Two documents, one handover
The previous lesson established why an agent needs a knowledge base. This lesson is about building the actual documents and connecting them. There are two: a standing-facts document (the vault entry for this task) and a brief (the instruction the agent reads each time it runs). They are kept separate because the facts stay constant across many runs, while the brief can be adjusted for a particular execution without touching the facts.
Together, they form the complete handover. The agent reads the facts first, then the brief. When both are well written, you can hand the task to the agent without an introductory explanation each time. That is the practical payoff of everything in this section of the course.
The interview pattern
Most people find it difficult to write a standing-facts document by sitting down and listing what they know. The easier route is an interview: ask yourself a short sequence of questions, write the answers, then organise what comes out.
The questions that reliably surface what the agent needs are these:
- What is the task, in one sentence? Not what the output looks like, but what the task actually is.
- What must always be true of a correct result? The non-negotiable standards, in your field’s terms.
- What varies from run to run? The inputs that change each time and the decisions the agent must not make on its own.
- What has gone wrong before? Past failures are often the most durable standing facts, because they capture the edge cases that general training cannot anticipate.
- What does the agent never touch? Scope boundaries stated explicitly are less likely to be crossed quietly.
Write the answers in plain language. Do not compress them into bullet fragments. A sentence the agent can read without ambiguity is more useful than a neat list it might misread.
Turning the answers into a document
Once you have answered the questions, group the answers under four short headings: What this task is, What a correct result looks like, What varies, and What is out of scope. That framework is the standing-facts document. It does not need to be long. Two pages of clear sentences outperforms ten pages of compressed notes, because the agent reads it in full and acts on what it reads.
The brief is shorter still. It has four parts: a pointer to the standing-facts document, the current state (what the agent is receiving this run), what you need (the specific output), and any constraints particular to this execution. The brief changes; the standing-facts document changes only when your understanding of the task changes.
The monthly financial-document run
Consider the monthly task of processing a set of financial documents: receipts, invoices, statements, all arriving in different formats, all needing to be matched, categorised, and reported accurately. The standing-facts document for this task would capture what a correct categorisation looks like in your context, which suppliers appear every month and how their invoices are formatted, what the agent must flag for human review rather than decide on its own, and what falls outside this run entirely. The brief for any given month would point to that document, describe the specific batch of documents being processed, and state the output required: a reconciled summary, a list of anomalies, a report ready to file.
A person picking up this task for the first time, reading both documents, would understand the job. So does the agent. That is the test: if a competent colleague could execute the task from those two documents alone, the documents are ready.
What you can do now
Take a task you run regularly. Work through the five interview questions. Write the answers in full sentences and organise them under the four headings. Then write the brief: current state, what you need, any constraints for this run. Keep both documents somewhere you can load them at the start of every execution. From this point, every lesson in the course builds on the assumption that these two documents exist and are in good shape.