Keep Reading the Prompts

On Friday, a longtime colleague and I were getting a new AI tool set up on her work computer. I already had it running on mine. She was getting it installed on hers, and I was helping — which mostly meant I was sitting next to her saying Continue, Continue, Continue. She was reading every prompt.

My colleague leads a large team responsible for some of the most-visited real estate on our company's website. She's not a developer. She's a strategic senior leader. But she stays close to the work — not to manage it, but to understand it.

My own work with AI is in a different register: inside-the-firewall analysis, strategy, productivity, and activation. I've been using this particular tool for over a month. I know its edges. I've built a high level of trust with it, earned over many reps.

So during the install, when the tool surfaced each change it wanted to make to her machine, my instinct was to push her through. I wasn't being reckless. I was rushing her for several reasons: I'd been using this tool for over a month and I know the shape of its prompts. And the work were were setting it up to help with is the same low-risk, firewalled territory I'd been living in.

She was doing something else. She was reading each prompt. Each proposed change. Each plan the tool laid out. The code it was generating. The commands it was about to run. I'd have burned through the install in twenty minutes. She and I worked on it twice that time.

Being fast was right, for me. Being more deliberate was right, for her. And the gap between those two "rights" is why I'm so glad we have the right people like her minding our big storefront.

Her seat is different from mine. She's accountable for a surface where one wrong push, one cut corner, one undetected assumption can turn into a very public, very expensive problem. When you carry that kind of accountability, you don't develop the click-through reflex. You develop the opposite reflex. You read the prompts because the prompts are where the tool gives you a chance to catch something before it happens.

She was also reading because she wanted to understand. She's the kind of person who doesn't like to have tools she can't explain. So some of what looked like caution was the learner in her, taking the moment to grok what was being installed and why. That's not slowness. That's depth.

The people who are best at using these tools, in my experience, aren't the people who move the fastest. They're also not the people who move the slowest. They're the ones who read the stakes of their own seat, and match their pace to it.

For anyone whose role has real stakes — you're accountable for something that, if it breaks, matters — here are seven things I watched her do that I think are generalizable.

1. Read the prompts. Every one. When a tool asks whether to proceed, that's not friction. That's the moment the tool is handing you a review opportunity you couldn't afford to build yourself.

2. Look at the plan before the action. Most serious AI tools will describe what they're about to do before they do it. That description is the cheapest review you'll ever get. Plans are cheap to examine. Actions are expensive to undo.

3. Understand the output — don't just receive it. If the tool wrote code, read the code. If it made a plan, understand the plan. Accepting an artifact you can't explain is a form of delegation, and there are things you cannot delegate.

4. Know what's irreversible in your environment. Every accountable role has a short list of actions that can't be taken back. Before the AI is operating in that environment, your list should be explicit — to yourself, and where possible, to the tool.

5. Keep experiment and production on separate surfaces. Where you play and where your real work lives shouldn't be the same ground. The trust you build in one does not automatically transfer to the other.

6. Read the output with the skepticism you'd bring to a new hire's first memo. Fluency is not accuracy. The tool sounds confident when it's wrong. That's probably the single most important thing to internalize.

7. Do the reading yourself. Your accountability doesn't transfer to whoever is sitting next to you doing the clicking. If it's your seat, you read.

A lot of the AI conversation online lives at two poles: "you'll be left behind if you don't jump in", and "these tools are dangerous". Neither pole is useful if you're the person who actually carries outcomes.

What I saw on Friday was the useful middle. You adopt. You absolutely adopt. These tools are too powerful, too useful, and moving too fast to sit out. But you adopt with both hands on the wheel. You keep reading the prompts. You stay the person doing the work, even when the tool could do it without you looking.

I wanted to click Continue. She wanted to read, and I take great comfort in that. The difference between us isn't something to resolve. It's the engine. It's what propels the work forward while keeping it honest and safe in this brave new world.