Give the crew the script.
AI agents give confident wrong answers when they cannot see your company's facts. This module shows you how to build a shared knowledge base they can rely on, one approved file at a time. No data team needed.
- Production
- The context gap
- Director
- You
- Crew
- AI
- Take
- 01
- Scene
- 05
- Part
- II
By the end of this module you will be able to:
- Explain why agents give confident wrong answers, and what context engineering means.
- Name the five kinds of knowledge an agent needs, and how to test each.
- Build a first knowledge base in two weeks, starting from real questions.
- Keep it accurate with five habits, and check it is ready before go-live.
This module follows the guide section by section. Keep it open as a one-page reference while you work.
Why agents sound sure and still get it wrong
An agent that does not know your services, prices or policies fills the gaps with plausible guesses. It sounds sure of itself either way. The fix for most wrong answers is better context: approved facts the agent can find and trust.
Context engineering means deciding what the agent should know, writing it down once in plain files and keeping it current. Anthropic's engineering team uses the term for the discipline of choosing what information a model works from. You do not need a data team to start. You need an owner, a clear structure and a regular review.
In a June 2026 survey of 101 enterprises, 57% had traced an incorrect AI agent answer to missing or inconsistent business context. Only 25% had a governed context layer in production.
VentureBeat, VB Pulse survey, June 2026Five kinds of knowledge an agent needs
Every agent that speaks for your company needs the same five kinds of knowledge. Each comes with a simple test.
| Kind | Includes | Test it |
|---|---|---|
| Company factsFacts | Services, products, locations, history and approved figures. | Every number has a source and a date. |
| Voice and styleVoice | Tone, words to use and avoid, and strong examples. | Ask it to rewrite a sample in house style. |
| LimitsLimits | What the agent must never claim, share or promise. | Ask a question it should decline. |
| How work gets doneProcess | Steps, templates and who approves what. | Ask it to walk through a task step by step. |
| Where to lookSources | An index that routes each question to the right file. | Ask which file the answer came from. |
| The owner | Knows which files are current, and who approves changes. | "What did the agent get wrong this week, and which file fixes it?" |
A pile of documents versus a knowledge base
The most common first attempt is to upload everything. It feels thorough. It gives the agent old and new versions side by side and no way to tell which one is right.
| A pile of documents | A knowledge base |
|---|---|
| Every document uploaded at once | A small set of approved files to start |
| Old and new versions side by side | One current version, dated |
| Long documents on many topics | One topic per file, with a clear title |
| Anyone can change anything | Named owners and defined permissions |
| Answers with no source | Answers that name the file they used |
Write it down once. Every agent benefits.
Build the first version in two weeks
Start with the questions people ask most, not with every file you own.
- QuestionsList the 20 questions staff and clients ask most often.
- FilesWrite one short file per topic that answers them.
- IndexAdd an index that says which file covers what.
- TestAsk all 20 questions. Fix the file behind every miss.
- OwnGive each file an owner and a review date.
A workable rhythm over ten working days: two days to collect the questions, five days to write the files, one day for the index and two days to test, fix and assign owners. Plain text or Markdown files work well because people and AI can both read them.
Title: Managed backup service Owner: Service lead Last reviewed: (date) What it is: two or three plain sentences. Who it is for: the client types it suits. Approved figures: each with its source and date. Never say: claims, promises or names that are off limits. Related files: pricing, onboarding process.
Keep it accurate, then connect
A knowledge base is accurate on the day you finish it. These five habits keep it that way. Count the ones in place. 0 to 1: start with the index and owners. 2 to 3: add a weekly error review. 4 to 5: open it to another team.
- Log wrong answers. Each miss points to a gap.Try it: keep a list of answers that missed.
- Fix the file. Correct the source, then retest.Try it: "Which file should have covered this?"
- Date everything. Old facts cause confident errors.Try it: add a last-reviewed date to each file.
- Name owners. Someone answers for each topic.Try it: assign an owner to every file.
- Keep limits current. New rules go in when they change.Try it: "What must the agent never say?"
Before go-live
- Top questions are covered.
- Every file has an owner.
- Limits are written down.
- Answers name their source.
Expand when
- Test answers are consistently right. Add the next 20 questions.
- Owners update without reminders. The routine is working.
- People ask it first. It has earned trust.
- Another team asks. Connect their agent to the same base.
Keep a person involved when
- Answers go to clients. A person approves first.
- Prices or numbers appear. Check them against the source.
- Private data is nearby. Set who can read and who can write.
Be fair and clear.
A knowledge base is a set of written rules: what is true, what is off limits and who decides. Writing them down is an act of fairness. Everyone, people and agents, works from the same facts, and nobody is caught out by a rule that lived only in someone's head.
Fair also means consistent. When a limit applies, it applies to every agent and every team the same way. When a file is wrong, you fix it in the open and say which answers it affected.
- State your limits plainly in one file anyone can read.
- Give every file a named owner, so questions go to a person, not a void.
- When an answer was wrong, log it, fix the source and tell the people who relied on it.
A shared knowledge base behind every agent
I co-built a company's AI knowledge base and wrote its agent identity and tone-of-voice standard. It gives the company's AI agents and Claude projects one approved picture of the business, and each approved update strengthens every system connected to it.
- Files
- 86 interlinked Markdown files, one topic each
- Routing
- Query routing points each question to the right file
- Voice
- A tone-of-voice standard every agent follows
- Access
- Defined permissions: some projects read, some write back approved updates
Start your knowledge base
Draft the first pieces of your starter knowledge base. You can finish the files over the next two weeks.
Check your understanding.
Module 5 in four lines.
- Wrong answers usually come from missing context, not a weak model.
- Five kinds of knowledge: facts, voice, limits, process and sources.
- Start from the 20 questions people ask most. One topic per file, dated, with an owner.
- Agents draft, people decide. A person approves anything that goes to a client.
List the 20 questions people ask your team most often. That list is the first day of the build.
Further reading on my site
Pass the knowledge check to complete this module.