Give your AI your rules, and watch it answer like you

Jay WestOut of My Head · about 14 min

By the end of this step you have:

A "How I decide" note — your answers, reasons and exceptions in one document you own, the one document this course builds on
The note set where your AI reads it every time, with a standing instruction that hands back what is yours instead of guessing
A before and after on the same test question, scored, the first time it sounds like you

The idea

No Question Log yet? Step 1, "Your last five questions", starts it, or take your last five from memory; the note works without it.

An AI does not know your business. Ask it whether to discount a client and it answers for the average business on the internet. It only has the context you give it.

So give it the context once. Your log already holds it. Gathered into one note, it lets the AI answer from your rules; given raw scraps, it guesses.

Keep the note in a document you own — a Google Doc, Word, Notion, Obsidian — so it moves with you when the AI you use changes.

An example

The owner of a two-location dental practice keeps being asked by the front desk whether to waive a missed-appointment fee. Asked cold, ChatGPT gives her a paragraph about goodwill that could apply to a gym.

She adds her note to a project — waive it the first time, never for a second miss inside a year, always for a documented emergency — and asks again. The answer names the first-time rule and asks whether this is a second miss, the question she would have asked.

Do this now

Ten minutes. Make a document called "How I decide" (for a team, or new in the job? "How we do it here"). Started the note in "One place your answers live", or already keep a CLAUDE.md, project instructions or a rules file? Add only what is missing there. Press "Copy what is settled" below and paste it in. Under it, three lines — what you sell, who your best customers are, and this year's goals in order of priority (an internal team? who you serve, and what they need from you this year; several businesses? one line each, and which comes first). Then a "Where things are" section at the top — your prices, schedule, who does what, the facts people ask for most; facts need no because. Then give it to your AI where it reads it every time — in ChatGPT or Claude, a Project; in Copilot or Gemini, a pinned file or a Gem; on a free plan, paste the note at the top of the chat. Someone else's company? Use the AI it approved, never a personal free plan; nothing approved yet, keep the note as a shared document and do the AI parts by hand. Made a project in "Your first AI, set up in ten minutes"? Replace what you pasted there with this note. Building a custom GPT? The note goes in Knowledge and the instruction in Instructions; "Write instructions your AI cannot misread" has the full shape. Add the standing instruction — "Answer from my note and the files in this project; where they disagree, the note wins. If they do not cover it, say so and do not guess. Anything about pricing exceptions, hiring or firing, or a legal matter, hand back to me."

Test it on the AI you already use.

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

Your test question: one a person on your team really asks

Testing an instruction, a prompt or a brief instead of an answer?

Paste the instruction and one real input. Run the copy in a new chat, then run your old instruction on the same input in another, and mark which one you would send.

Test it

Ask your test question inside the project, then in a fresh chat with no note. Paste both answers in the box, mark each, and tick "without my note" for the fresh one. The gap is what lived only in your head.

If you already have documents, or years of notes

Already keep project instructions or a rules file? Put it first in each Project, or in the CLAUDE.md or AGENTS.md every session reads, and add the becauses and the hand-back instruction there. A second brain of notes and emails is the source, not the note: run the prompt below on it and keep only rules you confirm.

Add the three documents your team opens most under your note. Keep each kind in its own file, named by kind (CASE STUDY, PROSPECT CALL), with the label repeated atop every section — a project or custom GPT reads pieces, not whole files. Add to the instruction, "Name the kind of each source, and never cite a prospect call as delivered work."

Years of notes, sent replies, tickets or session transcripts (keep the file and timestamp beside each rule)? Paste five of one kind under this prompt and confirm each rule before it goes in the note — "Below are cases one person in my business handled. Write the rule you think each one follows, in one line. Group the cases that share a rule, and write each rule as: the rule, when it applies, the exception, the case numbers. Mark a rule shown by one case as weak. Add no rules the cases do not show."

If your answers live in a changing Google Sheet, PDFs and Word files, recorded calls, revenue totals, or everything your company holds

Your answers live in a Google Sheet that changes? If your AI can connect to Google Drive and read the sheet, connect it so it reads the sheet itself. A custom GPT, Gem or project built from an uploaded file holds a copy — re-upload it on a set day each week, and put that day in the note.
A folder of PDFs, Word files and sheets? Upload them as files in the shared project, or into NotebookLM, which reads PDFs and Google Docs as they are. Your note goes on top so the AI knows which source wins; for totals, see below.
Recorded sessions or talks? Transcribe each one (many meeting and phone tools can), keep one transcript per file, named by date, client code and theme, beside the video link, and add the folder to NotebookLM. It answers with the passage, and the transcript's timestamp takes you to the moment. Your own lookups are the log: each time you go looking for something, log what you looked for and where it was (me | what I looked for | where I found it). A week of them is the ten questions to test the notebook with. Confidential client sessions are personal details, and those never go into a chat as they are: get the client's consent first, or keep the transcripts on your own machine (Ollama with Open WebUI, or AnythingLLM).
Totals, such as revenue? Upload the CSV or sheet into a chat whose AI can run analysis on files (ChatGPT and Claude can on most plans); a tool that reads files only as text, such as NotebookLM or a vault-chat plugin, will not add them up reliably. Numbers kept as properties on notes in a vault? Point Claude Code or Cursor at the vault folder and ask it to total them, since it runs the sums as code, or copy a Dataview or Bases table out as a CSV and upload that.
Everything your company holds, searchable across its tools? Start with the AI your suite already has, if your plan includes it (Copilot in Microsoft 365, Gemini in Google Workspace, Atlassian Rovo, Slack AI), or a cross-tool search such as Glean, and trial it against a week of logged lookups: the one that finds those wins. Want the decisions written down too, with access rules? That is the Decision Week, and five days of this log is its spec.

If your answers live in an Obsidian or Roam vault, a repo, a Notion teamspace, SharePoint or tickets

Asking across a whole collection — a repo, an Obsidian or Roam vault, a Notion teamspace, a SharePoint library, tickets or ITGlue? That is search: Notion's own AI search, or your AI's Notion connector, for a teamspace; Claude Code or Cursor for a repo or vault folder, NotebookLM for a Markdown export (a Roam or Obsidian export is one file per page, so merge the export into one file per year first (a free notebook takes a limited number of sources), then add those), a SharePoint agent built on the library (check your Copilot licenses), the platform's own AI for tickets. Give the tool your note first, so it knows which source wins and when to ask ("if two SOPs match, ask which system"); for a repo, put it in CLAUDE.md or AGENTS.md.

If your team asks it too, you roll connectors out to a team, or you move Projects to a company workspace

Your team asks it too? A shared project needs a team plan (ChatGPT Business, Claude Team); on a free plan, share the note. If the team edits it in Notion or Google Docs, connect your AI to that page (Claude and ChatGPT have connectors for both) rather than pasting a copy — the document is the original.
Rolling connectors out to a team? If your plan lets an admin add them once for the whole organization, add them there, so nobody pastes keys or config. No admin route, or the vendor needs a key per person? One person pulls the numbers on a schedule into a shared Google Sheet, and everyone's AI reads that sheet through its Google Drive connector. Not sure which connectors are worth it? Log a week of the team's lookups by hand; the top three kinds are the ones worth the setup. Not the admin? Send the admin the week's lookup tally ("Copy kinds and counts only") with the three connectors it names.
Moving work Projects to a company workspace? Rebuild each from documents the company owns, run your test set in both, and leave the personal ones behind.

If it must stay on your own machines, or you use the API or Azure OpenAI

Must stay on your own machines? Ollama with Open WebUI, or AnythingLLM, runs the note as the system prompt and can index a vault or repo folder beside it. The setup, in AnythingLLM: install it on one machine (the desktop app) or one in-house server (its Docker version, which takes several users); choose Ollama as the model provider and pull a model; make a workspace, paste the How I decide note into its system prompt, and add the vault or repo folder as its documents. On the server, add your team as users.
The API or Azure OpenAI? The note is the system message. A note of a few pages fits in any current model's instructions, so it needs no database or Redis. Retrieval (Azure AI Search, or "on your data") is only for the document library. Hiring a developer? Give them the note, Instruction v2 from "Write instructions your AI cannot misread" and Test set one from "Test your AI before you trust it". That is the spec, and the test set is how you accept the work.

Hand it over

Send the side-by-side to one person — a partner, your ops lead, another owner.

Where it goes wrong

Pasting everything you have and expecting the AI to sort it out. Keep personal details out — the note is 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-0359, L-0161, L-0358, L-0047)

Your first step · step 1 is next

Next: step 1, which the rest build on →