My AI Forgot Everything Every Morning. So I Gave It a Rulebook.

A simple setup for doing serious research with AI, one that remembers what you learned, keeps track of who said what, and gets a little better every day. No coding needed.

The problem

A while ago I started researching a problem space I did not know well. If you have done this, you know what it involves. You need to understand how the space works, who the existing players are, what has changed recently, and most importantly, you need to talk to a lot of people who live in that world every day.

I did what most people do. I opened a chat with an AI and started asking questions. The answers were good. So I kept going, one chat after another.

Two weeks in, I had a mess.

• I had dozens of chat windows and could not remember which one had the useful thing in it.
• Every new chat started from zero. The AI did not know which companies I had already looked at, who I had already spoken to, or what I had already ruled out.
• My call notes were in one place, my competitor list in another, and the links between them lived only in my head.
• When three different people told me roughly the same thing, I only noticed by accident.

The AI was not the problem. The problem was that I was using it like a very smart stranger, when what I needed was a research assistant with a memory and a set of house rules.

So I built one. This post walks through the setup, from the simplest version you can start with today to the layers I added as my research grew. You can copy it for almost anything: researching a market, a job search, a thesis, even a big purchase.

The big picture

The whole setup has three parts.

1. A folder of notes. Plain text files, one note per thing you are researching: a person, a company, a conversation. I use a free app called Obsidian to view and connect them.
2. Claude, working inside that same folder. It reads and writes the notes directly, so nothing lives in a chat window that will disappear.
3. Instructions that Claude reads every time. One file that explains how I work, and a few small “skills”, which are recipes for jobs I do again and again.

You look at the notes. Claude maintains them. The instructions make sure it maintains them the way you would.

The rest of this post builds this up one layer at a time, then points to what you can add later.

The case study

To keep it concrete, I will use a made-up example throughout.

Say I want to understand small independent fitness studios: the local yoga, pilates and boxing studios with one or two locations. I want to know how they run day to day, what software they already pay for, what has changed for them recently, and what frustrates them enough that they would pay to fix it.

That means reading, mapping competitors, and talking to a lot of studio owners and instructors. Exactly the kind of research that falls apart in chat windows.

Layer 0: One folder, two tools

What you need:

• Obsidian (free, obsidian.md). Create a new “vault”. A vault is just a normal folder on your computer. Call it something like fitness-research.
• Claude Code. This is the version of Claude that can work with files on your computer. You can use it from the Claude desktop app (the Code tab) or from a terminal (the text window where you type commands). Point it at the same folder.

That is the entire trick. Both tools look at the same files. When Claude writes a note, it shows up in Obsidian a second later. When you edit a note in Obsidian, Claude sees your edit the next time it looks.

Nothing is locked in either tool. If you stop using one of them tomorrow, you still have a folder of readable text files.

Layer 1: Think in things, and the relations between them

This is the most important idea in the whole post, and it has nothing to do with AI.

A chat is organised by time. What you asked first is at the top. But research is not about time. It is about things and how they relate to each other.

For my fitness studio research, the things are:

Just three. Start with three.

Now the part that makes it work: the relations.

A person had a conversation. A conversation mentions a company. A person uses a company’s product. A person introduced me to another person.

In Obsidian, a relation is just a link. You write the name of another note inside double square brackets, like [[Maya]], and it becomes clickable. Here is what the conversation note looks like:

---
type: conversation
person: "[[Maya]]"
date: 2026-03-03
---

# Call with Maya, 3 March

Maya runs a two-room pilates studio. She uses [[Booking App A]] for
class bookings and a spreadsheet for instructor pay.

She said the worst part of her week is Sunday night, when she
works out what to pay each instructor. It takes her about three hours.

She offered to introduce me to [[Leo]], who runs a boxing gym nearby.

Why not folders? It is tempting to make a folder for each company and put the conversations inside. Do not. One conversation can mention three companies. One person can talk to you twice. A folder can only put a note in one place, but a link can connect it to as many things as it relates to.

What Obsidian gives you for free:

• Backlinks. Open the Booking App A note and Obsidian shows every note that links to it. So without doing anything extra, I can see every conversation where a studio owner mentioned Booking App A.
• Graph view. A visual map of every note and every link. It is great for spotting that one person is connected to everything, or that a company nobody talks about is sitting all alone.

Layer 2: Let Claude keep the links tidy

Links are powerful, but keeping them correct by hand is boring and easy to get wrong. This is exactly the job I gave to Claude.

When I finish a call, I paste my rough notes and say “log this”. Claude then:

1. Creates the conversation note.
2. Checks whether a note for that person already exists. If not, it creates one.
3. Links the conversation to the person, and the person back to the conversation.
4. Creates a company note for any company mentioned that I have not seen before.
5. Tells me what it created and what it linked, so I can check.

Every so often I also ask it to check for broken links, meaning a link to a note that does not exist, or a person who is mentioned in a conversation but has no note of their own. It takes seconds and keeps the whole thing trustworthy.

But for Claude to do this the same way every time, it needs to know how I like things done. That is the next layer.

Layer 3: The rulebook, AGENTS.md

Every time you start a new session, Claude starts fresh. It is like hiring a very smart assistant every morning who has never seen your work before.

So I wrote them an onboarding document.

It is a single file called AGENTS.md, sitting at the top of the folder. The AI reads it at the start of every session, before doing anything else.

Why that name? AGENTS.md is an open, shared convention. Most AI assistants that work with files, including ChatGPT’s Codex, Cursor, GitHub Copilot and Gemini, look for it. Claude Code has its own file for the same job, called CLAUDE.md, and it also reads AGENTS.md when there is no CLAUDE.md in the folder. So one file named AGENTS.md works with almost any tool, and if you switch tools later, your rules come with you.

(If your version of Claude Code is older and does not pick it up, create a CLAUDE.md file whose only line is @AGENTS.md. That tells Claude to read the other file.)

The file does not hold my research. It holds how I work: what the project is, where each kind of note goes, how to link them, and the handful of rules that must never be broken. Here is a simple version for the fitness studio project.

# Fitness studio research: how to work in this folder

## What this is
I am researching small independent fitness studios: how they run,
what software they pay for, and what frustrates them.

## The notes
- people/        one note per person. Name the file after the person.
- companies/     one note per company.
- conversations/ one note per conversation, named date-person,
                 e.g. 2026-03-03-maya.md

Every conversation links to its person. Every person note lists
their conversations. Always link both ways.

## Rules
- Never invent anything a person did not say. If you are unsure,
  write "unclear" and ask me.
- Keep quotes word for word. Keep my paraphrases clearly marked
  as paraphrases.
- If I described my own idea to someone before they described
  their problem, note it. Their enthusiasm after that point
  counts for less.
- Before creating a company note, check one does not already exist
  under a slightly different name.
- Every fact about a company (price, funding, launch date) needs
  a source link.
- Write in plain sentences. No jargon.

Two tips that helped me a lot:

• Keep it short at first. Ten rules you actually care about beat fifty you copied from somewhere.
• Keep “how I work” separate from “where things stand”. I have a second file, README.md, with my current status and a dated list of decisions (“15 March: decided to focus on studios with under 200 members”). I only ever add to that list, never edit it. That way I cannot quietly re-argue something I already settled three weeks ago.

Layer 4: Skills, your recipes for repeated jobs

The rulebook tells Claude how to behave in general. The second piece of the setup is for the jobs you do again and again, like logging a call, which have the same steps every time. Instead of explaining those steps in every session, you write them down once, as a skill, and after that you only need to name the job.

A skill is a small instruction file that describes one job, step by step. It lives in a folder called .claude/skills/, with one folder per skill and a file inside called SKILL.md. Claude sees the list of skills you have and uses the right one when your request matches.

My log-conversation skill, for example, asks me for anything missing, writes the conversation note, creates or updates the person note, links every company mentioned, and ends by listing every note it touched. Here it is, simplified.

---
name: log-conversation
description: Turn rough notes from a call or chat with a person
  into a conversation note, and create or update their person
  note. Use when I say "log this call" or paste call notes.
---

# Log a conversation

1. Ask me for anything missing: who it was, the date, how we met.
2. Create conversations/<date>-<name>.md with:
   - What they do, in two or three sentences.
   - What frustrates them, using their words where possible.
   - What they already pay for, and roughly how much.
   - Anyone they offered to introduce me to.
3. Create or update people/<name>.md. Add a link to this
   conversation under "Conversations". If they offered an
   introduction, add it under "Intros offered".
4. For every company mentioned, link to its note. Create the note
   if it does not exist yet, with just the name and a one-line
   description.
5. At the end, show me a short list of every note you created
   or changed.

The difference between a skill and the rulebook:

• AGENTS.md is always on. It applies to everything.
• A skill is for one specific job. It only comes into play when that job comes up.

How to write your first one: do the job with Claude by hand once or twice. When it comes out the way you like, say “turn what we just did into a skill”. Then read what it wrote and fix anything that is off.

Putting it together

That is the core setup. Four pieces, and you can have all of them running in an afternoon:

Start there. Use it for a couple of weeks before adding anything else.

Further reading: layers to add later

Once the core is working, your research will start asking for more. These are the layers I added, roughly in order. None of them is required. Add one only when you feel the need for it.

Tables from properties. Every note can carry a few labels at the top, such as status: talked or city: Austin. Obsidian’s built-in Bases feature turns these into live tables, set up by clicking, with no code. I use one for “people I still need to contact” and one for “introductions I have not followed up on”. Tell Claude the exact allowed values in AGENTS.md, because a note labelled Status: Talked instead of status: talked quietly disappears from the table.

A new kind of note when a pattern appears. After a dozen conversations I noticed several people describing the same frustration in different words. That pattern did not belong to any one conversation, so I added an insight note: a one-line claim, linked to every conversation that backs it up. The rule I added with it is that one enthusiastic person is not a trend. An insight only counts as strong when several different people said it without me prompting them.

More skills. Each time a job started repeating, it became a skill. For example, one finds the companies already trying to solve an insight, one collects what changed in the space this week, and one researches a person and drafts a message to them. They all write into the same linked notes, which is why the setup stays connected instead of turning into separate piles. Claude drafts, I send. Nothing goes out under my name without me.

Every mistake becomes a rule. This one is a habit, not a feature, and it is why the setup keeps getting better. When Claude gets something wrong, do not just fix the output. Ask why it happened, and update AGENTS.md or the skill so it cannot happen the same way again. For example, I added “check the date on the page itself, never on the search result”, and a small tone file built from the edits I made to its drafts. One caution: a rule written in a file can still be skipped when things are rushed. For anything that truly must not happen, like sending a message, take the ability away instead of asking nicely.

Using it for something else

The fitness studio example is just one shape. The same setup works for almost any research that involves people, options and evidence:

Pick your three starting things. Link them. Write down how you want them handled. Let Claude do the tidying. Add a layer only when you feel the need for it.

What it does not do

I want to be honest about this part.

This setup does not do the thinking for me. It cannot have the conversations, and it should not decide what matters. What it does is make sure that when I do talk to people, nothing they say gets lost, nothing gets exaggerated, and every new session starts from everything I already know instead of from zero.

It also keeps me honest in a way I did not expect. The mistake I worry about most in research is not lying. It is optimism: remembering a compliment as real interest, or counting one excited person as a trend. A system that makes me link the evidence for every claim makes that much harder to do by accident.

If you try it, start small. Three kinds of notes, one rules file, one skill. You will know what to add next.