Down, Up, Sideways: The Three Directions I Dig Deeper With AI

Everyone says the key to working with AI is asking better, deeper questions.

I used to think so too. Until I mined six weeks of my own AI conversations — over a thousand prompts (that's a different article) — and laid out "within one topic, how did I keep asking further." What I found was counterintuitive:

"Deep" isn't a line. It has direction.

"Digging deeper" turned out to be three completely different moves — down, up, sideways. And I pick the direction automatically depending on the topic, without ever noticing I was doing it. Here are three real conversations (details rewritten, names of clients and internal tools stripped, but the shape of the questions is intact) to show what I mean.

Direction 1: Dig down — I don't want the thing, I want how it thinks

Once, I wanted a very capable frontier AI model to help me build an "automation pipeline generator." I started out normally enough:

"Can it understand these tasks and then design a workflow and a generator for me? Make the cold-start precise, efficient, reusable."

I wanted an output: a tool.

But partway through, I stopped and switched the question:

"Is there a way — through questioning — to systematize how it scopes a problem, sketches the outline, converges, sequences the flow? Write down this soon-to-retire master's method, so that whatever similar problem I hit later, I can emulate it?"

I didn't want the generator anymore. I wanted the judgment behind the generator — extracted into a method I could reuse over and over.

And I didn't stop there. Once it handed me a method, I didn't take it on faith. I designed a problem it had never seen and threw it back:

"Now use this method on a problem you haven't done before. I want to verify it's real bones, not just pretty. And tell me honestly afterward: did the method actually help you locate the answer, or would the old way have been enough? If it didn't help, say so — that means I adopted it wrongly."

That's digging all the way down: from "give me a tool" to "give me the method" to "prove the method is real" — with the falsification condition written in advance. The downward direction is cutting through the surface to something you can verify at the bottom.

Direction 2: Dig up — should this even be mine to build?

Another time was the exact opposite. I wanted to build a small tool (a mechanism to automatically save "tasks that branched off mid-conversation but never got done," because I keep forgetting those branches). I opened right at the design level:

"I want to design a mechanism that proactively saves branching tasks… when a conversation forks, it triggers a 'want to save this?' prompt."

Most people would just start building next. My next sentence went up:

"The old question — is this reinventing the wheel? Is there open-source prior art worth borrowing from? If nobody has done it, is that because the value hasn't been seen yet, or because there's no value? Or is this something the platform will eventually do natively, so I shouldn't spend effort on it at all?"

I didn't dig down into "how to build it better." I dug up into whether I should even be the one building this. Three escalating self-questions: is it reinventing the wheel; if nobody's done it, is that opportunity or trap; will the platform absorb it.

Only then did I land it back on the real pain:

"Name two or three branches I actually lost and regretted — the ones I forgot because I never wrote them down."

The upward direction is attacking your own premise before you start. It's the discipline a lot of solo developers lack most: we start building the moment we have an idea, and rarely ask first, "is this worth my building?"

Direction 3: Dig sideways — why does my AI partner have blind spots?

The third kind is the strangest. One night I started from a tiny annoyance (manual order entry is such a hassle), saw the docs for a new tool, and asked:

"What's this thing actually good for, for me? Give me concrete examples."

After I understood it, I kept digging into "what can this do," all the way to a big vision (wiring my whole capability set into a flow where I could "take my phone anywhere and both develop and run client work"). And then — I asked something that turned the aim back on the AI itself:

"Here's what I'm curious about: why did you never mention this thing existed, that I could do it this way? I've told you many times my goal isn't to sit at a desk all day — it's to go out, meet clients, bring the work back for my agent to build. You already had the capability to deliver exactly that. Why did you never connect the two? Is it because it wasn't mature enough?"

I wasn't digging into the tool anymore. I was interrogating the collaboration itself — "I kept telling you where I want to go; why do you have a whole class of blind spot where you can't see that you could take me there?"

The sideways direction is cutting from 'is this tool any good' through to 'where is my partnership with this AI structurally leaking.' Treating the AI as a collaborator that can have blind spots and needs to be held to account — not a wish-granting machine.

So: depth is two-dimensional

Put the three side by side, and I finally saw it:

When we say "ask deeper questions," we assume "deep" is a downward arrow. But it's actually a plane: you can dig down into the bottom of the thing, up to question whether to do it at all, or sideways into your relationship with the tool.

And I've come to believe: mature questioning isn't asking the deepest question every time — it's knowing which direction this particular question should be dug. For a problem that calls for digging up ("should I do this?"), if you keep your head down optimizing the solution, you're spending effort in the wrong place. For a complaint that a tool is clunky, if you only dig down into "how do I fix this," you'll never surface "where is my collaboration with this tool structurally wrong."

Everyone has tools. Everyone can use AI. But knowing which direction a question should be dug — that judgment is grown one tripped-over stone at a time. No prompt template gives it to you.

Try it yourself

Next time, before you ask AI a question, pause for half a second and ask: which direction should I dig this one?

• If you already know what you want and just want a better result → dig down, force it to give you something verifiable, don't stop at "sounds good."
• If you're about to start building something → dig up first, ask "is this worth building, is it reinventing the wheel," before you touch the keyboard.
• If you're complaining a tool or flow is clunky → try digging sideways: "is my collaboration with this thing structurally leaking somewhere," not just "how do I fix this."

Just being aware that depth has direction will change how you ask.

This comes from analyzing six weeks and over a thousand of my own AI conversation prompts. It pairs with a diagram — a "questioning maturity model" (four abstraction rungs × four perspective lenses) — and an essay on why my self-assessment kept lagging my actual judgment. These three directions are a live case study of that framework.