Teach Claude Everything It Doesn't Know About You

I counted the AI setups I've abandoned in the last two years.

Workflows that ran for a week and stopped. Projects with a skill I built once and never touched again. Sessions that started strong and faded because every time I opened a new chat, I was starting from zero.

Not because the tools were bad. Because the tools had no idea who I was.

Every conversation I've ever had with an AI started the same way. I explain the context. I explain my goals. I explain my format preferences, my constraints, my current project status. Then I get an output, close the tab, and do it all again tomorrow.

The agent knows my tasks. It doesn't know what I'm building toward. It doesn't know what I'd say no to. It doesn't know whether a given opportunity aligns with where I'm trying to go or pulls me off course.

An agent with no identity isn't an assistant. It's autocomplete with ambition.

Andrej Karpathy called this the shift that defines AI development now: you don't program a system by writing logic. You program it with data. Your identity, your goals, your decision framework - those are the code. And most people have never written a single line.

This guide is the setup. Three files that make Claude actually know you. A memory layer that compounds over time. And Claude in Chrome as the action layer that lets it move in the world filtered through your judgment, not just your instructions.

Most AI setups are technically complete and personally empty.

Skills. Connectors. Scheduled tasks. The tool layer is there. But the tool layer is serving nothing in particular, because nothing in the system knows what matters to the person using it.

Claude has your calendar if you've connected Google Calendar. It knows your task list if you've connected Notion. It can read your emails if you've toggled on Gmail.

What it doesn't have: why any of those things exist. What you're building toward. The decision criteria you use when two good options compete. The things you'd say no to even when they look attractive on paper.

Without that, Claude optimizes for the task in front of it. With it, Claude can optimize for your actual life.

The difference sounds abstract until you feel it. A meeting summary that notes the follow-up items isn't the same as a meeting summary that flags the one item that actually matters for the goal you're steering toward this quarter.

Same model. Completely different output. The only variable is whether Claude knows who it's working for.

Not twelve. Not a complex framework. Three documents that load into every Claude session and change what the model has to work with.

Who you are, how you think, what you're building toward.

Where it lives: click your initials in the bottom-left corner → Settings → Profile, for your global identity. Project Instructions for project-specific context.

What goes in it:

Keep it under one page. Specific is better than comprehensive.

One thing I added to mine that changed everything: the last section. What I say no to. Claude now flags when something conflicts with my stated priorities rather than just executing the task.

Your actual missions, not your task list. This is what the agent steers by.

Where it lives: click "Projects" in the left sidebar, open your working project (or create one). The knowledge base sits on the right side of the project's main page under Files, click the "+" there.

I update this file every quarter. It takes about 20 minutes.

What Claude can do without asking. What it always escalates. Where it writes.

Where it lives: same project, right-hand panel, under Instructions, click the "+" there.

Read wide, write narrow. Claude can access everything. It can propose anything. It writes to exactly one place, a folder you review before anything leaves it.

The three files are the foundation. Memory is what makes it compound.

Claude's memory system lives in Settings → Capabilities, under Memory. Two toggles matter:

"Search and reference chats", Claude can look back at previous conversations for context. Turn this on.

"Generate memory from chat history" (marked Legacy in the current UI), Claude extracts patterns, preferences, and recurring details from your conversations and stores them as memory items. Turn this on too, but audit it.

That second part is the key. Claude will store things automatically. Some of those things will be accurate and useful. Some will be approximate or irrelevant. The ongoing maintenance - reviewing what's been stored, deleting what's wrong, adding what's missing - is what separates a memory system that compounds from one that drifts.

My rule: once a week, open Settings → Capabilities → Memory and spend five minutes on it. Delete anything inaccurate. Add anything important that Claude didn't capture on its own. Over time, the memory list becomes a compressed version of you - the most important things about how you work, what you're building, and what you care about, distilled into a form Claude loads every session.

After three months of maintaining mine, I noticed something. I'd stopped explaining myself. Not because I stopped caring about the context - because the context was already there. Every session started already knowing what I was working on, what I was trying to avoid, what kind of output would actually be useful.

That's the compounding effect. It's slow at first and then it's everywhere.

Here's where the identity layer meets the world.

Claude in Chrome is a browser extension that gives Claude the ability to read, navigate, click, and fill forms in real time alongside you. It requires a paid plan, Pro, Max, Team, or Enterprise, free accounts can't use it.

Setup takes about five minutes:

Here's the part that changes when you've done the work in the previous sections:

A Claude in Chrome with no identity context is a browser automation tool. Useful, but dumb. It does what you tell it.

A Claude in Chrome loaded with your identity file, your goals file, and your trust boundaries is something different. It researches filtered through what you actually care about. It surfaces information relevant to where you're going. It acts the way you would if you had more hours - because it knows what you'd do with them.

Three setups worth building:

Click the clock icon in the upper right of the extension panel. Set the cadence and the time. Write the task:

The brief that comes back isn't generic news. It's filtered through your goals. It surfaces what matters to you, skips what doesn't, and flags anything that needs your attention - all before you've opened your laptop.

I walk Alfie three times a day, usually starting around 7am. The brief is ready by the time I'm back. Six months ago that was the first thing I'd spend 30 minutes doing at my desk. Now it's two minutes to read and move on.

Before a purchase, a hire, a commitment, or a choice between two options, give Claude in Chrome a research task filtered through your decision criteria:

The research comes back structured around your actual decision framework, not around what seems generally relevant. That's a different kind of output - and a much faster path to a decision you'll actually stand behind.

Any task that requires knowing your background - filling a bio, writing an application, completing a profile, drafting outreach - becomes dramatically faster when Claude in Chrome has the identity file loaded.

Claude reads the page, fills what it can from what it knows about you, and stops where it needs your judgment. The form that would take 20 minutes of looking up your own information takes three.

This is what the full system looks like running.

Morning: The brief is already there. Two minutes to read, one minute to flag anything that needs attention today. The identity and goals files mean it's already filtered - you're not sorting through noise.

During the day: Every Claude session inside your project loads your identity, your goals, your trust file. Every output is filtered through where you're going, not just what you asked. You stop explaining yourself.

When a decision comes up: Claude in Chrome gathers the research. Your decision criteria are already in the system. You get a structured recommendation, not a wall of information to sort through yourself.

End of week: Five minutes in Settings → Memory. Delete what drifted. Add what the week taught you. The system gets slightly more accurate. Over months, that compounds into something that feels like working with someone who has been your thinking partner for years.

The agent doesn't become powerful because it can think. It becomes powerful because it finally knows what you want it to think toward.

If this changes how you think about your setup, share it with one person still explaining themselves from scratch every morning.

That's all I'm asking.

Become a paying subscriber to get access to this post and all other premium content.