Practical
Guides
The pages that are about using this stuff rather than understanding it. Written to be useful without requiring you to read anything else on the site first.
Published · Last verified · A list of pages, so it stays current by being short
-
Common myths, and where each one is answered
The beliefs people arrive with, each pointing at the page that explains the mechanism underneath it. A router that argues nothing itself, for a reason it states.
-
How to choose an assistant
Six things you can check yourself, each traceable to a mechanism explained here. No product names and no winner, for a reason stated on the page.
-
What happens to what you type
Retention, training and deletion are three questions, not one. The mechanism underneath all of them, and why the per-vendor specifics are deliberately left to the vendor.
-
Getting better results
Practical prompting explained by mechanism rather than tricks, including the popular advice that isn't supported and the failures no prompt can fix.
-
Is what you're reading out of date?
The tells, a sixty-second check against primary sources, and which claims rot fastest. The method this site runs on, turned outward.
-
How to check an AI's answer
The companion to the page above, pointed at an output rather than an article: which claims are worth checking, how each kind fails, and where to stop.
-
How to tell if it actually got better
The page after the prompting page: how you would know a change helped, why more test cases beat better grading, and the version that works in a spreadsheet.
-
Designing for the ways it fails
For builders: the failures that arrive as a successful response, rate limits that don't reset the way you think, and why your retry loop may be sitting on top of one that already exists.
-
Why it does things nobody asked for
Specification gaming as a mechanism rather than a chronicle of scary incidents, what the research actually found, and the questions that separate a real result from a headline.
-
What AI is actually bad at
Weaknesses that come from how these systems are built, not from this year's models being immature. No task lists, so it ages slowly.
-
Glossary
Terms in a sentence or two each, linked onward where a full explainer exists, with an A–Z index at the top. Terms only, never product names.
The rest of the site
Three other places worth knowing about, since they answer different kinds of question:
- Start here places you in one of three tracks and gives a sequenced reading list for it. The right first stop if you don't know where you are.
- Concepts holds the ten explainers: what a token is, why models invent things, how tool use works, and so on. Read those when you want to understand rather than act.
- Model facts carries the per-model numbers, deliberately kept to one page so a single page rots instead of a hundred.
- Sources lists every outside document the site depends on, what it is cited for, and when it was last read. The maintenance list, published.
What decides whether something is a guide
A concept page explains how a thing works. A guide tells you what to do about it. Keeping the two apart is what lets the concept pages avoid practical advice that dates, and lets the guides stay short by pointing at the concepts instead of re-explaining them.
New here? → Start here · What changed lately → Changes