Track: writing and thinking
Start here whatever your job is. Writing is where these tools are most obviously good, where mistakes are cheapest, and where the habits you build transfer to everything else.
There is a fork at the start of this track and it decides how useful the whole thing is: are you using AI to write for you, or to think faster? The first is what most people try and it plateaus quickly. The second is where it keeps paying, and it is what this track is mostly about.
---
Stage 1 — the fix-it stage
You have written something. It is not quite right.
This is the easiest win available and a lot of people never get past it, which is fine — it is genuinely useful.
What to try first. Paste in something you have already written and ask for one specific thing. Not "improve this" — that gets you a beige version of what you had. Ask for:
- "Cut this to 150 words without losing the point about X."
- "This paragraph is doing two jobs. Split it."
- "Where is a reader going to stop and re-read a sentence?"
- "Argue against this. What is the strongest objection?"
What goes wrong. You ask for improvement in general and get something smoother, longer, and less yours. The tell is that it now contains the words delve, leverage, robust or in today's fast-paced world. That is what "make it better" means to a model: make it sound like the average of everything.
When you have got it. You reach for it to answer a specific question about a specific piece, not to run a piece through it.
---
Stage 2 — the blank page stage
You have nothing yet.
What to try first. Never ask for the finished thing. Ask for the parts:
- "I need to explain X to Y. What are the three things they will not already know?"
- "Give me five different angles on this. One sentence each. Do not write any of them."
- "What is the question this piece answers? Give me four candidates."
Then you write it. The AI did the part that is hard because it is unstructured; you did the part that needs judgement and voice.
What goes wrong. Asking for a first draft and editing it. This feels faster and usually is not — editing somebody else's structure into your argument takes longer than writing your argument, and the result reads like it was edited rather than written. It also skips the bit where writing forces you to work out what you think.
When you have got it. You use it before you start and after you finish, and barely in the middle.
---
Stage 3 — the context stage
You give it something to work from.
This is the step that separates people who find AI mildly useful from people who find it transformative, and it is not a technique. It is just: give it more of your actual situation.
Compare "write an update about the project" with "here are my notes from the last three weeks, here is who this is going to and what they already know, here is the thing I am worried about telling them. Draft the update." The second one is a different tool.
What to try first. Take your existing notes — meeting notes, a document, a week of a journal — and put them in with the request. Ask for something that could only be produced from that material: "what did I actually spend this week on", "which of these decisions have I not written down anywhere", "summarise this for somebody who missed all of it".
What goes wrong. Two things. Sending too little, which is far more common — people ask a question in isolation when the answer depends entirely on context they did not include. And sending too much unstructured material, which buries the point; a model reading twenty pages will weight the middle less than the ends.
When you have got it. Your first instinct on a hard request is "what does it need to know", not "how do I phrase this".
This is also the whole idea behind AIOS: you write notes normally, and the context goes with the question without you assembling it each time.
---
Stage 4 — the reusable stage
You stop retyping.
You will notice you keep asking for the same shape of thing. The weekly summary. The tightened paragraph. The plain-English version of something technical.
What to try first. Write that request down properly, once — with the role, the audience, the format and the tone in it — and keep it somewhere you can find it. Then improve it whenever it disappoints you, rather than rewriting it from scratch.
A good saved prompt after a few months of small fixes is dramatically better than anything you would type off the cuff, and it is the closest thing to a durable asset in this whole field.
What goes wrong. Keeping them in a chat history you cannot search. They need to be somewhere you would actually look.
---
The tools
Any of the big chat interfaces will do everything on this page. Choosing a model covers picking one; for writing, the differences between the top few are smaller than the difference between a good request and a bad one.
Whatever you already write in. Most editors and document tools now have AI built in. It is worth using when the alternative is copying text back and forth, and worth ignoring when it starts suggesting sentences you did not ask for — that trains you to accept text rather than choose it.
Something that holds your notes. The stage-3 jump needs your own material to be somewhere it can be pulled from. Any notes app, a folder of markdown files, or AIOS if you want the context assembled for you.
---
Where it genuinely does not help
Anything that has to be true and specific about the recent past. It will produce a confident, plausible, wrong version of what happened in your project last month. Give it the notes or do not ask.
Your voice, unless you show it. Asking for "my style" does nothing. Pasting in three things you have written and asking it to match them does quite a lot.
Deciding what to say. It will happily generate the piece you should not have written.
---
Next: Research and learning is the natural companion to this one. Or go back to all the tracks.