Dropping an AI agent into a business without documentation is the modern version of hiring someone and never telling them what the job is. You will get motion, not progress. Maybe not even motion.
Start with the general overview
You cannot fully document something you have not used yet. So the first pass is a general overview - your best guesses about the system, the user, the data. This document is not the final spec; it is the raw material the final spec gets built from.
For this stage, I run the document through Claude. It is genuinely good at turning rough notes into structured documentation. Use it.
The 5 questions the brief has to answer
- What are the recurring tasks we do over and over?
- Which processes eat the most human supervision time?
- Which decisions must stay with a human, no matter what?
- Where does the data come from, and where does it go?
- How do we measure success?
If you cannot answer those five questions, do not deploy the agent yet. Build the documentation first. The agent is the easy part.
The order is: docs, then automation
The teams I have seen succeed with AI automation follow the same order every time: write the brief, run it past the model, refine it with the people who actually do the work, then deploy. The teams that try to skip to deployment end up rebuilding the documentation six months in, twice as expensive, with half the trust.
Documentation is the spec. The spec is the system. If the spec is wrong, the system is wrong. AI does not fix a bad spec - it just runs a bad spec faster.
