Four things nobody explains, so you can stop guessing and start building smarter.
Why does your AI "forget" halfway through a project? Why does it work great in one chat and fall apart in the next? Which app builder should you actually use? And how is it clicking around your browser? Today we answer all four.
A model is the brain behind the app. Claude, ChatGPT, and Gemini are the apps. Inside each one are different models: bigger ones that think deeper, and smaller ones that move faster and cost less.
Best for building apps, hard problems, and long projects. Slower and costs more.
Great balance of smart and fast. Most daily tasks live here.
Fast and cheap. Good for simple, repeated tasks.
Think of the context window as the model's desk. Everything it's working with right now has to fit on that desk: your messages, its replies, the files it opened, the screenshots it took. The size is measured in tokens, roughly three quarters of a word each.
When the desk fills up, something has to go. The AI either summarizes the older stuff to make room, or starts losing track of what you said at the beginning. That's why a long chat suddenly "forgets" your rules.
About 150,000 words of working space, on the free plan.
About 20,000 words of working space, on the free plan.
The desk. Gone when the chat ends or gets cleared.
The filing cabinet. Still there next week, in a brand new chat.
Memory settings remember facts about you, but only in that app. Projects hold instructions and files, but only for that one project in that one app. Switch tools and it's gone.
CLAUDE.md and AGENTS.md are files inside your project folder. Move the folder, push it to GitHub, open it on another computer or in another tool, and the instructions come with it.
| Tool | What /init creates | Reads AGENTS.md? |
|---|---|---|
| Claude Code | CLAUDE.md | Only if CLAUDE.md points to it |
| Codex | AGENTS.md (there's no "Codex.md") | Yes, it's Codex's main file |
| Cursor | Nothing on its own. You ask. | Yes |
| Copilot | Nothing on its own. You ask. | Yes |
/init is short for initialize: set it up for the first time. The AI reads your whole project and writes the file for you. The slash tells it "this is a command."
Update AGENTS.md with what we just decided and anything you learned in this session that you'd need to know next time. Keep it short. Remove anything that's no longer true. Make sure CLAUDE.md contains only: @AGENTS.md
When you pick a folder for your AI, it doesn't have to be on your Desktop or in Downloads. You can pick a folder from Google Drive too. Your files stay in the cloud, so they don't fill up your laptop.
Slides, syllabi, PDFs, videos, writing. Pick the folder from Google Drive and let the AI work on it from there. Just make sure you're online.
Don't put code in Google Drive. Syncing can break it. Keep it in a normal folder, push it to GitHub, and when you're done, delete it from your computer. You can clone it again anytime.
Type an idea, get an app. These tools are amazing for getting started. The question is what happens when real people start paying you.
| Tool | Best at | Database & hosting |
|---|---|---|
| Lovable | Beautiful designs from a prompt | Connects to Supabase, hosts for you |
| Bolt | Fast full apps right in the browser | Connects to Supabase, hosts for you |
| Replit | Everything in one place: editor, database, hosting | Built in, lives on Replit |
| v0 (by Vercel) | Clean front ends and pages | Deploys to Vercel |
| Claude Code / Codex | Real production apps you fully control | Your GitHub, your Supabase, your Vercel |
Testing an idea fast. No setup, nothing to install, results in minutes. Perfect for a prototype or a demo.
Credits that run out fast when it gets stuck in a fix loop. Harder to customize. Your app lives in their house.
The AI sees the screen, moves the mouse, clicks buttons, fills in forms. Works on almost any website. But it's slower, and it can get lost when a page changes.
A direct line into the tool, no clicking needed. Faster and more reliable. Think of it like installing an app on your phone.
Claude can look at your tables, check your security rules, and read your logs when something breaks.
Claude can check your deployments, read error logs, and tell you why the live site isn't working.
Claude can save your work, open pull requests, and keep track of every change.
My live site isn't working the way it does on my computer. Using the Vercel and Supabase connectors, check my latest deployment, read the error logs, and confirm my Supabase keys are set in Vercel. Tell me in plain language what's wrong before you change anything.
Missed a week? See all past live sessions →