Teach it once

Where this fits

This is the third piece in a set. Before you start: settings and plans covers setup. Putting AI to work covers the doing: prompts, artifacts, connectors.

This one is about making it hold together. What the brain actually is under the hood, how to find the conversation you had three weeks ago, how to keep the whole thing from turning into a junk drawer, and how to build a skill properly instead of hoping.

Read this after you have used Claude for a few weeks. It will not land before then.

Part 1: The brain, actually explained

People say "brain" and "knowledge base" and "project" like they are the same thing. They overlap, but knowing the difference is what stops you being confused later.

The four things feeding Claude context

At any moment Claude is working from four different sources. They behave differently and most frustration comes from not knowing which one is failing you.

1. Project instructions. A box you write once inside a project. Always loaded, every single chat in that project, whether you need it or not. This is the job description.

2. Project knowledge (the files). The documents you upload into a project. Claude reaches into these when your question calls for them. This is the reference library.

3. Memory. What Claude picks up about you as you go, across conversations. Working separately from your files.

4. Chat search. Claude going back and reading your old conversations when you ask it to.

Instructions are always on. Files are pulled when relevant. Memory builds passively. Search happens when asked.

So what is "the brain"?

The brain is instructions plus files, inside one project. That is it. There is no button for it. It is just a name for what you have loaded into a project.

When someone says "build your company brain," they mean: put enough into a project that Claude stops needing to be told the basics.

The thing nobody tells you about knowledge files

Claude does not read your whole file every time. It pulls the parts that look relevant to what you asked.

Which means how you write the file changes how well it gets used:

• Headings matter enormously. A file broken into clear labeled sections gets used far better than the same content as one long block.
• Long documents lose their middle. A ninety-minute transcript pasted raw is close to useless. The beginning and end come through. The middle gets thin.
• Cut before you upload. Ten tight pages beat sixty sprawling ones. Always.

If a file is long, do this before uploading it: paste it into a chat and say "reorganize this into clear sections with headings, cut anything repetitive, and keep every specific detail." Upload that instead.

Your chats do not talk to each other

This is the single most common surprise. Two chats inside the same project do not share what happened in them. They each see the instructions and the files. They do not see each other.

So if you work something out in a conversation and you want it to persist, it has to leave that conversation. Either into a knowledge file, or into memory, or you have to go find it again.

The habit worth building: at the end of a good session, say "summarize what we figured out here as a document I can add to my project knowledge." Then add it. That is how a brain actually grows.

Part 2: Finding your way back

You will have hundreds of conversations. Here is how not to lose them.

Chat search

You do not go hunting through a list. You just ask:

What did we work out about the Henderson pricing?

Find the conversation where we wrote the intake questions.

Let's pick up where we left off on the newsletter.

Claude goes and reads your old chats and pulls what it finds into the conversation you are in now.

Turn it on properly. Settings, then Memory, then "Search and reference chats." Without it, searching only looks at conversation titles, which means anything you did not name well is effectively gone.

This one is paid plans only.

A scoping detail worth knowing: searching inside a project stays inside that project. Searching outside covers your non-project chats. So if you cannot find something, check whether you are standing in the right place.

Memory

Separate from search. Memory is what Claude quietly learns about you as you work.

Settings, then Memory, then "Generate memory from chats." On by default for personal accounts.

Three things worth knowing:

• You can put things in on purpose. Say "remember that I never work Fridays" and it saves it.
• Each project has its own memory space. What it learns in your client project stays there.
• You have a pause and a reset. Pause keeps what is there and stops adding. Reset deletes everything permanently and does not come back.

Name your chats

Claude names them for you and the names are fine, not good. Thirty seconds renaming the ones that mattered will save you an hour later. Rename the conversation the moment you realize it produced something you will want again.

Part 3: Staying organized

The junk drawer problem is real and it arrives around month two.

One project per area of your life or work, not per task. Household. Your parents. A big project at work. If you are making a new project every time you have a new question, none of them will ever get smart.

New chat for a new task, new project for a new area. Chats are cheap. Start a fresh one whenever the subject genuinely changes. A conversation that has wandered through four topics gives worse answers than a clean one.

Pin the two or three you live in. They move to the top of your sidebar.

Prune your knowledge files quarterly. Old prices, a doctor you no longer see, a school schedule from last year. Claude believes what you gave it. A file you forgot to remove will confidently repeat last year's information forever. Put it on the calendar once a quarter, open each project, and ask yourself what in here is no longer true.

Keep a "what changed" file. One document in each project, updated when something material shifts. It is the cheapest way to keep a brain current without redoing it.

Part 4: Skills, done properly

Putting AI to work introduced skills. This is the real version.

What a skill actually is

Anthropic describes skills as folders of instructions, scripts and resources that Claude loads when a task calls for them.

Note the word folder. A skill is a folder containing at minimum one document called SKILL.md, and optionally examples, templates and small scripts.

The cleanest way to hold the distinction:

A project says "here is what you need to know."
A skill says "here is how to do this thing."

A project is a room you work in. A skill is a tool you carry between rooms. Your skills are available in every project, all the time. Your project knowledge only exists in that one project.

Compared with a saved prompt you keep in a notes file, a skill triggers itself. You do not have to remember to go get it.

The anatomy

A minimal skill is genuinely this small:

---
name: weekly-briefing
description: Builds a one-page look-ahead of the coming week from calendar events, school and caregiving deadlines, and open to-dos. Use when the user asks for a weekly briefing, Sunday look-ahead, week plan, or what's coming up.
---

# Weekly Briefing

[your instructions here]

Two required fields. name has to be lowercase with hyphens, under 64 characters, and it has to match the folder name. description has a length limit too, so keep it to a few sentences.

Everything below the second --- is your instructions, in plain English. Keep the whole thing under about 500 lines. If it gets longer, split the extra into separate files in the same folder and link to them.

The single most important thing on this page

The description is the entire trigger.

Claude does not read your skill to decide whether to use it. It reads only the name and description, and matches those against what you asked for. The body only loads after that decision is made.

Which means the most common failure is this: someone writes a beautiful, careful skill and gives it a lazy description. The skill never fires. They conclude skills do not work.

The formula: [what it does]. Use when [the words a person would actually say].

A good description names the job and lists the words people use for it. A bad one says "Helps with documents" or "Does stuff with files."

Write it in third person. "Builds a weekly briefing," not "I can help you build a weekly briefing."

Stuff the trigger words in. If you would say briefing, look-ahead, week plan or schedule to describe the same job, put all of them in the description.

How to tell whether it actually fired

Watch for "Using [skill name]" as Claude works.

This is the whole debugging technique. If you do not see it, the skill did not trigger, and the problem is almost always the description.

Before anything works: one setting

Settings, then Capabilities, then turn on "Code execution and file creation."

Skills will not work without it. If yours seem to have vanished, check this first.

Then find your skills under Customize in the left sidebar. Anthropic has moved this menu more than once, so if Customize is not there, look under Settings for Capabilities or Features.

Three ways to build one

1. Talk it into existence. The best route for most people, and it is officially supported.

Start a normal chat. Say what you want. Upload your templates, your past examples, the two outputs that were exactly right. Claude will interview you, then build the folder and hand you back a file you can install.

I want to build a skill for my Sunday look-ahead. I'm going to paste in the last three I was happy with. Ask me questions until you understand what makes a good one, then build the skill.

2. Reverse-engineer a session that worked. Finish something you are happy with, then say "turn what we just did into a skill so it runs the same way every time." You already know what good looks like, because you just made it.

3. Write the file yourself. A text editor and the lines above. No terminal needed.

Installing one

From the directory: Customize, the + button, the Skills tab, Browse skills, Install. That is the easy path and it is the one we would use.

From a file, watch this one thing, because it is the most common failure and it has nothing to do with AI:

Zip the folder, not the contents.

Right-click the folder itself and compress it. If you open the folder, select the files inside and compress those, you get a broken skill and a confusing error. On a Mac it is one wrong click apart.

Refining it

This is where most people stop, and it is where the value is.

Write your tests before you write the skill. Before building, write down three real things you will ask it to do, and what a good answer looks like for each. Then build. Then run those three. You now know whether it works instead of guessing.

Use two Claudes. It takes two browser tabs.

Claude A helps you build the skill. Then you test it on Claude B, a completely fresh conversation with the skill installed. Bring the failures back to A.

Why the fresh one matters: Claude A has the whole conversation in its head. It will appear to succeed for reasons that have nothing to do with your skill. Only a stranger tells you the truth.

Expect the first version to be about half right. Correct it, tighten the instructions, run your three tests again. A skill that gets used every week is worth five rounds of this. A skill you built once and never refined is a skill you will stop using.

Do the job by hand four or five times first. A skill built before you understand the job just automates your uncertainty.

Part 5: Finding skills other people built

Where to look, in order

1. The directory inside Claude. Customize, the plus button, the Skills tab. Curated by Anthropic. This is where a beginner should start and most people never need to leave it.

2. Anthropic's own repository: github.com/anthropics/skills. The official collection. Anthropic says these are for demonstration and learning, so test one before you rely on it.

3. Partner skills. Companies that built official ones. Canva and Zapier are the two most useful for everyday life and small business.

4. Claude Academy at academy.claude.com. Free official training.

5. Community collections on GitHub. Search "awesome claude skills" and you will find several large lists. Some are genuinely useful. Read the next section before you install anything from one.

The safety part, which matters more than it sounds

Anthropic warns plainly that a malicious skill can direct Claude to run code, reach sensitive files or send data elsewhere, and that installing one deserves the same care as installing software.

Think of a skill as closer to an app than a document.

Two things it can do that a document cannot: run code, and give Claude instructions you did not write. That second one is the sneaky part. A malicious skill does not need any code at all. It can just contain a sentence telling Claude to do something you would not want, and Claude will read it as instructions.

Five rules for a non-technical person:

1. Stick to the in-product directory and Anthropic's repository unless you have a real reason not to. Almost everything you need is already there.
2. Open the SKILL.md and read it before installing. It is plain English. This is the one kind of software you can actually inspect yourself. If there is a section you do not understand, do not install it.
3. A skill with no scripts folder is much lower risk. Instructions only. Look at what is in the folder.
4. Ask Claude to check it. Paste the SKILL.md into a fresh chat: "What does this skill do, and is there anything in it I should be worried about?" It takes ten seconds.
5. Never install one containing a password or an API key. Even the author's own. It tells you everything about their judgment.

And one on value, since it comes up: if somebody is selling you a skill, check whether it is sitting in Anthropic's public repository for free first.

The short version

• The brain is instructions plus files, in one project. Nothing more mystical than that.
• Claude reads the parts of your files that look relevant. Headings and brevity do real work.
• Chats in a project do not see each other. If it matters, get it out of the chat and into a file.
• Turn on chat search. Name your conversations. Prune quarterly.
• One project per area of your life. Not per question.
• On a skill, the description is the whole trigger. Write it for the words a person would say.
• Watch for "Using [skill name]" to know it fired.
• Zip the folder, not the contents.
• Write your three tests before you build. Test on a fresh Claude.
• Read a skill before you install it. It is English, and you are allowed to.

When you get stuck

Ask Claude, in plain words, including "explain that like I have never done this before." It is better at explaining itself than any guide is, including this one.

Settings, menu names and plan rules were accurate as of October 2026 and Anthropic changes them often. If a screen does not match this guide, trust the screen.