AI
From first prompts to agentic workflows, learned on Claude — AI as a real working partner, not a toy.
How It Actually Works
Mechanics before technique. Nearly every mistake people make with AI grows out of a wrong picture of what's inside — so replace the picture first, and the rules that follow stop looking like superstition.
- What a Language Model Actually IsThe model doesn't look an answer up — it rolls to one across a landscape worn in by millions of pages. Almost everything else about it follows from that.
- Tokens and the Context WindowThe model has no memory, only a desk. Everything it knows right now is lying on that desk — and the desk has edges.
- How the Model Was TaughtReading gave it knowledge; feedback gave it manners. Its most annoying habit — agreeing with you — comes from the second stage, and there's a fix.
- Why It Invents Things So ConfidentlyA hallucination isn't a malfunction — it's the same operation as a correct answer, minus the luck. Which is why confidence tells you nothing, and why you check by cost.
How to Talk to It
Not magic phrasings but four working habits: state the brief, feed the material, treat the first answer as a measurement, and leave the work somewhere to happen.
- The Anatomy of a RequestA weak answer is usually a weak brief. Every decision you leave out, the model makes for you — and it makes it by rolling to the average.
- Context Beats WordingTwo people send nearly the same request and get very different answers. The difference is almost never the phrasing — it's what was lying on the desk.
- The Dialogue Is the MethodThe first answer is rarely the one you want — and it was never supposed to be. It's a measurement that tells you exactly which part of the brief was missing.
- Hard Tasks: Give It Room to ThinkThe model computes as it writes — the text is its scratch paper. Most failures on multi-step problems come from demanding the answer before the work.
Trust and Verification
The stage that decides whether AI helps you or quietly costs you: where its ability actually ends, how to check without cancelling the savings, what must never be sent — and how not to hand over your own thinking along with the work.
- The Map: Where It's Strong, Where It Isn'tThe same person calls it brilliant on Monday and useless on Friday, and is right both times. The ability isn't uniform — and the edge is jagged, not straight.
- How to Check Without Losing the PointVerifying everything cancels the savings; verifying nothing eventually costs a lot. The way out is to stop checking answers and start checking the few points inside them that can be checked.
- What Not to Send, and What Not to TrustPasting is the whole convenience of the tool, which is exactly why so much ends up pasted that shouldn't be. And traffic runs both ways: text the model reads can carry instructions.
- Not Forgetting How to ThinkThe risk isn't that AI makes you stupid. It's that it removes exactly the difficulties that were doing the teaching — and the fix is a rule about ordering, not about abstinence.
Everyday Work
The four places this actually pays off — studying, writing, code and documents. Same mechanics as before, now as concrete moves you can use tomorrow morning.
- Learning with ItYou have a tutor available at three in the morning, infinitely patient, on any subject. Most people use it as a reference book — which is the one role where it teaches you nothing.
- Writing with ItAsk it to write the text and you get something readers recognise in three lines. The value isn't in generating prose — it's in everything that happens around the prose.
- Code with ItCode is the one field with an instant external judge: it runs or it doesn't. That makes the model strongest here — and most dangerous, because running and being correct are different things.
- Documents and ResearchA hundred-page PDF and three questions you need answered. This is the model working in its most reliable mode — provided you ask the document rather than the model.
Beyond the Chat
Tools, a permanent setup and agents that act rather than advise — and then the assembly: which of your tasks climbs which rung, and where to stop deliberately.
- A Model with ToolsSearch, code execution, files, connected systems — none of it changes what the model does. It changes what happens to the text it writes.
- Your Permanent SetupYou've explained who you are, what the project is and how you want it written twenty times now. The model has no memory — but the desk can be laid automatically.
- Agents: When It Acts Instead of AdvisingAn agent is the same model, the same tools, one thing added: the feedback loop closes without you. That's what makes it powerful and what makes it go wrong at scale.
- Building Your Own WorkflowNineteen articles were not nineteen rules — they were one way of seeing. This one turns it into a system you actually run, and marks the line you shouldn't cross.