Two questions. Both are things I typed to an AI myself. Six weeks apart.
June 24th, I asked:
"That checkout button does nothing. No console error either. Help me find the cause."
July 23rd, I asked:
"I need you to correct my thinking. I'm not offering the kind of automation-plumbing service that everyone already does — what's the real value I provide? Please recalibrate me."
Same person. Same tool. Same AI. But those two questions are asking from two completely different floors of a building.
The first one: I know where the problem is (the button's broken), I want you to find the cause.
The second one: I don't know who I am, I want you to correct how I see myself.
I pulled all 1,427 prompts I typed to an AI between mid-June and late July. I set out to see "have I changed." What I found wasn't about how much I know technically — it's that the abstraction level of my questions climbed, one rung at a time. And this ladder is something nobody ever taught me. I only noticed it existed by looking back at the record.
So this is me pulling that invisible ladder apart. Because I'm increasingly convinced: real skill with AI isn't whether you can write a clever prompt — it's which rung your question is standing on.
The bottom rung. You already know the shape of the answer; the AI is just your hands.
This rung is useful and necessary — I still ask this way every single day. But if you only stay here, the AI is just a version of you that types faster. As specific as you are, that's as specific as it gets; whatever you can't think of, it won't think of for you either.
Mid-June me lived mostly on this rung. My prompts back then averaged barely 200 characters — short, direct, just giving orders.
One rung up, you state what you want, and hand over the "how."
June 26th I typed this, and it still feels honest to me:
"The pieces are actually there, they're just not wired together. I'm pretty weak at this wiring-together part — so, got a good approach?"
What's the difference? On rung 1 I give the shape of the answer; here I only give a direction and hand over the solution space. That takes a bit of courage — you have to admit you don't know how, before you can ask "got a good approach."
From this rung on, the AI starts handing you things you couldn't have thought of. Because you didn't pin it down with a specific instruction. On the third rung, you stop solving "a problem." You solve a whole class of problems.
June 23rd, I wasn't trying to fix a bug — I was trying to stop a kind of bug from ever happening again:
"I want to design a skill that proactively saves branching, unfinished tasks. When a conversation drifts into a fork, can it trigger a prompt asking whether to save it?"
July 3rd was even clearer — I wanted to extract a model's judgment into a method:
"Looking at it from a higher level: how does it scope a problem, sketch the outline, converge, and sequence the flow… make this endless stream of problems systematic."
On this rung, I go from "the person solving problems" to "the person designing the problem-solving machine." And the most interesting part — I often use my own flaws as the starting point of the design. I know I lose track of branching tasks, so I design a mechanism to catch them. I turn "what I'm bad at" into a spec.
In my logs from late June to early July, the share of prompts that "talk about method, governance, systematizing" jumped from 13% to 36%. Not because I suddenly got smarter — because the position I was asking from moved up a whole floor.
The top rung. The hardest. I only climbed onto it recently.
Here, you're not asking about the world — you're asking about yourself. You want the AI to be a mirror, correcting your judgment about yourself and your direction.
Why is this the hardest rung? Because on the first three you're still holding the wheel — you know what to fix, what approach you want, what class of problem to prevent. On rung 4, what you hand over is "my judgment about myself might be wrong." You have to admit you might not see yourself clearly, before you can ask "please recalibrate me."
This takes no technical skill — it takes a very mature self-awareness: knowing your blind spots need an outsider to light them up. And an outsider is exactly what a solo developer lacks most.
I used to assume it was the AI getting stronger, so I could ask harder questions. Looking back at the record — no.
It's that each rung shows you the next rung exists.
I fixed the same kind of broken button too many times on rung 1 — that's what made me ask, on rung 3, "can we design a mechanism to prevent this." I built too many systems on rung 3 without pushing a single one out — that's what made me hit rung 4: "am I pointed the wrong way, recalibrate me." The pain of each rung is the door to the next. You don't teleport to rung 4 — you have to trip over the same stone on rung 1 enough times first.
So if you spend most of your time right now on rung 1 asking "fix this," that's completely normal. I do too. The point isn't to force yourself to skip rungs — it's to not stay there pretending that's all there is. Next time you catch yourself fixing the same kind of thing a third time, try asking one level up: "is there a way to stop this class of problem from coming back?" — and you've just stepped onto rung 2.
I want to be clear about one thing, because it's counterintuitive.
A lot of people think "good at using AI" = "good at writing impressive prompts." It's not. My 1,427 prompts are full of typos, casual phrasing, "what do you think?", "I'm just confused." My prompts are not "engineered" at all.
What's actually improving is my judgment about which rung to put a question on. The same bug, asked on rung 1 as "fix this," versus on rung 3 as "design me a mechanism that prevents this class of bug," gives you wildly different things. And knowing when to move a question up a rung — that judgment is the real skill.
It's also why I'm not that worried about "AI is so strong now, everyone can do it, so what's my value." Everyone has the tool. But the person standing on rung 4, who knows to ask "recalibrate me" instead of just "fix this" — that's judgment you climbed to one rung at a time, over years of tripping over stones. No prompt template hands you that.
This comes from a script I ran over my own 6–7 month history — 171 sessions, 1,427 prompts. From the same data I also wrote "I Thought I Was Wasting My Time — While Designing a Cross-Model Controlled Experiment," about the gap between capability and self-assessment. This piece is about how the capability itself grows, one rung at a time.
本文原載於我的部落格： How I Ask AI Changed: From "Fix This" to "Recalibrate Me"