Let the AI answer first, and check only what it hands back

Jay WestOut of My Head · about 13 min

By the end of this step you have:

Your rules where the people who ask can reach them without you — a shared AI project, a team GPT, or the note itself
The "ask it first, then me" rule sent, and the line it must never cross written into its instructions
The first count of questions answered without you, beside the ones that still reached you
A fifteen-minute weekly check that replaces reading everything, and a rule for every wrong answer

The idea

If you have not started your Question Log yet, step 1, "Your last five questions", starts it, and "Give your AI your rules" makes the note this step shares. No note yet? Start a document called "How I decide" now; the steps add to it. No log yet? Take your last five questions from memory.

A written answer works while you are not in the room. It only pays off if the people who ask can reach it. The AI is the fastest way for them to reach it, because they ask in their own words and get your rule back — and it needs what a new hire needs, which is what to do, what not to do, and when to stop and get you.

Then check the exceptions, not everything. Toyota built looms that stopped themselves when a thread broke, so one person could watch many machines and look only at the ones that stopped. Handing something over and never looking again is hoping; reading every answer puts you back in the middle.

An example

A twelve-person property management company sets up one shared Claude project for the office, with the owner's note and five rule cards. The instruction — answer from these rules only; a rent reduction, a legal notice, or anything the rules do not cover goes to Dana.

In the first week the office manager asks it about late-fee waivers, pet deposits and a contractor's invoice. It answers two from the rules and sends the invoice to Dana, correctly. On Friday Dana reads only what it handed back and what still reached her — and the invoice question becomes a rule the same day.

Do this now

Ten minutes. Pick where the asking will happen. On free accounts, or where the company has not approved AI, share the "How I decide" note itself as a document — that works today. Sharing an AI project needs a paid team plan (ChatGPT Team or Business, Claude Team); a custom GPT, a Gem, or an engineering team's CLAUDE.md or AGENTS.md also work. If customers do the asking, it is the AI you draft replies with. Add your note and rule cards and the standing instruction from "Give your AI your rules", with your never-delegate list as the hand-back topics. Then, below, log anything this week that was decided without you, beside what still reached you.

Log what was decided without you this week, beside what still reached you.

Your Question Log: day 1 of 5 · 0 questions reached you · 0 rules written

Test it

Ask it three questions as your team or a customer would — one a card covers, one right on a limit, one from your never-delegate list. It should answer the first, flag the second, and hand the third back to you. If it answers the third, tighten the instruction.

Hand it over

Send one message — "Before you ask me, ask this. If it says it is mine, bring it to me." Use the first month's log as the case for a team plan.

If most of your log is about people — leave, conflicts, pay — give the shared AI only the rule cards for those, written by role ("a team member asks for unpaid leave"), never names or cases.

If two people's agents are drifting apart, you built an agent the team cannot reach, or the team's AI cannot reach your CRM

Two people's agents drifting apart, each with its own context? Keep the shared note in a git repo both clone; each agent proposes changes on its own branch, and a person merges them. The full route is under the two-agents heading of "One place your answers live, kept current by your log".

Built an agent the team cannot reach? Split it. The note and the rule cards go where the team asks, because any surface can read a document. The actions — the mailbox, the CRM, the scripts — stay with the agent where its tools are. The shared record lives in a system both already reach, such as your CRM, not in a chat.

If the team's AI cannot reach the CRM either, make it reachable from a hosted place, not your machine — the CRM's own connector, if it has one, or a remote MCP server every surface can call. Until then, the agent writes to a shared sheet both surfaces already open.

If you are moving the team's Projects to a company workspace, or AI must stay on your own machines

Moving work Projects from a personal account to a company workspace? If your tool has no way to move one Project at a time, recreate each Project's instructions and files in the company workspace from documents the company owns, rerun the five questions from "Test your AI before you trust it" in both, and leave the personal ones where they are.

Must stay on your own machines? Ollama with Open WebUI, or AnythingLLM, runs the note and can index your vault folder beside it; "Give your AI your rules" has the setup, step by step, under its own-machines heading.

If you work alone and your customers do the asking, nobody asks you, or your AI sessions or agents do the asking

Working alone? The people asking are your customers, and the free route comes first: your note open beside your inbox, and the AI drafting each reply from it for you to send. The check is the same count — how many replies went out without a rewrite.

Nobody asks you? You are the one asking. Put the note and your idea pages in one project, and ask it what you decided about the thing in front of you. Your AI sessions or agents do the asking? Every question one brings you is a row, and the check counts what they settled from the note without you.

Every Friday, fifteen minutes

Three lines — answered without you, still reached you, anything wrong. Anything that reached you twice gets a rule. Any wrong answer gets its rule fixed and a date on it. Silence is not a check; if nobody logs what was decided without you, you cannot tell a working system from one nobody uses.

Where it goes wrong

Putting personal details about customers or staff, passwords or anything under a confidentiality agreement into an AI tool your company has not approved. Keep the note to how you decide, not who it was about.

This step is written for this course from the research in the full curriculum. The research behind it, on howtothink.ai (L-0532, L-0525, L-0401, L-0527, L-1707)

Your first step · step 1 is next

Next: step 1, which the rest build on →