Agents, in plain English.
You keep hearing the word agent. It sounds technical. It isn't. Here you don't read about one — you assemble one, piece by piece, and see what each piece changes. Drag, tap, and see for yourself.
So… what is an agent?
An agent is an AI you've given standing instructions to. You write down, once, who it should be and how it should help — and from then on it behaves that way every time, for anyone who talks to it. A little worker you describe in words.
If you can describe a job to a new colleague in plain English, you can write one. Below, you assemble the actual instructions, line by line, and see what each changes. The question to hold onto: when a reply is bad, whose fault is it — the AI's, or the instructions'?
Whose fault is a bad answer — the AI, or the ask?
The AI below is the same every time; only what you tell it changes. Start vague, add a piece at a time, and read what comes back. Does the ask explain the difference — or not?
Your ask
Tap a piece to add it. Each one tells the AI something it can't guess.
What the AI gives back
From a one-off ask to a reusable agent
That birthday request was a one-off — you asked once, you got an answer. An agent is the same trick, but written down once and pointed at a job instead of a single question — and it talks to strangers, not just you. So it needs a couple more pieces: the facts it must stick to, and the limits of what it's allowed to promise. Let's build a real one.
Build one, piece by piece
Meet a shoe shop that wants an agent to handle refund questions. Below, it has no instructions at all — so when a real customer writes in, watch what it does. Then add each instruction and watch what each one changes.
The agent's instructions
Add the pieces in order, or in any order — see what each one fixes.
A real customer writes in…
Which pieces actually changed the answer?
A pattern worth noticing: most agents — support assistant, writing buddy, study helper — are some mix of these seven plain-English pieces, aimed at a different job:
- Who it is — give it a role.
- Its job — what it helps with (and what it shouldn't touch).
- The facts it must follow — your policy, your prices, your truth.
- Its limits — what it must never promise or do.
- When it's unsure — say so, and hand off to a human.
- Its tone — warm, plain, brief.
- The answer shape — how the reply should look.
Honest note: the replies above are representative examples, not live AI calls — the move without the noise. Want the real thing, surprises and all? Run any version on a live model in chatty.
Now make it yours.
Two ways to take it further — run it for real, or dig into how it works.