Summary
Most people hit the same wall with AI: it gets slower, forgets things, and starts confidently telling you things that are wrong. The instinct is to blame the model. It's almost always the data underneath it.
The setup that causes it looks like this — a project folder stuffed with markdown files, client notes, meeting notes, a file called important-stuff.md. Every question you ask becomes a search for a needle in a haystack, and once the pile gets big enough you stop controlling what the AI reads. It grabs whatever it guesses is relevant, you can't see what it grabbed, and you can't ask anything precise. Built-in memory doesn't fix it either: memory is a diary, not a database. It remembers vibes, and you can't query a vibe.
The fix is a split. Facts go in a database, notes stay in markdown. A database has a schema — this column is a date, this one is a number, this one can never be empty — and anything that doesn't fit the rules doesn't get in. That's the entire difference. Markdown accepts anything you type, which is why AI loves writing it and why it falls apart at scale. Nothing stops the same client's name being written three ways in three files.
This walkthrough goes end to end: connect Supabase to Claude through the official connector (OAuth, about 30 seconds, no API keys), point Claude at your existing markdown mess and have it propose a schema, approve the plan as a business owner rather than an engineer, and migrate. Your original files are read, never edited. From there the daily habit is a single question — is this a fact or is this a thought? Facts go in the brain, thoughts stay in the notebook.
One honest caveat included in the video: a schema stops structural garbage — duplicates, missing fields, wrong formats — but it does not stop a wrong fact. Tell it a lie in the right format and it will store that lie beautifully. The rules make your data consistent. You still have to make it true.
What you walk away with
Markdown has no rules; SQL is nothing but rules. That's the whole concept — everything else is setup.
Facts belong in the database, thoughts belong in markdown. When you're unsure which one something is, ask Claude where it should live.
The Supabase connector in Claude is OAuth now. No API keys, no config files, about 30 seconds.
You never write SQL. Claude writes every line; your only job is reading the proposed schema like a business owner and deciding whether it matches how your company actually works.
Start with four tables that matter, not twenty. Tables have to earn their existence.
A schema enforces consistency, not truth. Wrong facts in the right format still get stored.
Chapters
- 0:00Intro: it's not you, it's your data
- 1:20The problem: why markdown breaks at scale
- 3:31The solution: what SQL actually is
- 7:14Setup: connect Supabase to Claude in minutes
- 10:40Your first tables: markdown chaos becomes clean rows
- 13:33Daily use: faster, sharper answers every day
- 16:28Maintenance: make it compound, not rot
- 18:00Outro: SQL is the brain, markdown is the sticky note
Lines worth keeping
“Memory is a diary, not a database. It remembers vibes, and you can't query a vibe.”
“Markdown has no rules. SQL is nothing but rules. That's it. That is the whole lecture.”
“A schema stops structural garbage, but it doesn't stop a wrong fact. Tell it a lie in the right format, and it'll store that lie beautifully.”
“The database is the brain, and the markdowns are just the sticky note.”
Mentioned in the video
The hosted Postgres used throughout. Free tier is half a gigabyte, which is hundreds of thousands of notes if you're storing text.
The Supabase connector lives under Settings → Connectors. Supabase is free; the connector needs a paid Claude plan.
What Supabase runs under the hood. You can run it on your own laptop for free if you'd rather not use a host.
Full transcript
The whole video in text, split by chapter. Lightly repaired for caption errors, otherwise exactly what was said.
Intro: it's not you, it's your data
If you clicked on this video, I'm guessing you've been working with AI for quite a while now. For personal projects, for your business, or messing around trying to create the next Silicon Valley unicorn app. But, you've hit a problem. Your AI is getting slower, it's forgetting things, it's confidently telling you stuff that's just outdated or wrong. And the Google Drive, ClickUp, Notion, all of those plugins are not cutting it anymore. They're slow and they are not reliable. But, you're smart enough to realize that it is a data and organization problem. One that you can solve instead of just writing this whole thing off as, "Yeah, I tried AI. It doesn't work for me." If what I just described it sounds like the situation that you are in, grab your popcorn, fire up Claude, and let's get right into this tutorial. And just a quick disclaimer, I never finished college, I don't have a software engineering background, no technical background at all. Like, I studied business management for 2 years in college, then dropped out. The point being, you don't need to be Einstein to follow along. If you're like myself and you like learning and trying new things and getting faster, better, stronger at what you do, then this will be perfect for you because I'm telling you this way of working with AI and just having a data-driven approach to business in general has been such a game-changer.
The problem: why markdown breaks at scale
The problem. Let's talk about what's actually going on. Right now, your setup probably looks something like this. A Claude project with tons of markdown files, client notes, meeting notes, that one file called important stuff to markdown. And every time you ask AI a question, you are asking it to find a needle in a haystack. Here's the part that nobody tells you. Once a project gets big, you don't control what the AI reads anymore. At least, it's very hard to guide it in that direction when there's so much slop, aka unstructured data. It just grabs whatever it guesses is relevant. You can't see what it grabbed, you can't steer it, and you definitely can't ask granular questions like which of my clients hasn't paid this month. You can't trust the answer because the files live in six different places written in six different ways. Some of you are thinking, "Well, doesn't Claude have a built-in memory?" It does, and memory is great at what it does. But, memory is a diary, not a database. It remembers vibes, and you can't query a vibe. You can't ask a diary every invoice over $500. It would be sorting through so much information, and it would be very unreliable and not efficient. Look, every business that makes great decisions runs on data, not feelings, not I think we're doing pretty good this month. And the best part is you already have the data. It's just scattered across tons of documents in different places where nobody, including your AI, can actually make use of it. So, here's the model I want you to install in your head. It's the whole video in one picture. A core data layer, one structured place where your facts live, and markdowns get demoted to what it's actually good at, quick notes and instructions that tell your AI how to navigate the structured data. So, we got facts in the database, notes in the docs. So, every piece of data has its own lane. By the end of the video, my goal is to have your brain thinking in if-then statements rather than vibes. And that's what it really comes down to. We are going from a very probabilistic and assumptions-based model to something that is much more deterministic that we can actually rely on.
The solution: what SQL actually is
So, let's get to the solution. So, what's the structured place? A database. And the language our database is speaking is called SQL. So, you've maybe heard names like SQLite, MySQL, Postgres. It's different flavors, but it's all the same idea. And here's the only thing you actually need to understand about SQL. Are you ready? Markdown has no rules, so think docs. Markdown has no rules, so think docs, unstructured data. SQL is nothing but rules. That's it. That is the whole lecture. A markdown file literally accepts anything you type into it. That's why AI loves writing markdown. It physically cannot mess up, but that's also why it falls apart. Nothing stops your AI from writing the same client's name three different ways in three different files. Nothing stops a due date from being next Tuesday in one doc and the 14th in another. A database has a schema, so rules. This table has these columns. This column is a date. This one is a number. This one can never be empty. If something doesn't fit in the rules, it doesn't get in. So, let me just show you. Here's Claude logging a new client into markdown, and look, it made a second file for a client that already exists. The phone number's gone, and the format matches nothing else in the folder. So, it's not Claude's fault, but there were no rules to follow. Now, this same request into the database, one row, every field where it belongs, and the duplicate never happens because the rules won't let it. Now, to be fair, because if I don't say it, the nerds in the comments are going to attack me, a schema stops structural garbage, duplicates, missing fields, wrong formats, but it doesn't stop a wrong fact. Tell it a lie in the right format, and it'll store that lie beautifully. So, the rules make your data consistent. You still have to make it true. So, where does this database live? Honestly, anywhere you want. You can download Postgres right now, run it on your laptop, totally free, no cloud involved. And if that's you, go for it, because that is what Supabase, which is the SQL provider that we're going to use, they are using PostgreSQL under the hood. So, we're essentially just paying to run that same software on their servers. But if you're like myself, and you'd rather not keep a laptop open 24/7 and have Wi-Fi connected, that is where Supabase comes in. So, it's our SQL database hosted and always there, which unlocks my favorite thing. I'm out somewhere, I pull out my phone, and I ask Claude, "How much does a client owe me?" or something like that, and it just answers because it has the Supabase connector through Claude. So, my entire company data in my pocket always. The free tier gives you half a gigabyte of storage. That sounds really small until you remember that text is really small as well. That's hundreds of thousands of notes. So, why would I advise just storing text to start? And then if you really need photo, video, or other file storage, then you might need to upgrade to one of their higher plans later on. But even still, I'd encourage you to just use Supabase or PostgreSQL for storing text-related things. One really minor thing, if you have a project on Supabase and you don't use it for a few weeks, they actually will just pause your instance, but I'm sure it's relatively easy to just get it back up and going again. You might be thinking now, okay, are there other providers in which I can host this database through? Well, yes, you could do this on Google Cloud. Um you could do your taxes via fax, but Supabase is honestly just really easy to use. I've tried doing SQL instances through Google Cloud. It's possible, but their whole setup is a lot more involved.
Setup: connect Supabase to Claude in minutes
So, I just recommend stick with Supabase and let's get into setup. Let's build this thing. A quick note before we start, I'm going to do this in the Claude desktop app. Terminal person, everything still applies here. I just didn't want to scare anyone off using the terminal. And if you don't know what I'm even talking about when it comes to a terminal, then just don't worry about it. Step one, we are going to head to Supabase here. Let's just click sign in. If you don't already have an account, just create an account. And if you do not have an organization created, which I'm guessing you do not, you're just going to go here and click new organization. And you know what? I'll just go through the steps with you so there are no guessing games. Plan, beautiful, create new organization. Right here it's going to ask about a GitHub repository. So, what this means is we're going to have the ability to implement version control so that changes to our database are actually logged and that we can go back to specific commits is what it's called. But, it's something I would not worry about right now and we can get to later. So, let's just make sure this is all good. We're going to type in a database password here and we are going to enable row-level security. I always like to keep things as secure as possible and then we are going to create a new project. And for any reason if you're feeling hesitant right now, this is just a sandbox. If you end up not wanting to use this, it's not like you are signing your soul over to, you know, Supabase. You know, you can always go back. We're not doing anything crazy here. So, I just wanted to get that sigh of relief out of the way before we continue. Set up a Claude project stuffed with exactly the kind of messy markdowns we all have. That's on purpose. It's about to become the star of the show. Okay, step two. And this used to be the painful part. So, API keys, configuration files, an afternoon of Stack Overflow. Now, watch. In Claude, we're going to go to settings, connectors, find Supabase, hit connect. It's going to bounce you to the browser. Just click authorize and done. So, it took about 30 seconds. No API key, no configuration file. That is typically stuff that you have to do with other softwares, but they make it really easy for Supabase. So, if you're bracing for the scary, really technical part of this video, it's over. It's done. I would say that is the most of it. Okay, so we are in Claude here and what we're actually going to do is head over to settings, connectors. As you can see we already have the Supabase connected, but if you do not, you would just search Supabase and then you would click connect. And as you can see there are certain things that have approval and don't have approval. It's totally up to you. And then just to test it out, I am going to type in here Hey. So as you can see, Claude is using that connector tool right now. Last step just to prove it works, we're just going to type in what tables do I have in my Supabase project. And there it is. An answer straight from the database, which is empty obviously. We're about to fix that. And just one sentence on the money side of things just so nobody feels ambushed. The Supabase side is free. The connector wants a paid Claude plan. And if you clicked on this video, odds are you're already on pro. Just thought I would mention it. So for most of you, this new brain doesn't cost anything extra.
Your first tables: markdown chaos becomes clean rows
Your first tables. So we are connected but empty. And this is where most tutorials leave you hanging with now design your schema or, you know, now do the thing. But I am going to walk through all of this with you. And we don't really need to design anything. It's not rocket science. So here's what we're going to do. You know your business, Claude knows databases. So stay in your lane and this will literally take 10 minutes. So here's my sample project, a fake little agency. And I'm betting this pile looks familiar. Client notes in one file, half an invoice tracker in another, a to-do list from March. And I'm actually guessing that you have a lot more junk. So let's clean this house. The prompt we're going to use is actually super simple. We're going to say look through all these files, propose how this should be structured in my Supabase project. Don't build anything yet, show me the plan first. That is all we're going to say. So look what comes back. A client's table, projects, invoices, notes. And now you have the only job you get in this entire process, which is read that plan like a business owner, not an engineer. Does this match how your company actually works? Do you call them clients or accounts? Can a project belong to two clients or just one? So, you don't need to check the SQL code, but just make those logical decisions of, you know, oh, how many projects typically fit under a client? Is it one or is it many? And then from there, you'll have a really strong foundation to build off of. And notice what happened to the original files. Absolutely nothing. They get read, never edited, never deleted unless you want to. So, once the plan looks right, just say the magic words, build it, and migrate my files in, and it'll do exactly that. And now the moment this whole video exists for, ask it something you could never trust your pile of information to answer. Look at that. That is not an AI vibing through 50 documents to find what you're looking for. That is a query. So, if you ask the same question tomorrow, you will get the same answer. If your stuff lives in spreadsheets rather than markdown, it's the same process. Just just get that CSV file, hand it to Claude, and it'll get it put in Supabase. Okay, two pieces of advice before we move on. Piece number one is start small. You don't need 20 tables, you need four that matter. Tables really have to earn their existence. Write a short directory file that it can use to navigate your database. Things like what tables exist and what data goes where. This little file is the map that makes every conversation in the future just instant when it's wanting to access, read, and write from your database. So, if you take anything away, remember the model. The database is the brain, markdown is the map. And maybe you have some other miscellaneous markdown files, that's okay, but I would say trying to structure your data is always going to be the better long-term play.
Daily use: faster, sharper answers every day
Daily use. So, what does this feel like day-to-day? One word, querying. And a query is just a question with rules. So, when you ask, "How much did I invoice in June?" Claude does not go read your documents. It writes a tiny piece of SQL code, sends it to the database, and gets back the exact rows that answer your question. It's nothing else. You never write SQL. I just want to be crystal clear about that because if that was the case, I would not be here doing this tutorial. You just supervise it, Claude writes it, and you probably notice by now I haven't written a single line of code in this entire video. So, everything we've talked about till now leads up to the question, "Okay, so how does this actually beat my current setup of messy markdown files and random spreadsheets?" It's speed and sharpness. A query pulls exactly what you asked for, but when you have a ton of random data, your AI is having to read through so much of that just to get an answer, and that might not even be reliable. Because I think people forget that AI is not a magical black box. It has tools. It has to read. It has to write. And it can't just magically ingest tons of data. You know, it actually has to scan through things. So, the smart approach to working with AI is figuring out, "Okay, how can I get my AI what it needs the fastest?" And on top of that, how can you make sure that the data your AI is receiving is factual? When you're doing the regular method of huge lake of random unstructured data, it is going to eat through your AI tokens much, much quicker, and your sweet, precious context window is going to get filled up so quickly. So, this is also why I'll tell you not to build your core data layer on something like Notion or Google Sheets. I've actually tried this before, and it's very bad. Because first off, your AI doesn't like reaching for those connectors because it's just inherently slow. And the other piece of it is that those are not SQL databases. I'm not sure what database system they're using under the hood, but it's genuinely really slow. So, I thought go big or go home, and that's why I went with pure SQL using Supabase. Now, the daily habit, and it's one question. Is this a fact or is this a thought? So, a fact would be something like an invoice, a deadline, a client's number. This type of stuff goes in the database. A thought, a rough idea, a draft, this stays prose. And what prose just means is unstructured. So, this would be something perfect for markdown. Recap, facts in the brain, thoughts in the notebook. And when you're not sure, just ask Claude, "Hey, where should this live?" It'll get it right basically every time. Oh, and when your notes really pile up, databases can even search by meaning rather than keywords. And it's called vector search. I'm not going to get into it this video, but it's extremely cool. And this is one of those long-term effects when you use something like SQL.
Maintenance: make it compound, not rot
Once you start thinking in structured data, you want to track everything. Because for the first time, tracking actually pays off. But two warnings to keep you honest. Supabase is a tool, not your free ticket to having Jarvis. I love that philosophy, but a tool cannot run your business at the end of the day. It makes your judgment and decision-making much faster and much more reasoned. Second, maintenance is real, but it's small. Just every month, every week, just take a few minutes to think about ways in which you could be optimizing the structure of your data. Maybe you add a few tables, maybe you remove a few tables, but the point is that we're keeping this core data layer very truthful and always trying to find those new data points that we can start to track. And you can literally just talk to Claude about it. So, ask it, "Hey, how can we improve our database?" It'll tell you what tables to keep, which ones to drop. Um and then you can just go through files, you know, maybe some markdown files, see what data in there is useful that you'd like to enrich into your database, and what stuff that you can just toss. It's like cleaning your house. If there's not a place for it, just throw it out. Like, the live laugh love thing you have sitting on your toilet that your mom just does not want to get rid of, you know it should not be in your house, you just throw that baby out. Don't let your brain rot. Let's keep it very healthy, meaning this is very truthful data, because at the end of the day, if we are going on this very data-centered play, then we have to make sure that it's very factual.
Outro: SQL is the brain, markdown is the sticky note
All right, let's close the loop. You clicked on this video with an AI that is slow, forgetful, and confidently wrong. And I hope you're leaving now knowing it was never the AI. It was the data underneath it. Structure the data, and the same AI gets smarter. So, here's just the entire architecture in one sentence. The database is the brain, and the markdowns are just the sticky note. Thank you so much for watching. I really appreciate it. I hope you learned a lot in this video, and let me know what types of video ideas you guys would like to see next. I'd love to keep diving further on this topic of structured data and working with AI in an efficient manner. So, let me know. Again, I really appreciate it, and I hope you have a good rest of your day.
Let's make something worth watching.
Tell us what you're trying to make. We'll come back with how we'd do it — same day, no pitch deck.