I Blame Tony
The gist: when does someone outgrow a chat window?
Tony — same Tony — has been building things. Real things. Multi-file, multi-component, talking-to-APIs things. He's still doing it through a chat window mostly. Pasting files in. Pasting code back out. Playing middleman between the AI and his project.
I got home, got stoned, and said to myself, "Self, where are the cookies, and how big is the average site?"
So I Asked
I didn't open a calculator. I didn't write a script. I said — and I'm barely paraphrasing here — "Hey, how big is this project compared to a context window?"
The AI ran six terminal commands. Simultaneously. Scanned every file, counted every line, measured every character, ranked the largest files, mapped the directory structure. All while I sat here and watched.
Here's what came back:
| Files | 821 |
| Source files | 387 |
| Lines of code | ~84,000 |
| Estimated tokens | ~1.93 million |
Cool. Now here's the fun part.
It Doesn't Fit
Every AI has a context window — the maximum amount of stuff it can hold in its head at once. Your code, your question, its answer, the whole conversation. Everything has to fit.
| AI Chat | Window Size | How Much of This Site Fits |
|---|---|---|
| ChatGPT | 128K tokens | ~6.6% |
| Claude | 200K tokens | ~10% |
| Gemini | 1M tokens | ~52% |
On ChatGPT, the AI can see less than 7% of this project. Even Gemini's million-token window — the biggest one available — only holds about half. And that's before the conversation itself starts eating into the space.
This isn't a massive enterprise app. It's a personal website.
The Relay Race
If you wanted those numbers from a chat window, here's what that looks like:
- Ask the AI to write a terminal command
- Copy it
- Open a terminal
- Paste it
- Run it
- Copy the output
- Paste it back into the chat
- Repeat five more times
Fifteen to twenty minutes of being the AI's hands and eyes. And every paste pushes older context off the desk. By the time you've given it what it needs, it's already forgotten the first thing you showed it.
What Actually Happened Here
I asked one question. The AI:
- Ran 6 commands in parallel — not one at a time, simultaneously
- Scanned 821 files across 20 directories
- Calculated token estimates, divided by each model's context window
- Produced the percentages, the tables, the comparison
- Wrote this article's first draft
My contribution: one sentence and some feedback on tone.
That's the difference between a chat window and a workspace. The AI didn't just know things. It did things. Searched the project. Read files. Ran commands. Built something from what it found.
Same AI. Same brain. The only thing that changed is whether it has tools or a text box.
The Point
There's a moment — you'll feel it — where you spend more time managing the AI's context than actually building. Where you paste a file and realize the AI already forgot the one you showed it two messages ago.
That's when you need a workspace. Something like a Codespace, Cursor, or any IDE with an AI integration. The AI can see your project, search it, edit files, run builds. You stop being the relay runner and start being the one calling the play.
Tony's at that moment. If you're reading this and nodding, F-U I was going to watch YouTube when I got home. (Don't fire me, plez.)
And if the right tool doesn't exist? Build it. Preferably using an agentic IDE. Because obviously.
Thank you for your time. I hope you have an awesome day.