Reading
The first questions everyone asks about AI
Don't worry about the non-stop news, focus on where this is going and how to ride the wave
The core idea
AI isn't perfect — it's reactive, and it'll do what you ask. The problem is you often don't know what to ask.
That's because an AI chat is not a normal conversation. It's a meta-conversation — a conversation about a conversation.
That will make more sense as you read on.
The skill you need to master here is learning to invoke AI on a whim, to get an instinct for how it responds to you, and to gain a natural fluency to bridge the gap between your thoughts and AI's actions.
My advice is intentionally simple and vague, for a good reason: the tools will change. What you need to build is a fundamental philosophy and world-view about what is happening right now and how to operate in a frontier you cannot fully comprehend yet.
1. "I want to get into AI but I don't know what to do"
Answer: Get the AI to teach you the words for what you are asking it to do.
You likely already know the long-form version — you have multiple sentences and framings that describe what you want. But you don't have the single word or phrase that captures all of it.
So you task the AI with information compression: take your multiple descriptions of intent and compress them down to a single term, sentence, or direction.
Example
Sometimes you are not missing a summary — you are missing a name. You have been circling the same idea across a long thread, and it still feels fuzzy. So you ask the model to do the vocabulary work for you:
Context
Prior messages in the thread were long — you still didn't have a compact word for what you were circling (the critical path idea, before it had that label).
Is there a word for this? We're kind of orbiting around a concept and a practice.
That user message is a question about the current conversation — which is one kind of question you ask inside a meta-conversation (a conversation about the conversation).
A tight name for what we've been orbiting: critical path — or in the same neighbourhood, main effort / rate limiter.
Commentary
In the real thread, the model had the conversation plus workspace context — e.g. a doc it could read — and treated your question as: give me a handle for what we've been circling. The exact words vary every time; the structure is the point — which is why the bubble above is only an example, not a transcript.
Takeaway
You supply the long, messy intent; the model compresses it into a name you can reuse — including in conversations with yourself.
2. "How do I get started? I don't know how to set things up"
Whether you have zero programming experience or twenty years of it, the first unknown is the same: where do I work?
Step 1 · Unblock your environment
Set up a place where you can work daily without blockers. I used a mono-repository: one parent folder for everything. Open one instance of your AI-coding software in one place — no context-switching to another folder.
Step 2 · Unblock your thinking
Write down every single idea, immediately, into a README.md. Even the dumb ideas — they free cognitive room in your brain so you can move on to new ideas. After 10, 20, 100 ideas you will have practised a new skill: ideation. Let the AI give brutally harsh critiques.
Takeaway
Environment removes friction. Writing removes fog. Do both before worrying about what to build.
3. "I'm writing down ideas — but what should I ask the AI?"
You need another skill: meta-conversations.
To answer your question directly: "What should I ask?" is the wrong framing. The better question is "How should I engage in a conversation with AI?"
Here is a single thread that walks through three moves — framing, pushing, and reframing — inside one conversation:
Context
Move 1 · Framing. If you don't provide a frame in your first message, you get a generic reply. Set the contract up front.
I have an idea I want to discuss. I need a way to build a solution that solves for XYZ — but don't just agree with me. Push back, challenge my assumptions, and critique me harshly. The idea itself isn't valuable; the use-cases downstream are valuable.
(The reply tends to match that contract: real pushback, explicit trade-offs, pressure on whether downstream use-cases hold.)
Move 2 · Push harder. You got a response — now steer it deeper.
Good — but I need us to dig deeply into the fundamental foundations and reason our way back up to the use-case. Dig so deep we touch bedrock and justify every aspect of the design.
(Denser, more grounded answer — sometimes including the model faulting its own previous reply as too shallow.)
Move 3 · Reframe for another stakeholder. Grant the current argument and look at it from a different angle.
Now package this from a user's perspective and create a pitch that delivers them value. Anyone can have the idea and the code; those are not the value. The value is what problem this solves and for whom.
(The thread shifts to buyer- and user-language: problem, audience, outcome — aligned with your proposal from two major stakeholder angles.)
Takeaway
Frame → Push → Reframe. Three moves inside one thread. Each one is a meta-conversation move: you are steering the conversation itself.
The loop
Those steps turn "dumb ideas" into useful feedback and train you to have meta-conversations with AI. The entire practice is a cycle:
Repeat. Each cycle sharpens both the idea and your skill at steering AI. The ideas get less dumb. The conversations get more useful.
What to do right now
- Open a folder. Put everything in one place — a mono-repo, a single project, whatever. Remove the friction of "where does this go?"
- Write down your first idea. It doesn't matter how dumb it is. Get it out of your head and onto the page.
- Open a conversation with an AI and frame it. Tell it what you want, how you want to be challenged, and what success looks like. Then push harder.
For architecture reviews, consulting, or demos: aia.works/call-me-back