Better output and less back-and-forth start with getting aligned before the first draft.
Many teams work with AI the same way they’d work with a new agency in its first month. They ask the AI for a campaign plan or a launch post and get back something that looks finished but reads generic. Then the back-and-forth starts, one tweak and one regenerated version at a time, until the draft is close enough to use. Readers have gotten good at spotting AI-generated writing, to the point where the em dash has become a running joke, so a lot of that back-and-forth goes to scrubbing out the giveaways.
Anyone who has onboarded an agency has been through the same thing. The agency has real expertise, but it doesn’t know yet what matters most to you or which tradeoffs you’re willing to make, so it fills the gaps with its best guesses. Every revision is the two of you getting aligned after the work has already been made.
With AI, that’s the model we’re running today. With an agency, we recognize those early revisions as alignment happening late, but with AI, many teams accept the back-and-forth as the cost of using the tool. I think the cause is the same in both cases. The work started before the two of you were aligned on it.
Getting aligned means the AI knows the things about the work that only you know. It knows what the work has to accomplish and who it’s for. It knows which of your inputs wins when two of them disagree, and what you’re willing to leave out. When it knows all of that before it starts, what comes back sounds like you, because every decision in it is yours.
How I get aligned with AI
Something I’ve found useful for getting aligned with AI is the Grill Me skill from Matt Pocock. Matt is an engineer who teaches developers to build with AI, and Grill Me came out of that world, but it works on almost anything. You start the way you normally would, by telling the AI what output you want and giving it your starting point. Then, instead of letting it build right away, you ask it to grill you. It works through every decision that goes into that output and offers its recommended answer with each question, until the two of you agree on all of it. Only then does it build.
Skills aren’t limited to Claude Code, either. You can upload them to Claude’s regular chat app too, so you don’t need to be a developer to use this one.
Why the AI should ask the questions
The giveaways that make output read as AI-generated come from the places where you and the model aren’t aligned. When the model doesn’t know what matters most or which decisions still need to be made, it fills every gap with its own defaults. Defaults are reasonable by design, which is exactly why a piece built out of them reads like it could have come from anyone.
A better prompt is the usual answer, and a good prompt is still the right starting point. The limit is that a prompt can only carry the decisions you already know you need to make. The decisions that make output generic tend to be the ones you don’t know are still open, and you can’t write a prompt for a decision you don’t know you haven’t made. Those gaps are hard to see from your side and easy to see from the AI’s side, so it makes more sense to let the AI ask.
The questions do more than inform the AI. They bring up angles you hadn’t considered and contradictions in your own thinking, like two goals you’ve been holding at once that can’t both lead. By the end of a session, the AI understands the work, and you usually understand it more clearly than when you started.
What this looks like on a launch post
Here’s a hypothetical. Say you’re launching a new feature, and you need the launch post plus shorter variations for LinkedIn and for the sales team’s outreach to prospects. You have plenty of inputs, including product marketing’s positioning doc, notes from a dozen customer interviews, the sales deck, and the release notes from product. You hand all of it to the AI and ask for the full set.
What comes back is polished and generic. The launch post calls the feature faster, more accurate, and more affordable, the LinkedIn version leads with speed, and the sales email leads with cost. That’s what unmade decisions look like. The positioning doc leads with speed, the customers who asked for the feature wanted accuracy, and the sales deck leads with cost. The documents disagree, and the AI resolved the disagreement differently in each piece instead of telling you about it. Nobody decided which benefit leads, so you spend the next hour going back and forth, fixing one version and watching the others drift.
Run the same request through Grill Me and the AI reads the same documents, then starts asking. The first question might look like this.
Your positioning doc leads with speed, but the customers who asked for this feature wanted accuracy. Which one should lead? I’d recommend accuracy, because it’s the benefit buyers described in their own words.
A few questions later, it might raise something you hadn’t thought about.
Three of the customer interviews mention a competitor that launched something similar in the spring. Should the post address it or leave it out? I’d recommend one line that addresses it directly, since prospects are already making the comparison.
Later questions work through the variations, like whether the sales email should lead with the same benefit as the launch post. All of these are decisions that were yours to make, and none of them would have made it into a prompt, because you didn’t know they were open until the AI read everything and asked. Once they’re answered, the AI builds the full set, and what comes back needs a light edit. The variations agree with each other because they’re built on the same decisions.
Judgment belongs up front
Writing is where this is easiest to see, because the unmade decisions show up in the sentences. The same gaps exist in a campaign plan or a budget recommendation, anywhere a lot of information has to be weighed and a call has to be made. Those outputs just hide the gaps better.
I think the bigger shift is in where your judgment goes. In the back-and-forth version of AI work, judgment shows up at the end, when you read the versions the AI hands back and pick the best one. That feels like judgment, but it’s closer to ordering off a menu someone else wrote. The AI already made the decisions behind every option, so you’re choosing the best of what it decided. Judgment belongs at the start, in deciding what the work has to do. When those calls are made up front, every option that comes back is built on your decisions, and choosing between them gets easy.
That matters more as AI takes on more of the production. Handing off more of the work only works if your thinking is solid and the AI is aligned with it, and grilling handles both at once. Matt makes the same point about his own skill, that the main thing a session leaves you with is clearer thinking. Your job is to provide the judgment, and the AI can supply the individual words.
What’s the catch?
The catch is that each recommended answer is the AI’s default, shown to you before it gets used. Seeing the decision is the whole value. If you accept every recommendation without thinking it through, you’ve handed the decisions back to the model, and you’re ordering off its menu again.
Most questions come with a recommendation and a few options. Picking one of them is fine when it fits, but you don’t have to stay inside them. When none of them are right, give a completely different answer. This is where dictation helps. I’ll often talk through how I feel about a question for a minute or two and let the AI make sense of it, especially when its recommendations missed the point.
Whatever you learn along the way about how you like to work can be saved as a memory, so the next session starts with some of that understanding already in place.
When AI handles the production, we still have to make the judgment calls. I believe we need to make them before the first draft instead of spending them on picking the best of what the AI hands back.