The prompts that actually moved the needle
TL;DR: The prompts that moved the needle weren’t sharper task instructions. They were six steering moves, repeated: plan first, ask system questions, verify don’t trust, add one constraint at a time, point at resources, anchor in values.
After shipping a product where AI did nearly all the work, I went back and read my own prompts to see which ones mattered. I expected the wins to be clever task instructions: sharp specs, the right phrasing for “build this.” They weren’t.
The prompts that actually moved the needle were never about the task. They were about steering.
The task prompts (“implement this endpoint,” “write these tests”) were interchangeable; the AI was going to do those fine almost regardless of wording. The leverage was somewhere else entirely: in a handful of steering moves I made over and over without noticing. Here they are.
1. Plan first, build second
My most-repeated prompt, in different clothes: “Tell me the plan and your recommendation, then wait for my OK.” “Charter first, build second.” “Don’t build yet; tell me what you understood.” This one prompt prevented more wasted work than anything else, because it catches a wrong assumption while it’s still a sentence instead of a thousand lines of code.
2. Ask system questions, not just task questions
The biggest outcomes all started by stepping out of the task: “How should we incorporate this tool into our workflow?” “When should you use a sub-agent versus doing it yourself?” “What’s the best-practice way to manage this?” A good system question pays off across every future task. A good task prompt pays off once.
3. Verify, don’t trust
“Are all the docs actually up to date?” “Do you really update this as you go?” Challenging a claim instead of accepting it caught real drift (stale documents, skipped steps) that a polite “looks good” would have buried. Make the AI show evidence, not assert it.
4. Extend the model one constraint at a time
Instead of writing a giant spec, I’d add a single constraint per turn: “Add this as a gate.” “I’ll have a different rule for that.” Each step stayed small enough to review, and the design converged without a big up-front document.
5. Unblock by pointing
When the AI said “I can’t, because…,” the highest-value reply was usually a pointer: “That already exists here, use it.” “Here are the official docs.” You supply the missing fact; the AI does the rest. Half of “the AI can’t do X” is really “the AI doesn’t know where X is.”
6. Anchor in your values
“I want to delegate as much as possible.” “Favor low-stress and reversible.” Stating what you actually optimize for makes the AI weigh trade-offs by your lens instead of generic defaults, and it’ll start flagging the burdensome option before you have to.
What about “think hard”?
I almost never typed “think harder” or “ultrathink.” The deep reasoning came from the moves above: a good system question and a plan-first gate force careful thinking far more reliably than a keyword does. The one explicit phrase I do use is “what’s the best practice”: it reliably pulls honest methodology instead of a quick answer. Thinking keywords have their place (genuinely hard design trade-offs), but they’re a complement to steering, not a substitute.
The takeaway
If you’re working with AI, start a running list of your steering prompts: the meta-moves, not the task instructions. Notice which ones consistently produce good outcomes and reuse them deliberately. Mine fit on one card: plan first, ask system questions, verify don’t trust, add one constraint, point at resources, anchor in values. The task prompts take care of themselves. The steering is where you earn your keep.