On this pageJump to a section
Ask any AI assistant to "write a follow-up email to a lead." You will get something polite and forgettable.
Now give it your offer, your price, how you actually talk, the three objections you hear every week and the one promise you will never make. Ask again. Same model. A different universe of output. The difference is not intelligence. It is context, and context is the part of this you own.
Why does context matter more than the model?
Because the model knows the world and nothing about you.
It can write well in general. It cannot know that you stopped selling a package, that your refund rule has conditions or that your clients dislike exclamation points. Every time you leave that out, it fills the gap with something plausible, and plausible is the dangerous kind of wrong.
What the builders say
Anthropic's engineering team describes context engineering as the set of strategies for curating and maintaining the optimal set of tokens during model use, which goes beyond writing a prompt. It also says context is a finite resource with diminishing marginal returns, and that models can lose accuracy as the window fills, which it calls context rot.1
a finite resource with diminishing marginal returns
That has a plain consequence for a small business. Dumping every document you own into a tool is not the strategy. Choosing what it sees is. (The groundwork is in Get Your Data Ready for AI Before You Automate Anything.)
A different Anthropic conversation, posted on February 13, 2025, adds a twist to the moat idea. One of the team put it this way: "If the models get smarter, all of our moat's gonna disappear. That means you're building the wrong thing."3 I read that as a reason to invest in your context and your process, which get better when the model does, and not in anything a smarter model will simply do for free.
VideoAnthropic
Tips for building AI agents
Three Anthropic people talk through workflows versus agents, tool design and how to measure results.
What goes in the knowledge base?
Six parts. Each is a folder or a clearly named file.
- Offers. What you sell, who it is for, what is included, what is excluded, the price and the terms. One file per offer.
- Procedures. Your SOPs: how you onboard, deliver, follow up and report. Written as steps, not essays, the way SOPs for AI agents describes.
- Voice. How you sound. Include a short description, ten sentences you would say and five you would never say.
- Answers. The questions customers ask most, with your approved answers.
- Proof. Results you can document, with dates, context and what was typical. Only things you are allowed to publish.
- Boundaries. What the tool must never claim, never promise and never do without asking. Your AI policy is the parent document.
Start with boundaries
If you build only one piece first, build the last one. A short list of rules like "never state a result unless it is in the proof file" prevents the worst failures, and it costs an hour.
Voice is examples, not adjectives
Telling a model to be "warm and direct" produces generic warm and direct. Showing it ten real sentences of yours produces you. Collect messages you are proud of and label what makes them work.
How should you structure it so AI can use it?
Think like a good librarian.
One topic per file
A file called "everything about us" forces the tool to read all of it every time. Ten short files with clear titles let a tool or a person pull only what the task needs. Anthropic's guidance on agents points the same way: keep lightweight pointers and load detailed information at the moment it is needed, rather than packing everything in up front.1
Put the answer first, then the detail
Open each file with a two-sentence summary. Follow it with the details, the exceptions and the date it was last confirmed. If someone, human or machine, reads only the first lines, they should still get the right answer.
Tell it what is current
Add "last reviewed" and "owner" to every file. An out-of-date price in a knowledge base is worse than none, because the tool will repeat it with total confidence.
How do skills and agents change the picture?
They make the knowledge base modular. (Agent Skills and the End of the Mega-Prompt is the longer version.)
Anthropic announced Agent Skills in October 2025. It describes a skill as a folder of instructions, scripts and resources that Claude can use when relevant, and says the system loads only the minimal information and files needed for the task.2 The idea is not specific to one product. It is a pattern: package a procedure with the exact context it needs, and let the tool reach for it when the task matches.
A procedure and its context travel together
Instead of one giant document, think in packages. A "new client onboarding" package contains the steps, the welcome message, the intake form and the boundaries that apply. A "weekly report" package contains the sources, the formula for each number and an example of a good report. Each package is easy to test and easy to fix.
How do you keep it from rotting?
A knowledge base is a garden, not a statue.
| A knowledge base that rots | A knowledge base that works |
|---|---|
| Written once, never touched | Reviewed on a schedule with named owners |
| One huge file | Small files with clear titles |
| Claims with no source | Proof tied to dated records |
| Everyone has access to everything | Access matched to the tool and the task |
Review it when you learn something
Every time the AI gets something wrong, ask which file should have prevented it, and fix that file. The mistake is your feedback loop. After a few weeks, the errors you see change from "it doesn't know us" to "it needs one more rule."
- The AI gets it wrongNote the mistake
- Find the fileWhich one should have prevented it?
- Fix the fileAdd the fact or the rule
- RetestRun the same task again
Where I stand
When one person runs the business, as I have since 2024, the knowledge base is the nearest thing you have to a team handbook, and the agents are the team. Whoever writes the handbook decides how well the team performs.
Anyone can rent the same model. Nobody can rent your context.
Start small this week. Write the boundaries file and one offer file. Test them by asking a tool to draft a reply, then fix what is wrong. Your context is the moat, and you build it one file at a time. Building that handbook for a client's business is the heart of the systems I build for clients.
Keep building.Ekai
Questions people ask
What is an AI knowledge base for a business?
It is a set of plain, well-organized documents that tell an AI tool how your business works: what you sell, how you deliver, how you sound, what you answer and what you must never say. It is the context the tool reads before it writes or acts.
What is context engineering?
Anthropic describes it as strategies for curating and maintaining the optimal set of tokens given to a model, going beyond writing a single prompt. In practice it means choosing what the model sees, and keeping that set small and relevant.
How big should a knowledge base be?
As small as it can be while still covering how you work. Anthropic calls context a finite resource with diminishing returns, so favor short, focused files over one giant document.
What should never go into an AI knowledge base?
Passwords, payment details, government identifiers and sensitive client or patient information. Keep secrets and regulated data out, and write down which tools may see which documents.
Sources
- 1
- 2
- 3
