AI Improv, August: every question, answered

First, a confession. I promised I'd answer every question from the chat, and then I did not save the chat. Zoom did exactly what I told it to do, which was nothing. That one's on me and my settings.

So here's what I've got: every question that got asked out loud, answered properly. If yours was in the chat and it's not here, hit reply and ask me again. You will get a real answer, not a link to go read something.

Names are stripped. Nobody gets called out.

Next one is Tuesday, September 22, 7 to 8 PM Central. Follow me on Lu.ma and you’ll be notified when I post the invitation.

Lovable. It's all-in-one vibe coding, which is a fancy way of saying you type what you want and it builds it.

The free tier gives you a handful of credits a day, and one credit is roughly one prompt. That's enough to find out whether the idea works. It is not enough to build the whole thing in a weekend.

My advice: build something small and dumb first. A one-page thing you don't care about. Learn how it responds to you before you point it at the real project.

No. Pick one and get good at it.

The reason I show you several is so you know what's out there, not so you feel behind for using one. I bounce between them because my whole business runs on this stuff. If you're using ChatGPT for everything and it's working, that's a complete answer.

If you do want to try others, make a throwaway email address and use that to sign up. Then the marketing emails go somewhere you never look.

Go to chatgpt.com in your browser. That's the real one.

The app stores are full of things that look like ChatGPT and are not ChatGPT. If you're going to download an app, check that the publisher says OpenAI. If it says anything else, back out.

Honestly though, start in the browser. Everything works there, and some things only work there. Custom GPTs are a good example. If someone shares one with you and you open it in the app, it can just refuse. Open the same link in a browser and it's fine.

I stayed browser-only for about a year and a half. It cost me nothing.

It's not gone, it moved.

The picker got reorganized around how hard you want it to think instead of model names, so now you see options like Instant, Thinking, and Pro. The old list lives one layer down. Click the model name at the top of your chat, then click Configure in that menu. That's where you'll find legacy models, the thinking-effort setting, and the toggle for whether it auto-switches on you.

Worth knowing: a lot of the older named models have actually been retired over the past year, so some of what you remember genuinely isn't there anymore. What's left is under Configure.

Great question, and the answer is more interesting than yes or no.

It knows today's date because the date is handed to it at the start of your conversation. It's not remembering it and it's not looking it up. It's reading it off a note that got stapled to the top of your chat.

So both versions of the question point it at the same note. What changes is the mood. Ask it what it thinks and you invite hedging. Ask it what the date is and you usually get the date.

That difference is small here. It is not small when you're asking about something real. "What do you think about this contract" and "what does this contract say" will get you two very different answers.

Behind enough to burn you.

I'm not going to print a specific cutoff date here, because it's different for every model and it changes constantly. Any number I give you is wrong by the time you read this.

Here's what to do instead. Before you let it walk you through anything that involves a real tool or a real company, ask it flat out:

I lost an hour to this once trying to set up a CRM. It confidently walked me through a version of the product that didn't exist anymore. Not because it lied. Because nobody told it to go look.

The other move that saves me constantly: screenshot what's actually on my screen, paste it in, and ask "what do I click next." Now it's looking at reality instead of memory.

I'm on a high tier now, but only because my whole business runs on this. I ran on the entry-level ($20/month) paid plan for well over a year and it was plenty.

Check their pricing page for current numbers, because these change often.

What I'd actually suggest: pay for one, not three. My setup for a long time was one paid ChatGPT and one paid Claude, and I'd play them off each other. Heavy lifting and data crunching in ChatGPT. Writing and polish in Claude. Claude is the Shakespeare of these tools. It's also terrible at images, so don't ask.

Yes, and this is where my recipe earns its keep. It's called ROCO.

R is Role. Who should it be? Make it up. "You're a chief of staff who builds board decks for medical device executives." I invent roles constantly.

O is Objective. What do you actually want. Not what you don't want. "A polished deck I can put in front of a board" beats "something not too corporate."

C is Context. All the messy extra. Who's in the room, what they care about, what they hate, what happened last quarter. Most of us give way too little here. We're used to keeping our thoughts to ourselves.

O is Output. What shape you want it back in. Seven slides. A Word doc. Three bullets.

These 4 (ROCO) are the ingredients for my recipe in building a prompt. And it works with every AI model. Change any one of the 4 ingredients and you’ll get something different back.

Put those four things in and you'll get something usable on the first try. Then push on it.

Two more things on decks. Gamma will build you one from a couple of sentences, and it got dramatically better this year. And both ChatGPT and Claude now connect to Gamma and Canva directly, so you can go from notes to a real deck without opening PowerPoint.

Check every statistic it gives you before that deck leaves your hands. It wants to make you happy, and a number that makes you happy is easy to invent.

Ask your security team before you put anything in it. Not me, not the internet. Them.

That's not me dodging. Anything touching PHI is a compliance question with real consequences, and the answer depends on what your company has already approved and signed. Some organizations do use Gamma. That doesn't mean yours has cleared it.

The safe version while you wait for an answer: build your deck with fake numbers and real structure. You'll still save the three hours. You just won't put anything sensitive anywhere it shouldn't be.

AI Improv runs monthly. The next one is Tuesday, September 22, 7 to 8 PM Central. Same format, always free, always unscripted. Come back.

The free stuff is at aiconfidential.net/freebies. Prompt kits, resources, all of it. No catch.

The ROCO Prompt Builder is a free GPT I built to help you write prompts that will get you decent results.

Joy Prompt Club is my paid community, $27 a month. About what you'd pay for a streaming service you don't watch. joypromptclub.com

Mostly Human is my weekly newsletter. If somebody forwarded you this, you can sign up at newsletter.aiconfidential.net.

You are the boss. The AI is an overconfident, overeager intern who graduated from Harvard, thinks it knows everything, and has never met you.

It doesn't know your job. It isn't in the room. It can't see what you see.

So tell it. That's the whole skill.

And never take the first answer. Push back once and watch what happens.

💜
Deb