Video plus voice beats a prompt

When I started handing real work to bots, I explained it the way most people do. I typed it. I'd write out how a task should go, read what came back, find what I'd left out, and write some more. The results got better slowly, and the tokens went fast.

At some point I noticed I was spending more effort describing my screen than doing the work on it. What finally worked better for me was simple: share my screen, record it, and then narrate my reasoning while I do the task. That's how I explain work to bots now, and I think it beats a prompt for most of what I hand off.

What a prompt leaves out

When I type instructions, I write down the steps. Open this, check that, send the reply. What I don't write down is everything in between.

The first thing that goes missing is the reasoning. When I do a task myself, I make a lot of small calls without saying them out loud. Say, why one customer ticket gets a short answer and another gets a longer one, or why a step gets skipped when something on the page looks a certain way. That's hidden reasoning. A typed prompt gets the steps across and drops the why, which is the part the bot needs when the next case is a little different from the one I described.

The second thing is the screen itself. There's extra context that comes from the UI: which field, which number, what the page looks like when something's off. Describing a screen in text is slow, and it's easy to get wrong. Write "the total" when there are two totals on the page, and the bot can pick the other one.

Neither of these is a writing problem. You can write a very careful prompt and still lose both, because typing forces you to translate what you see and what you think into words first. The translation is where it leaks.

Pointing while you talk

Think about how you'd show a new hire a task. You wouldn't hand them a page of instructions and leave. You'd sit next to them, do it once, and point. "This one, not that one." "When you see this, stop and check."

Those words don't mean much in a typed prompt. They make a lot of sense when the bot can see what you're pointing at.

That's most of the reason a recording works better for me. I do the task the way I always do it, my cursor goes where my attention goes, and I talk while I work. We're also building a way to draw on the screen in LIT, for when I want to be specific about one part of it. The bot gets my words and the screen they were about, together.

Prompt engineering alternatives that keep the reasoning

Most prompt engineering alternatives I come across are still text. Templates, longer system prompts, pasted examples, structured formats. Some of them help. But they all ask you to do the same translation first: turn what you do on a screen into a written description, then hope nothing got lost.

Recording skips that step. You don't describe the task. You do it, and say why as you go. The steps, the reasoning, and the UI context all land in one take.

That's the first thing I built LIT to do. You press a hotkey and record your screen while you talk. When you stop, you can send the recording with its transcript to your bot. Record the screen, send it to any bot.

I want to be honest about what this doesn't do. It doesn't make the bot right on the first try every time. Bots still get things wrong, and you still have to check the work. And a recording is only as good as the reasoning you say out loud, so mumbling through a task isn't much better than a thin prompt. What changes is how much of what you know actually reaches the bot.

The recording sticks around

There was a side effect I didn't plan for. Once I'd recorded a few tasks this way, I had something I never had with prompts: a library of how I do things. Each recording is a small lesson with the video and the transcript.

A prompt tends to live in one chat with one bot. It gets edited, buried, and eventually forgotten. A recording of me doing the task and explaining it stays useful after that chat is gone. I can send the same one to a different bot if I want to try another provider, without explaining it all again.

That's the part that changed how I think about it. A prompt is something you write for one bot, once. A recording is closer to training material you keep.

Try it on one task

Pick a task you've written instructions for more than once, the one where the bot keeps getting something almost right.

Next time, don't rewrite the prompt. Start a screen recording, do the task yourself, and talk through it like someone new is watching over your shoulder. Point at things. Say why you skip the step you skip. Then send it to your bot.

Pay attention to what it asks you afterward. If the questions are different from the ones your prompt used to cause, that tells you what the prompt was leaving out.

LIT · Teach your AI bot by showing it your screen