How to use AI, and keep your judgment.
Short answers to the questions people ask about using AI, from one person’s essays and practice.
These are short answers to the questions people ask about using AI. Each one comes from an essay on this site, or from how the essays and experiments here are actually made, and links to the longer version. They are one person’s answers, not a manual: Van writes the essays and builds the experiments, and every answer below is something argued in one or done in the other.
The short answer
- Bring it something you already care about, and ask for the pieces rather than the finished thing.
- Keep the judging: decide for yourself what is worth expressing.
- Check what actually happened, not the status it reports.
- Don’t let what is newly possible decide, by itself, what is now required.
- Stay close to the decisions that matter, and notice who claims the time it saves.
- Keep people near your work, even when the work no longer needs them.
Getting started
How do I start using AI?
Start with something you already care about but don’t yet know how to do, and ask for the pieces rather than the finished thing. In The Curiosity That Survived, Van loved three songs and wanted to bring them together without the music theory to do it. Asking an AI what to borrow from each, then for chords and a scale to explore, kept the curiosity alive long enough to pick up the guitar. An app that generated the whole song ended it instead.
I can bring AI an intuition before I’ve learned how to turn it into a question.
Why does using AI feel overwhelming?
Because the line between what you can and can’t do keeps moving, faster than you can redraw your picture of it. When a friend of Van’s first tried a coding agent, the question stopped being what the tool could do and became what else he had been wrong about. The essay calls the feeling possibility vertigo: it doesn’t mean every direction is worth pursuing, only that the old map can no longer tell you how far you can reach.
How often has “I don’t know how” quietly become “I can’t”?
How do I show AI to someone who doesn’t see the point?
Don’t lead with the most impressive demo. To someone whose craft it imitates, “Look what it can do” arrives as “you can get something like this without putting in all the time and effort you did”. Once a demonstration puts them on the defensive, they stay there for the rest of the conversation. Show them instead how it kept one of your own curiosities alive long enough to do something with it.
We’ve mistaken the most spectacular demonstration of AI for the most persuasive one.
Writing, reading and thinking
Can I use AI to write if I can’t write?
Yes, but the judging has to stay yours. AI hands you a version that looks almost finished in seconds, and a finished-looking draft doesn’t tell you whether you are rejecting the model’s attempt or abandoning your own idea. I Can’t Write tests itself by counting: five of thirty-one published experiments never opened, and two more that went live without any oversight.
It isn’t my guide for deciding what is worth expressing. I want to make that judgment myself, even when the model contributes the idea I’m judging.
ReadI Can't Write
How do you write with AI and still say what you mean?
Let the tool do the refining, and keep the attention for yourself. Each essay on this site starts as a question and a pile. Van explores the subject with ProseMap, a writing tool made in this workshop, which maps the topic out, keeps the passages that hold up and works on the prose, draft by draft, until the sentences say what was meant. Most essays show how many words were considered and what share was kept.
The work is still attention — deciding what deserves a second look and what does not.
How do I get more out of long AI answers?
Read slowly and send back only what matters. Assay, a free browser extension made here, lets you tap a word in a ChatGPT or Claude reply, widen it to the sentence or paragraph, collect a few, add a note, and send them back as one prompt. Nothing leaves your device.
When somebody actually thinks with a model, most of the work is editorial — reading a long answer and deciding which three sentences deserve a second look…
ReadAssay
Why are AI explanations sometimes impossible to understand?
Because the longer you work with a model, the more precise and compressed its vocabulary can become, when what you need is the long version. In Tiny Phrase, asking Claude where a phrase came from got an explanation of what it meant instead, in words that needed explaining themselves. One reply Van didn’t want to translate stopped what was maybe one of the more important projects Van had come up with; it hasn’t been touched since.
An answer to what it means, when you asked where it came from, is incorrect by nature, no matter how precise the answer is.
ReadTiny Phrase
Should I let AI connect my ideas for me?
It can, and the result can feel uncannily like you: two of the experiments on this site began when a model pulled old, loose interests together. The cost is that every connection it makes for you is one you will never make first. Orphaned Curiosities calls that the payment.
Every time the system assembles these pieces for me, I see the possibility sooner, but I lose the chance to be first.
Building with AI
Do I need to understand the code an AI agent writes?
Not all of it — some generated code may never need a human to understand it. But sometimes you must stop and look: before a change crosses a domain boundary, when an agent invents a persistent dependency, when part of the system starts taking on new responsibilities, or when the architecture changes shape without anyone deciding it should. Tests check what the system does, not what it should be.
Every change can be correct. The system they produce can still be wrong.
Can I build software with AI from my phone?
Yes, if the phone is only the screen. Van built and shipped a small puzzle game from a phone by typing prompts to a coding agent that ran in a persistent session on another machine, with the whole project, its tests and its deploy pipeline in context, then testing each preview on the phone. Idea to live took about 30 seconds per change.
My phone is not the machine. It is the interface.
What can one person build with AI?
The experiments on this site are one answer: tools, games and simulations, including the writing tool and the browser extension the essays are made with. They are built from Van’s own ideas, with AI doing most of the building; the site, the experiments and most of the plumbing are written with Claude Code. Each is put in front of strangers to see what happens, and gets a page with a picture and a note when it ends.
Work and skills
Will AI make my skills redundant?
Some of what you became good at may stop being scarce, and that loss is real. What the years built in you — sitting with confusion, taking big problems apart, distrusting elegant answers that don’t survive reality — comes with you into whatever is next. The essay’s advice for Monday is unromantic: adapt aggressively, use the machines, stay close to consequential decisions, protect your income, and don’t make the next moat your identity.
When the machine saves you time, notice who claims the saved time.
Should I use AI for everything it makes possible?
No. Some standards should rise — there is no good excuse left for guessing at something you can check in seconds. But a capability becoming cheap doesn’t, by itself, make it required. Before something becomes your new minimum, ask what would justify it.
Yes, I could have done more. No, that fact alone does not establish that I should have.
ReadThe New Minimum
If AI lets me work alone, should I?
You can, and some of the old dependence is worth losing. But shared work used to bring trust and friendship with it, unscheduled, and when the work no longer needs anyone else, the chance for that has to be kept on purpose. Van now shares a table with a friend from another company, and the old byproducts have started arriving again, uninvited.
The danger is confusing optional with unnecessary.
Should I save my AI usage quota or use it up?
Both can go wrong, because the quota resets whether you use it or not. A week of saving it can end in inventing work just to burn what is left, and running out midweek can leave you unable to work. In The Quota, as the backlog ran low, the last job got the highest effort and was of the lowest importance.
The last job got the highest effort, but was of the lowest importance.
ReadThe Quota
Agents and automation
Should I let AI agents run on their own?
Only knowing what you are handing over. Once you tell an agent to figure out the rest, it decides what success means, and checking its work tends to decay — from inspecting everything, to sampling, to a signature on a page nobody reads. Van’s own loop was small enough to stand outside and watch where it stopped, but a system doesn’t have to outgrow its limits for its limits to outgrow ours.
The judge remains human until the agent writes the test.
Can I trust an AI automation that says it succeeded?
Check what actually happened, not the status. In Completed “successfully”, a scheduled routine reported success every time it ran for two weeks while doing nothing, because it couldn’t reach the repository it was meant to maintain. The work was only alive when Van touched it.
The corpus was supposed to be maintained and kept alive by the machines. But it was in fact only being kept alive by my interaction.
What happens if we let AI keep working without us?
Probably nothing dramatic on any single day. The essay argues that stepping out happens a little at a time, every time it is cheaper, faster or just easier to let something keep going without us — and that the real question is whether we would still be standing somewhere we could tell.
The day the whole world went to sleep won’t arrive as a day.
Making sense of it
What is AI, really?
You may not need to settle that before you can talk about it usefully. Where in the World Are We? suggests treating AI less as a new someone than as a new somewhere: describe what you find when you use it, compare notes with people who use it differently, and start a shared map even while you disagree about what it is.
Common geography before common philosophy.