“Question 4” – Into AI, asking the right questions: Making AI Useful Is a Conversation, Not a Prompt

The first three pieces in this series argued that AI is more useful than the public conversation usually admits, and less magical than the marketing suggests. Question 1 looked at what AI can help you investigate. Question 2 looked at protecting yourself from the AI-generated noise that surrounds you. Question 3 looked at getting an honest answer from a tool that is trained to flatter you.

Each of those questions has a specific answer. This question does not.

Question 4 is about how to actually use AI, day to day, in a way that produces work worth doing rather than work that just looks like work. The honest answer is not a single technique or a clever prompt. It is a set of practices that develop over time, with use, through paying attention to what the tool is doing and what it is failing to do.

That answer does not fit in one article. So this question has six(ish) of them.

The premise

There is a particular pattern that gets repeated whenever someone tries to use AI seriously for the first time. They type a question. They get an answer. The answer looks fine. They use it. Sometimes the answer is good. Sometimes it is wrong in ways they do not notice. Sometimes it is technically correct but completely misses the point they were actually trying to make.

The instinctive response to a disappointing AI output is to blame the prompt. Better prompt, better answer. Find the magic words. Learn the prompting tricks. The internet is full of advice about how to phrase your question so that the AI gives you exactly what you want.

Most of that advice misses the point. Prompts matter. The phrasing matters. But the underlying truth is simpler and harder to act on. Making AI useful is not a prompt-writing skill. It is a conversation skill.

That distinction is the centre of Question 4.

A prompt is a one-shot instruction. Type something, get something back, accept the output, move on. A conversation is an exchange. You ask. The AI answers. You read the answer critically. You push back where the answer is wrong. You refine where it is vague. You redirect when it has misread your intent. You stop when the work is actually done, not when the AI says it is finished.

The first version of this article was going to argue that point in three thousand words and stop. As I started writing the supporting material, the supporting material outgrew the article. There were too many practical skills, too many failure modes, too many specific working patterns that the conversation-not-prompt framing required to be useful rather than abstract.

So Question 4 became a family. Six(ish) pieces, not including this one. Each piece takes one practical aspect of the conversation-not-prompt argument and develops it in enough detail to actually be useful.

The shape of the family

The six(ish) sibling pieces, in publication order:

4(a) Setting Up AI So It Actually Knows Who You Are deals with the configuration layer. Most users never set up the AI properly before they start working with it. The default behaviour is generic because the AI does not know who you are, what you do, or what you need from it. Configuration is the practice of telling it. The piece walks through how to do that in the products that support it, and why the time spent on configuration pays back across every conversation that follows.

4(b) The Bandwidth Problem deals with cognitive load. AI as a working tool is most valuable when it addresses the gap between what users can think and what users can output in the time available. The gap is universal but variable. It widens for some users more than others. The piece argues that AI is genuinely useful for that specific problem and that the value generalises beyond the neurodivergent specifics that produce the sharpest example of it.

4(c) Different Tools for Different Jobs deals with tool selection. The public conversation treats “AI” as if it were one thing. It is not. Different AI products are different tools, with different strengths, different default behaviours, and different specific failure modes. Multi-AI orchestration, choosing the right tool for the right work, is a practical skill that most users never develop because the marketing tells them they only need one.

4(d) How to Recognise AI Slop deals with the most visible failure mode. AI used without configuration, conversation, or judgement produces output that is fluent, plausible, and substantively empty. The piece documents what slop looks like, why it is produced, and how to spot it. It includes five evidence specimens generated specifically for the article. The reader leaves with the diagnostic skill that the rest of the family assumes.

4(e) The Failures That Survive Good Practice deals with the harder problem. Configuration helps. Conversation helps. Verification helps. None of them eliminate certain predictable failure modes that AI produces structurally. The piece documents six categories of failure that survive good practice and gives the reader the diagnostic skills to catch each one.

4(f) The Long Game deals with maintenance over time. The relationship between a user and an AI tool is not static. The tools change. The user changes. The configurations drift. The piece argues that long-term AI use is a maintenance practice, not a one-time setup, and walks through the specific skills that make the maintenance work.

4(?) 2b or not 2b, that is a pencil… , closes the family. It is different. You will understand why when you get there.

Why this family exists

Most public AI commentary falls into one of two failure modes. It either oversells the tool (AI will revolutionise everything, you must use it, those who do not will be left behind) or it dismisses the tool (AI is overhyped, the output is unreliable, it will not last). Both responses miss the practical question that most working users actually face: how do I use this thing well, given that it is here, it is improving, and the work I have to do continues regardless?

The Question 4 family is an answer to that practical question. Not a complete answer. The honest version of the practical answer is that nobody has a complete answer yet, because the tools are too new and the working patterns are still forming. But the practical answer is more useful than the marketing one and more useful than the dismissive one.

The family draws on around eighteen months of regular AI use across multiple tools, in multiple contexts, with regular failure modes documented in real time. Some of the failures are mine. Some of them are the AI’s. The piece on slop contains specimens. The piece on bandwidth contains autobiographical material that grounds the argument. The piece on long-term use contains examples drawn from this very series being produced.

The recursion is deliberate. The argument that making AI useful is a conversation rather than a prompt is demonstrated by the way the family was produced. Each article was drafted in conversation with AI tools, edited through multiple passes, corrected when AI mis-narrated the working pattern, refined when AI produced inflation or slop. The artefacts in your hands are the working pattern, not just descriptions of it.

What this question is not

A few things worth being explicit about, because the public conversation has trained readers to expect them.

This is not advice about which AI product to use. The family discusses multiple products and the rationale for using different tools for different jobs. The specific products will change. The principle of choosing the right tool for the right work will not.

This is not a prompt-engineering guide. There are prompts referenced throughout the family, and a few are reproduced in full. The argument is that prompts are not the point. The conversational practice around the prompts is what matters.

This is not a piece about whether AI is good or bad. The family operates from the position that AI is here, it is being used, and the practical question is how to use it well. Whether you should be using it in the first place is a question for somewhere else.

This is not a piece about AI safety, alignment, or the broader societal implications. Those are real questions. They are not the questions this family is trying to answer.

What this question is

The family is for users who are already working with AI tools, or who plan to, and who want to develop a practical working relationship with the tools rather than treating them as magic boxes or as marketing claims.

The pieces are not exotic. Most of what they describe is recognisable practical work that any thoughtful user could develop through enough trial and error. The family compresses that trial and error into something more readable than the experience of accumulating it personally.

The intended reader is someone who is paying attention. Someone who has noticed that AI output is sometimes excellent, sometimes terrible, and often hard to tell apart at a glance. Someone who wants the diagnostic skill to know which is which. Someone who is willing to do the work of treating AI as a tool that requires practice rather than as a black-box service that should just work.

The work is not glamorous. It is practical, periodic, occasionally tedious, and quietly valuable.

A note before the family begins

The pieces publish in order over the next week, each and every day at 12 noon AEST fully automatically. Each piece can be read on its own. Together they form a more substantial argument than any one of them makes individually.

The reader who reads only one piece will find something useful in any of them. The reader who reads all of them will find that the argument compounds in a way that matters.

4(?) lands at the end. Whether it changes how you read the earlier six is up to you.

The family begins on the Tuesday, 19th May 2026.