The Step Before the Step

I was two sentences into a request before I caught myself. My hands knew its shape before my head had finished checking it. I was typing “write me a Python script that renames the files in this folder to match our naming convention.” Reasonable enough. Then I stopped, cursor blinking, because that one sentence had already made three decisions for me. That it should be Python. That it should be code at all. That I understood our naming convention well enough to hand it over. I held down backspace and typed something else instead: “I’ve got a folder of messily named files and a convention they’re supposed to match. Before you write anything, what’s the best way to handle this, and what do you need to know from me first?” It came back asking whether this was a one-time cleanup or a recurring problem, because those are different tools. One time. It told me to skip the script entirely and use a rename pattern built into the tool I already had open. Ninety seconds, no code, done. I’d geared up to write software and the machine basically said “my dude, right-click, batch rename.” The best prompt I wrote that week was the one I erased.

I want to talk about that erasing, because I’ve come to think it’s the most useful habit I’ve picked up with these tools, and almost nobody does it on purpose.

Start with one of the first tricks anyone teaches you about prompting: ask the AI to write the prompt for you. You describe what you want in messy human terms, the machine hands back the clean version, and it works. Here’s the part of that lesson nobody tells you. The moment you let the model write the prompt, you’ve admitted it’s better at the step before your step than you are. And if that holds one step back, the only question left is why you’d stop at one.

There’s a version of this that goes deeper than prompt-writing, and it’s the one I use most now, so I’ll put it up front. You can ask the AI how to get the AI to do what you want. Not “do this,” but “What should I be asking you so that you do this well? What do you need from me? Where do requests like this usually go wrong?” You make the model the expert on its own operation, which it turns out to be. Last month I needed a contract summarized and I almost typed “summarize this contract.” Instead: “I need to understand this contract before a call tomorrow. What’s the best way to have you help with that, and what should I be asking that I probably won’t think to ask?” It told me a flat summary was the weak move, that I’d get more from asking it to flag the three clauses most likely to cause me problems, list what was unusual versus standard, and mark anything a lawyer should see. I hadn’t known to ask for any of that. The plain request would have gotten me a plain answer, and I’d never have known what I missed. That’s the move: whatever step you think you’re on, back up one and get help there. Then look at where you’re standing and back up again.

What you’re auditing, underneath, is your own assumptions, and there are always more of them than you think. Every starting point is a stack of decisions you don’t remember making. Watch it run on the most familiar dead body in office life, the quarterly deck.

“Make me a deck on Q3 results.” Back up one step: what is this deck supposed to argue? The answer that comes back is a question of its own, does the audience already know the numbers or are you presenting them cold, and you realize you don’t actually know. Congratulations, that blank spot is the reason you backed up. Back up again: who decides what after they see it? Turns out it’s one VP deciding whether to fund a project, which means eleven of your fourteen planned slides are context she already has. Once more, the one that stings: should this be a deck at all, or a one-page memo with the ask in the first line and a meeting on her calendar? Full WarGames: sometimes the only winning move is not to build the deck. Three messages. Ninety seconds. The task changed twice and shrank by eleven slides before a single one got built. The request I almost typed was the tip of an iceberg of little decisions nobody had looked at.

We never used to work this way, and there was a good reason. Stepping back was the expensive direction. The step behind your step belonged to someone who cost more, the strategist, the architect, the person whose whole job was knowing which deck to build before anyone built it. Getting that input meant a meeting, a favor, or a budget line, so starting from wherever you were standing was rational, and pushing forward with your assumptions still boxed up was what professional speed looked like. Same story as the excuses in the last essay: the price of doing it right fell through the floor, and the habit never got the memo.

Which sets up the part I most want you to take: most people use AI to skip steps. The ones getting the results use it to add them. Cheaper steps, earlier steps, the ones that used to require a consultant and now require a sentence. Every demo you’ve ever seen sells step-removal, the one-prompt miracle, the four-hour job done in the length of a TikTok, and that is exactly why step-addition is where the advantage went while nobody was looking. More prep, more foundation, precisely because prep stopped costing what prep used to cost.

Here’s the practice, and it’s mostly a handful of questions you ask before you ask for the deliverable. “Here’s what I’m actually trying to do, how would you set this up?” “What do you need from me to do this well?” “What should I be asking you that I’m probably not thinking to ask?” The honest one: “What am I assuming here that I haven’t checked?” And the afternoon-saver: “Is what I just named even the right thing to build?” Then the recursion, which is the part that takes practice. You read the answer and, instead of running with it, you ask the question standing behind that answer. The file-rename becomes “one-time or recurring?” The contract summary becomes “which clauses, and unusual compared to what?” Each answer has another question hiding behind it, and you keep going.

Which raises the obvious problem: you can keep backing up forever, and forever is just a slower way to never start. So here’s where you stop, and here’s why. Each step back earns its keep only if it changes what you’d do next. The file-rename question changed the answer: script became no-script. But keep going past that and you hit questions whose answers don’t move anything: “What’s the deeper purpose of renaming these files?” There’s no purpose worth a sentence there, and asking it just spent thirty seconds to arrive back where you already were. That’s the tell. The moment a step back returns an answer you’d have acted on anyway, you’ve found the floor, and every step below it is costing you the exact time this whole practice was meant to save. Over-asking and under-asking burn the same afternoon from opposite ends. So back up until the answers stop changing the plan, then go. Two exceptions. Don’t scope a two-line email; some tasks are exactly the size they look. And watch the trap where the questions have stopped changing anything and you keep asking them because asking feels safer than starting. That’s not preparation anymore. That’s a waiting room.

Let me show you one more, a civilian one, because this was never only a coding trick. “Polish my résumé.” Everyone’s typed it. Back up: which roles is this actually aiming at, because a résumé pointed at everything lands nowhere. Back up again: what do you want the next three years to look like, since the résumé is only there to open a door onto it? And the fully recursive version, the one that changes it most: “What should I be telling you about my experience so you can help me point this in the right direction?” Suddenly the AI is interviewing you instead of running a Snapchat filter over the résumé you walked in with, and the résumé that results works because it stopped being step one. It became step four, and the first three steps were the ones doing the work.

If this instinct sounds like it wants to be written down and made repeatable, it does, and it has a name: a kickoff. There’s a whole framework built out of exactly this reflex over on the other site, if you want the system instead of the habit. But the habit is the thing. The system is just the habit with a checklist stapled to it.

Right now, working this way makes you look a little slow. You’re asking setup questions while the room is high on instant, and instant is what everyone’s selling. That’s the fringe stage, and it ends the way these always end, because your work comes out right the first time, and right-the-first-time compounds while everyone else is opening revision four. There’s a version of the standard question coming, the one that’ll sound as obvious in three years as “did you spellcheck it” sounds now: “did you scope it with the AI first?”

And here’s the feeling I want to leave you with, because it’s the argument that actually moves me, and it’s smaller than all of this. It’s the tiny, specific satisfaction of the deleted request. You’re two sentences into asking for the wrong thing, you stop, you back up, you ask the question hiding behind the one you were about to send, and you watch the whole task get simpler at the exact moment it gets better. That’s the click. The step before the step was never a delay. It was the work, showing up early.

The New Typing continues — the series lives at drjeffwurfel.com/writing.