The Agent Suggested Forty Hooks for This Essay and I Used None of Them

I asked my content agent to draft hooks for this essay. It gave me forty options. I scrolled through all of them. Not one was usable.

That's not a failure. That's the actual ratio when you work with AI agents at scale.

Most people building with AI right now are stuck on the wrong question. They want to know: "Can the agent do this task?" The better question is: "What's the hit rate, and is that good enough to change my workflow?"

If an agent gives me forty hooks and I use zero, that sounds like a waste. But here's what actually happened: it took the agent eleven seconds to generate all forty. It took me ninety seconds to read them. In those ninety seconds, I ruled out every angle I didn't want to take. I saw the obvious plays, the safe plays, the plays that would sound like everyone else writing about AI right now. I saw what the model thought would work.

And then I wrote this one.

The agent didn't write the hook I'm using. But it cleared the space for me to write it. That's a different kind of leverage.

People expect AI to work like hiring a junior writer who gets better over time. That's not the model. The model is: the agent generates a large surface area of possible, and you pick the true. You are not training it to know what you would pick. You are using it to find out what you would pick.

I could have spent thirty minutes staring at a blank page, trying to come up with the perfect opening. Instead, I spent ninety seconds seeing forty things that weren't it, and then I knew exactly what was.

The constraint people hit when they try to use agents for creative work is they want the first output to be good. It won't be. The first output is almost never good. The question is whether seeing the first output — and the second, and the tenth, and the fortieth — gets you to the good one faster than you would have gotten there alone.

For me, the answer is yes. But only because I'm willing to throw out thirty-nine hooks without feeling like I wasted the agent's time. The agent doesn't have time. It has tokens.

Here's what I think most people get wrong about working with agents: they treat every output like it cost something. Like the agent labored over it, like rejecting it means you're being picky or inefficient. You're not. You're just using the tool the way it's built to be used.

A pencil doesn't care how many times you erase. An agent doesn't care how many drafts you reject.

The workflow I've settled into is this: ask the agent to generate a wide set of options, scan them in under two minutes, reject all of them if none land, and then write the real one myself with all the bad angles now cleared from my head. Sometimes one of the forty is good enough that I edit it into shape. Most of the time, none of them are, and that's fine.

What I'm not doing anymore is opening a blank document and waiting for inspiration. The agent shortcut isn't "the agent writes it for me." The agent shortcut is "the agent shows me forty ways I don't want to say it, in eleven seconds, and now I know exactly how I do."

I used to think the goal was to get the agent to a place where I could just publish what it gave me. I don't think that anymore. The goal is to get the agent to a place where it makes me faster at doing the thing only I can do.

Forty hooks, zero used. Still worth it.

— Mir