ChatGPT Wrote Our Podcast Intro (and It Sounded a Little Rough)

The entire introduction of our podcast was written by ChatGPT and read by an AI voice clone of the host. It sounded a little rough around the edges.

That’s the version I want you to start with. We used it for the mundane, structural part of the episode so we could spend our energy on the actual message and the conversation. Low stakes. Easy to try. If you’ve been sitting on the sidelines waiting for AI to feel “ready,” this is your on-ramp.

Once you try it, the interesting question isn’t whether it can write a paragraph. It’s where it actually saves you time, and where you still need to stay awake.

Brooks keeps coming back to context switching. The research task is rarely the expensive part. Leaving the thing you were doing, hunting for a fact, and trying to climb back into the original thought is.

You know the workflow. You’re in a doc. You need one fact. You open a search engine. Then another tab, because the first result is thin. Then a third source, because the first two disagree. Ten tabs later, you’re scrolling around trying to remember what you were even writing.

Are you losing five minutes every time you need a fact?

Notion, Coda, and DevonThink are putting AI inside the workspace. You can generate a summary or pull action items without opening a new window. You stay in flow instead of breaking it to go consult a search engine. Single window. That’s the shift worth paying attention to.

Don’t take the first answer

Most people type a prompt, get a block of text, and call it done. Brooks treats that first output as a starting point.

He calls it layered prompting. Ask for a broad list first. Benefits of AI. A first pass at competitor analysis. A pile of blog topics. Then immediately follow up with a filter. Focus only on the time-saving benefits. One competitor. One angle.

The first prompt does the grunt research you didn’t want to do yourself. The second prompt is where it gets useful. You’re in an iterative dialogue, not a one-shot transaction.

Next time you’d normally bounce across ten tabs, try the two-step. Broad, then filter. Don’t accept the first answer.

Then I asked it to do the math

I had mixed units sitting there. One figure in cm, another in lbs, plus a third number. I wanted a converted result from a standard formula.

I put those figures into an LLM and asked it to run the calculation. It handled the unit conversion and applied the formula. Precise number. Seconds.

I wasn’t expecting that. I thought these tools were for drafting and rewriting. They’re also useful for this kind of data analysis. If you’ve got a complex formula or a mix of metric and imperial, prompt it to do the math. Then check the number. I can tell when a calculator just saved me a pile of conversion work.

The work you’ve been ignoring

Someone created an AI service that appeals parking tickets and property tax bills by generating formal legal letters. Success rates have been high enough to threaten a city’s ticket revenue.

People give up on that kind of task. Not because they can’t write a letter. Because they don’t have the time or the energy to fight a parking ticket with a formal document.

That’s AI moving into administrative work. The high-friction, bureaucratic stuff. What’s the equivalent on your list? The thing you’ve been ignoring because it’s annoying, not because it’s complicated?

Type the request. Skip the interface.

The future of this, as I see it, is plain English execution. Tools like Adapt AI let you type a request (“Add Brooks as a lead, tag as interested”) and the software clicks through the CRM and does it.

You don’t learn the interface. You don’t build a macro. The cognitive load of the admin work drops to zero. The learning curve is the thing that goes away.

Imagine never learning a new software interface again. That’s the direction. Natural language sitting on top of the tools you already use.

Look at what comes back

Speed is the easy part. Accuracy is the part people skip.

Brooks prompted an image generator for a “happy, peaceful Asian man near the beach.” The results were iffy. Stereotypical. The model reflected the biases in its training data.

If you’re going to use generated images or text in anything professional or public, look at them. Cultural accuracy. Demographic accuracy. Does this look like a real person, or like a collage of what the training data thinks that person looks like?

Build in a verification pass. Fast output, then a human check before it goes anywhere.

And while you’re being careful about the output, be careful about the input too.

Brooks’s warning: most current AI tools rely on a single underlying model (OpenAI). Your data and your creative outputs get processed by one entity. That’s a privacy risk and an IP risk. Centralized training data also means you inherit the bias and the copyright problems sitting in that pile.

Before you paste anything confidential, ask:

1. Where does this data go?

If you can’t answer that, don’t put anything sensitive in.

2. What happens to the IP?

Your prompt, their model, the thing that comes back out. Get clear on who owns what before you use it in professional work.

3. Am I okay with one vendor in the middle?

Single-vendor risk is easy to ignore until the tool is embedded in everything you touch. Especially if you work in a regulated industry or you make creative work for a living.

Embedded AI is useful. It is not a private notebook.

Pick the most annoying bureaucratic task on your list. The parking-ticket equivalent of your workweek. Run it through a tool you already have, use layered prompting (broad, then filter), and read the output before you send it anywhere.

One task. Today.