Last week I wrote about using AI to edit my videos. It did work, eventually. Eventually is the key word.

My first try was five paragraphs describing WHAT I wanted and HOW I wanted it done. I expected it to one-shot the rest. The results were pitiful.

A lot of people like to talk about one-shotting a task with AI. I'm happy for them. For trivial tasks, sure, you'll get most of what you want. For anything complex, one-shotting is playing roulette with your time. Sometimes you hit. Mostly you don't.

Those pitiful results I got? Those were what triggered the lightbulb above my head. Chasing the one-shot was just wasting my time.

I thought I already knew how to "teach" AI. Over the past year I was told I had to master skills, agents, prompt engineering, and context engineering. Turns out the AI can do most of those better than I ever will.

That's what went wrong with those five paragraphs. I gave it too little of the WHAT, and what I did give it was buried in the HOW. I was telling the AI how I wanted things built, because that's what I'd been doing for a year. Those days are gone.

So I stopped telling it HOW and started showing it WHAT. I walked the LLM through editing an episode by hand. I explained what I wanted and why. Sometimes I made the edits myself and had it review what I did so it could understand.

After about an hour of that, the video was the way I wanted it. Then I told it to review everything we had done in the session and build the pipeline to do it for the next episode.

Even then, it didn't just take off and run. We argued back and forth for a bit until it understood. That's because I end almost every prompt the same way:

"Interview me 1 q at a time until you understand."

It's a holdover from way back last year, I know. It's still one of the most valuable tools I have.

I used to say an LLM was like a really smart five-year-old with a bag of sugar. Maybe it's a ten-year-old now. You still have to be patient. You still have to explain. You still have to spend the time... just not nearly as much as writing the code by hand.

So stop chasing the one-shot.

  • Explain the WHAT.
  • Demonstrate when necessary.
  • Let it interview you until it understands.

Now let it figure out the HOW.

Originally posted on LinkedIn: "Stop Chasing the One-Shot"