Check what it says.
Trust a fluent answer after a check. A sum done by a tool, a claim against the line it rests on, and a correction written where it lasts.
new words
- Fluent
- Reads well. A model writes fluently whether it is right or wrong.
- Tool
- A command the AI runs, such as a calculator or a search, whose result is exact.
- Correction
- A fix to what the AI did, and to the instruction that made it do so.
the lesson
A model writes the likeliest next words, and the likeliest words read well whether they are true or not. So how an answer reads tells you nothing. Three checks do:
| The answer has | Check | Why |
|---|---|---|
| A number it worked out | "Do it with a command and show me the command" | A model reads numbers as pieces: How a model works, lesson 2 |
| A fact about your work | "Which file, which line?" Then open it | A quoted line is the source; the rest is a claim: lesson 5 |
| A mistake | Fix the instruction file, not only the chat | The next window makes it again otherwise: lesson 4 |
Ask the same question twice and the words can differ: the pick of each next token has chance in it, How a model works, lesson 6. A fact you will act on is one you have seen in its source.
With instibrain, each version of a page names its writer, and a page comes back from an earlier version: instibrain with Claude Code, lesson 5. Without it, keep your files in a folder that keeps versions, and read the change before you accept it.
try it
Ask your AI something about the folder you are in, then type this and press Enter:
Which file and which line says that? Quote the line.
you should see
- A path and a quoted line, which you open and find.
- Or an honest "I inferred it", which tells you the answer is a claim until you check it.