and name it
my-new-business-project.md. Use that exact name every time. A fixed filename becomes muscle memory, and the prompt can then refer to it directly.
Gather what your AI knowledge base needs
It holds the empty knowledge base, the intake forms, and
structure-for-ai.md, which is the whole structure flattened into one file so you can actually upload it.Open
intake-forms/00-source-inventory.mdfrom the pack. It asks what actually exists: whether there is a website and what state it is in, whether the Google Business Profile is claimed, which social accounts are live and which were abandoned in 2021, where the reviews are, what systems the business runs on.Most of it you can answer yourself in an hour without asking anybody. Mark what you cannot determine as “do not know yet” rather than guessing. A gap you can see is worth more than a plausible invention.
Collect the things the inventory listed. Save your main pages as text: About, Services, Contact. Copy recent posts from whichever social accounts are actually live. Export a few months of queries from Search Console if you have it. Drop in whatever documents already exist, price list, safety notes, standard terms.
Everything goes in the folder from step 1. You are not organising it yet, only gathering it.
Hand the whole pile to Claude
Upload three things:
structure-for-ai.md, yourmy-new-business-project.md, and what you gathered in step 6.This is why the pack ships a single flattened file. Seventy separate uploads is not something most tools will take.
It tells Claude what to fill, in what order, and what to do when your material does not answer something.
You have three things. structure-for-ai.md is the empty knowledge base, every file of it in one document. Every {{PLACEHOLDER}} in it is a blank to fill. my-new-business-project.md is a research report about my business, in my own words. The rest is what I gathered: pages from my own site, my documents, and what customers already say about me. Read all of it before you write anything. Then fill the structure from the other two, one section at a time. Start with the glossary, then identity, then what we sell, then who buys. Everything else after that. Three rules while you work. Use my words, not yours. If my own pages call something a “rough in”, the files say “rough in”. Do not substitute a more common term because it reads more smoothly. Where my material does not tell you something, write NOT YET ANSWERED and one line naming what you would need to know. Do not fill a gap with something plausible. A blank I can see is useful. An invention I cannot spot is not. Replace {{CUSTOMER_WORD}} and {{THE_WORK}} everywhere first, before anything else. When you finish, list every section you marked NOT YET ANSWERED, then tell me which intake form in the pack would close each one.
Then argue with it
Ask five questions only an insider could answer. What a specific customer type actually worries about. What work you refuse and why. How the job most commonly goes wrong.
Specific answers mean it worked. Fluent but general answers mean a section is thin, and the hedging tells you which. Treat fluent and general as a failure, not a pass.
The prompt ends by listing what it could not answer and naming the intake form that closes each gap. That list is not a failure, it is your next hour of work.
After that, add a file whenever you notice yourself explaining the same thing twice. That is the whole maintenance rule. Once you are running more than one of these at a time, my notes on CLAUDE.md for projects cover keeping them separate without duplicating everything.
Download the whole thing, free. The empty knowledge base with every section explained, eleven intake forms that produce the answers, a discovery call script, and a twelve phase checklist. About seventy files, nothing filled in, no email required.
What one of these files actually looks like
People imagine something technical. Really every file in an AI knowledge base is just a document with headings, and here is a genuine fragment from a glossary a plumbing company might write.
Rough in
First fix. Pipework installed before walls close up.
Never call it "the first stage" in front of a builder.
Snag
A small defect found at handover. Ours, not the customer's.
A snag is free to fix. A change is not.
Callback
Returning to finished work. We track these monthly,
because the rate tells us more than any review does.Notice what is happening there. It defines the term, and it also carries the judgment around the term, which is the part that never makes it into official documentation. Consequently the AI learns not only what a snag is, but that calling something a snag has a cost attached.
Write yours the same way. Definition first, then whatever you would tell a new starter in the second sentence. For applying this across whole projects rather than one subject, my markdown workflow for projects covers that ground.
Start with your own words
If you do one thing from this article, do this one. Above all, write down the words you use, and what they mean when you use them.
Every field has them. A plumber says rough in. Over on a dairy farm it is somatic cell count, while a nursing student says SBAR and a machine shop says first article. None of those mean anything to an outsider. Worse still, some perfectly ordinary words carry a very specific meaning inside your world that they simply do not carry outside it.
Consequently, an AI that does not know your words writes brochure language. Everybody who does the work can tell instantly, because the vocabulary is slightly off in a way that reads as an outsider guessing. Meanwhile an AI that does know them sounds like it has been around a while.
This costs you twenty minutes and changes everything downstream. For that reason I always build the glossary first, before anything else in the folder.
If you are new to structuring context this way, my guide on how to create a CLAUDE.md file covers the root file that ties the rest together.
Say it out loud first
People will tell you things in conversation that they would never write down. Naturally that holds for a client, and it holds just as much for you talking through your own goal into a voice recorder.
So record the conversation. Then transcribe it within a day, while you can still repair the garbled words from memory.
In a business setting, the first call is easily the most valuable hour of the whole build. Somebody will mention the customer type they quietly refuse, or the real reason prices went up last year, or a word for part of their work that appears nowhere on their website. None of that survives note taking, because while you write the first thing down they are already saying the second.
Therefore put the pen away and let the recording do the capturing. Your job during that hour is to notice, and noticing is impossible while you are transcribing by hand.
Afterwards, pull the vocabulary out first, before you touch anything else. That is the same sequence in our client onboarding process, where the brain gets built before anything else does.
The test almost nobody runs
You will finish your AI knowledge base and it will feel done. However, feeling done and being done are different things, and there is a twenty minute test that separates them.
To begin with, ask the AI five questions that only an insider could answer.
Not what does this business do, because it can answer that from the first file. Instead ask what a specific customer type actually worries about. Then find out what work you refuse and why. After that, push on how the job most commonly goes wrong, and on what your busiest month is and what breaks during it. Finally, ask it to explain part of your work using your own vocabulary.
Studying rather than working? The same test applies. Ask it which topic you keep failing and why, and what your instructor actually rewards.
Specific answers mean you built it properly. Conversely, fluent but general answers mean a file is thin, and the hedging tells you exactly which one. Go back, fill that file in, and ask again.
Treat fluent and general as a failure rather than a pass. That single habit is the difference between a system people trust and a folder that quietly gets abandoned in month three.
Everything before execute
I run a rule in my own Claude ecosystem where nothing irreversible happens until I type the word execute. Questions are free. Drafts are free. Plans are free. Action waits for one specific word.
In other words, everything in this article happens before that word. It is the unglamorous part, and it decides whether the executable task is worth asking for at all.
Because here is the thing about an AI knowledge base. It is worth the least on the day you finish it, which sounds like a criticism and is actually the whole point. Every conversation after that gets folded back in. Every mistake teaches it something. By month six it knows things you could not have told me in month one, simply because it has been paying attention continuously and no human does that.
By contrast, most systems peak at delivery and then decline. This one starts small and compounds, and that is exactly why it earns the effort. It is also why I would rather hire someone who already lives in the ecosystem than someone with a general AI background.
Frequently asked questions
A usable first version takes an afternoon. Four files get you most of the benefit: your vocabulary, the subject, your goal, and your constraints. Everything after that adds depth.
One long document forces the AI to read everything to answer anything. Small files let it find the two paragraphs that matter. Just as important, you can update one part without opening the rest.
No. The same eight part spine works for a nursing program, a diet, a class or a renovation. Business versions simply add customers, money and compliance on top.
Not at all. These are plain text files with headings, written in ordinary sentences. If you can write an email you can write these.
Say so in one line rather than deleting the file. A clear statement that you hold no licences and need none is useful information. A missing file is simply ambiguous, so the AI will either guess or hedge.
Name one owner and set a review rhythm. Update anything that changes the week it changes, then fold every new conversation back in. Without a named owner it is already abandoned.
Any of them. I use Claude and my files follow its conventions, so my notes on CLAUDE.md for projects go deeper there. Still, the structure itself is plain text and carries over anywhere.
Where to go from here
The pack above is deliberately industry neutral, so a plumber, a dairy farm and a dental practice all fit the same shape. Strip the business specific files and the same structure carries a degree, a diet or a renovation without any trouble.
If you would rather I built one with you, that is a large part of my AI business automation work, and it sits alongside the WordPress development side of what I do. Either way, get in touch and tell me what you are trying to teach it.
Author: Shane Clark
