Using AI properly

Workshop · 7 of 7 · Verify · Bound · Approve

The model is a brilliant colleague with no stake in being right. It writes with the same confidence when it knows and when it guesses, and it will do exactly what you asked even when you asked for the wrong thing. Working with it properly comes down to three habits.

Verify — trust is earned by receipts

Never ship a fact you didn't check, a number you didn't trace, or a claim whose source you can't open. Good practice is not “trust the AI” or “distrust the AI” — it is knowing which kind of output you're holding. Drafts, structure, code you'll test, options to consider: cheap to accept, cheap to fix. Facts, figures, quotes, legal and medical anything: verify before it leaves your hands. Ask for sources and confidence; the word “unverified” from an honest system is worth more than a confident answer from a flattering one.

Bound — decide what it may touch, before it touches anything

Before you connect an AI to your files, your inbox, your accounts, decide its limits on purpose: what it may read, what it may write, what it may never reach. Reading is cheap; writing is recoverable; sending, deleting, and paying are not. The boundary you set when calm is the one that protects you when you're rushed. If a tool cannot show you its permissions, it has permissions you don't know about.

Approve — keep the irreversible behind your yes

Nothing irreversible happens without your explicit approval. Emails go out as drafts you send. Purchases wait for your confirm. Deletions ask twice. This is not distrust of the machine — it's the same discipline you'd want from a human assistant on their first week, held permanently, because the cost of the rare mistake outweighs the seconds of the routine click.

These three habits are our own operating law, not just advice: every agent behind this site publicly declares what it may do, prohibits itself from sending anything autonomously, and puts every irreversible act behind approval. Check it at /agents/index.json. Hold every AI you use — including ours — to the same standard.