My first week of tokens went to teaching bots, not using them

When I first set up Grok Bot, I went through the work at Gravitate that annoyed me most, the stuff I figured a bot could do. Nothing fancy. The kind of work that lives entirely on a computer: an email that needs to go out, an update that needs to be posted, a visit to a product page to check something. If a bot has its own computer, I figured it could take those off my plate.

It could, eventually. But the first week didn't go the way I pictured. When I talked it through later on a recording, this is how I put it:

"My entire token allocation for the first week pretty much went to just me teaching the bots how to do parts of my job."

Then the tokens ran out, and I couldn't use the bots I had just trained to do the work I had trained them for. I had paid for a week of help and spent it on orientation.

If you pay for an AI agent so it can take over parts of your job, this is worth looking at closely. I don't think it's a bug in one product. I think it's how the work is shaped right now, and most of the advice about reducing LLM token usage skips it.

Teaching a bot is onboarding a new hire

What surprised me most is how much it felt like training a person. You give instructions, and they have to be good enough for someone else to do the task the way you want it done. The bot makes mistakes. You correct it. You explain again, a little differently. Eventually it gets there.

Nobody expects a new hire to produce on day one. You plan for onboarding. You know the first stretch is mostly you explaining how things work here, and you accept that because the person sticks around and pays it back.

I didn't plan for any of that with a bot. I treated the subscription like a tool I'd switch on and start using. It's closer to a hire that you pay by the hour, including every hour you spend explaining.

Teaching and doing come out of the same budget

This is the part that caught me off guard. With a person, onboarding comes out of your time, and their paycheck is the same whether they're learning or working. With a bot there's one meter. Every explanation, every correction, every "no, not like that" spends the same tokens the bot would have used to do the actual work.

So the more teaching a task needs, the less is left for doing it. In my first week, teaching won that fight completely. I was still training the bots when the allocation ran out, with nothing left to actually use them.

Where the tokens actually go

Looking back, the first explanation wasn't the expensive part. The correction loop was.

These bots are built to be confident and to act. That sounds great. The problem is they're not always honest, with themselves or with you, about the gaps in what you told them. They say "I got it," go do the task confidently, and do it wrong. Then you're stuck correcting them until it comes out the way you want.

That makes me the bottleneck twice. First when I'm explaining, because how fast and how clearly I can describe what I want sets the pace for everything after. Then again when the bot comes back with a result, because I have to figure out whether it misunderstood or I described the job badly.

Every round of that loop costs tokens. A vague explanation doesn't save anything. It moves the cost into the corrections, where you pay for it again and again.

How I'd try to reduce LLM token usage while teaching

Most tips for reducing LLM token usage are about the work itself: shorter prompts, smaller models, less context stuffed into each request. That's reasonable once the bot already knows the job. It doesn't do much during the teaching phase, because the cost there comes from round trips, not from how long each message is.

What helped me was changing the question. Instead of asking how to make each message cheaper, I started asking how to make the first explanation complete enough that fewer rounds come after it.

For me, the answer was to stop typing instructions. The most efficient way I've found to explain how I want something done is to share my screen, record it, and narrate my reasoning while I do the task. The bot gets the steps, the reasons behind them, and what was on the screen in one pass, instead of piecing it together from my corrections.

That's why I built LIT. You press a hotkey, record your screen while you talk, and send the recording with its transcript to your bot. Record the screen, send it to any bot. That's the whole idea.

I'm not going to put a number on what this saves, and it doesn't remove the checking. The bot still has to show me it can do the work, and I still have to look. What I'm sure of is this: how fast I can explain decides how fast a bot starts paying for itself. Explaining was the slow part for me, and a recording is the fastest way I've found to do it.

One thing to try this week

Pick one task you've already explained to your bot more than once. Before you explain it again, scroll back and look at how many messages it took last time, corrections included.

Then do the task yourself once with a screen recording running, and say out loud why you're doing each step, the way you would if a new hire were sitting next to you. Send that to your bot and see how many rounds it takes this time.

If it takes fewer, you've found where your tokens were going. If it doesn't, you've learned something about that task, and that's worth knowing before you spend another week of tokens on it.

LIT · Teach your AI bot by showing it your screen