Why doesn’t AI always do what I want. Question.

I asked Claude to “maybe take a look at renaming the variables in this file” and it renamed every variable in the file. That was Claude’s fault, sort of. Mostly it was mine.

Modern AI doesn’t just answer anymore. It does things. Claude Code edits your repo. Claude in Chrome clicks buttons. ChatGPT’s agent mode books tickets. The polite English construction “could you maybe…” used to be a request softener; now it’s a loaded firearm with the safety off.

The convention

If you haven’t read the book (or seen the 2026 film, in which Ryan Gosling shares the screen with a five-legged puppet and somehow makes it work), the setup: human astronaut Ryland Grace and Eridian engineer Rocky have to save both their species from a planet-eating space mould. They speak different languages. Rocky’s language has no rising intonation, so questions and statements sound identical to him. Their fix is brutally simple — tag every utterance.

“Your ship is broken, question.”

“My ship is fine, statement.”

Grace adopts it back. Within fifty pages it stops feeling weird and starts feeling like the most efficient communication protocol ever invented. Two words. No ambiguity.

Why this works on AI

Tag every prompt.

Refactor the auth middleware to use cookies instead of headers. Statement.

What would happen if I refactored the auth middleware to use cookies. Question.

The first is a command. The second is research. They are completely different requests with completely different outputs, and “could you refactor the auth middleware” splits the difference badly enough that you usually get both — a half-explanation followed by an aggressive code change that you didn’t ask for, plus an offer to write tests you also didn’t ask for.

Tagging the prompt forced me to decide, before I hit enter, which one I actually wanted. That turns out to be the whole game.

The honest part

Here’s where I tell you the magic words aren’t magic. Claude doesn’t parse “Statement” as a special token. There’s no hidden mode. If you removed the word it would behave roughly the same.

The technique works because it disciplines you, not the model. You cannot write “Statement” at the end of “could you maybe take a quick look at…” without noticing how mealy that sentence is. The convention drags vague intent into the open and forces a commit. Half of bad prompting is people not being honest with themselves about what they actually want, and Rocky’s convention is a surprisingly elegant cure for it.

Most prompt engineering advice fails for the opposite reason — it’s a ritual cargo-culted from a paper, performed on the model in the hope that the model will reward you. CRISPE this, chain-of-thought that, “you are a world-class senior staff principal engineer.” The Rocky protocol points the other direction. The work is on your side of the keyboard. Statement.

Disclosure

This post was drafted with Claude, edited by me. Statement.

Motivated by recently watching an AI confidently update something I asked about, and deployed to staging. The irony was not lost on either of us. The variable renaming incident at the top was real. The “safety off” metaphor was Claude’s and I left it in because it was better than mine.

If you try this and it changes anything for you, I’d be interested to hear. Question.