Your First Real Prompt

A good prompt is a precise spec: say exactly what you want, in exactly the shape you want it.

Building with AI · Lesson 4 · 10 min read

You ask an AI to “summarize this meeting,” and you get a nice paragraph — but your app needed a list of action items it could store in a database. The model didn’t fail; your request was underspecified. The first real prompting skill isn’t magic words — it’s stating exactly what you want, in a shape a program can use. You’ll practice that now, against a check that either passes or doesn’t.

Vague in, vague out

A model gives you what you asked for, not what you meant. “Summarize this” invites a paragraph because that’s the most likely continuation of that request. If you need a specific structure — a list, a table, JSON with named fields — you have to say so, explicitly. The clearer the target, the more reliably the model hits it.

Specify the shape, not just the topic

A strong prompt names three things: the task (what to do), the format (the exact output shape), and any constraints (length, fields, tone). When the format is machine-readable — like “return a JSON object with keys X, Y, Z” — you also get a free superpower: you can check the output automatically. Either it has the right keys or it doesn’t. That’s the foundation of everything in evaluation later.

Weak: “Summarize this meeting.” → a paragraph you can’t store.
Strong: “From the transcript, return a JSON object with keys summary (string) and actionItems (array of strings). No prose outside the JSON.” → structured data your app can save and verify.

Turning a fuzzy ask into a checkable one:
1. Goal: capture a standup update as data.
2. Decide the exact shape: a JSON object with name, blockers (a list), done (true/false).
3. Write the request stating that shape explicitly.
4. Now you can verify the result: does it parse as JSON? Does it have all three keys? Pass or fail — no guessing. That’s exactly the exercise below.