It's a rainy Sunday morning, coffee in hand, and you're doing something very adult: researching savings account interest rates. 

So you open Claude and type out a seemingly basic question: "What's the best interest rate I can get on a savings account right now?"

Claude answered and said something unexpected:

Knowledge cutoff of August 2025? That was nine months ago. There's been an entire winter, several Bank of Canada decisions, and a lot of life since then, and Claude has apparently been sitting in August waiting for pumpkin spice season to arrive.

Once you see this, some things start to make sense: the weird answers you've gotten, the questions that felt simple but somehow weren't, the moments where you figured you must have asked it wrong.

Classic "it's not you, it's me" situation. This is just how every AI model works. Nobody explained it to us, and there isn't exactly a disclaimer anywhere.

There's a lot of frustration about what AI tools actually deliver. It’s a genuinely powerful tool but let’s be real, it comes with some assembly required. Think of it like that Ikea table you bought. It looked incredible in the showroom. Got it home, opened the box, and suddenly there are seventeen steps, a screw that doesn't seem to belong anywhere and an allen key that hasn’t evolved, ever.

The assembly instructions nobody gave you? Here they are: 

1. It always defaults to average

AI was trained on the entire internet, including all the bad stuff. So when you ask it to make a slide deck with no further instructions, it draws from thousands of examples, good ones, mediocre ones, and of course bad ones, and the output is somewhere in the middle. It doesn't know what good looks like for you specifically. So it guesses.

🔨The fix: before you hit send, tell it who your audience is, what style you want, what brand assets you have, and what the goal is. The more you bring to the table, the less it has to guess.

2. It's frozen in time

As we saw with our Sunday morning savings rate experiment, every AI model has a knowledge cutoff. The training stopped at a certain point, sometimes six months ago, sometimes more, and everything that happened after that is a blind spot. It doesn't know what it doesn't know. It will answer like it has the full picture even when it doesn't.

🔨 The fix: turn on web search for anything time-sensitive, or paste in the current information yourself.

3. It's an overeager intern

You give it half a brief and it sprints off to start building before you've finished the sentence. It wants to impress you. And you end up with a finished product you have to throw away.

 🔨 The fix: tell it not to start yet. Literally type "do not start building until I give you the go-ahead." It works.

4. It hallucinates

It will state something completely wrong with total confidence, and because it sounds authoritative, you believe it. Dates, names, statistics, citations, it can get all of these wrong without breaking a sweat. This is getting better, but it still happens.

 🔨 The fix: verify anything that matters before you act on it.

5. Every new chat starts from zero

It doesn’t really know who you are, what you do, or what good work looks like for you. So if you're not telling it, it's guessing, and by now you know where guessing gets you.

 🔨 The fix: build a context document about yourself once, and attach it to every new chat. We'll get to exactly how to do this in the next section.

6. The longer you talk, the worse it gets

This one surprises people. If you use one single chat for every conversation, it starts to drift. It forgets the instructions you gave it at the start, goes off track, and sometimes delivers things half-finished. Researchers call this “context rot”.

 🔨 The fix: use multiple chats. Break your work into focused conversations. And get your most important work done early in the chat, before it starts to fade.

The solution across all six of these? Context. The more specific you are about who you are, what you need, and what good looks like, the less AI has to guess. And the less it guesses, the better your results.

You just learned that every new chat starts knowing absolutely nothing about you. So let's fix that tonight: get your AI to interview you. Ask it questions, one at a time, until it has enough to write you a context document you can reuse.

This is called a hot start document, and it's level one of actually getting good results.

Paste this into a new chat (not the same one you’ve been using for your meal prep planning!):

Here’s what mine looked like:

Answer honestly, let it ask, and when it's done it'll hand you a .md file you can attach to any chat or project going forward. Your AI will finally know who it's working with.

Want the full walkthrough with screenshots? It's on the Notion companion page.

You can now spot AI slop. Here's how. Smooth writing, zero substance, suspiciously perfect images. This piece breaks down the tells so you can stop falling for it. Read it here: techxplore.com/news/2026-05-ways-ai-slop

The Pope just published a 42,000-word letter about AI. Pope Leo XIV dedicated his entire first encyclical to artificial intelligence, calling for stronger regulation and drawing a direct parallel to the Industrial Revolution. When the Vatican weighs in, something is happening. Read it here: cbc.ca/news/world/pope-leo-magnifica-humanitas

A free cheat sheet for getting better at Claude. Ruben Hassid's Claude 101 covers everything from your first prompt to advanced workflows. Free, beginner-friendly, no strings. Bookmark it here: claude101.com

Sundar Pichai, CEO of Google, told The Rundown this week in three years, today's AI is going to look like a flip phone.

I thought about my first LG flip phone, peak 2000’s, where you wanted everyone to hear your ringtone because you paid $2 for it. Future us is going to laugh at the prompting era the same way we laugh at T9.

So yes, AI has a knowledge cutoff. Yes, it defaults to average if you let it. Yes, it sometimes acts like an overeager intern who didn't read the brief. But that’s just you learning to work with the flip phone. 

So tell us, how did you use AI this week? Hit reply and tell us. The best stories will be featured right here in a future issue.

Read the full interview: therundown.ai

Cinz.AI

Your unfair advantage, one week at a time.

Got a question, a topic request, or just want to say hi? Just reply to this email. I read everyone.