Prompting Is Just Writing Clearly (Sorry)

I'm going to save you a lot of money on courses with one sentence: being good at prompting is mostly being good at *writing clearly*. That's it. That's the secret. The forbidden incantations are just well-specified requests in a trench coat.

The model isn't a search engine and it isn't a genie waiting for the magic phrase. It's more like an extremely fast, extremely well-read new hire who is eager to help, has zero context about your situation, and will absolutely run in the wrong direction if you're vague. Your job is to brief them well.

Watch what specificity does. Here's the prompt everyone writes first:

You'll get a beige Wikipedia summary nobody asked for. Now the same intent, actually specified:

Same model. Wildly different output. You didn't unlock a secret mode — you just stopped making it guess. Every ambiguity you leave is a decision you've handed to a coin flip.

Most strong prompts have the same skeleton, whether you write it out formally or not:

• Role — who the model should act as ("senior security reviewer").
• Context — the situation and any constraints it needs to respect.
• Task — the one specific thing you want done.
• Format — exactly how the output should be shaped.
• Examples — a sample or two of what good looks like.

1. Role + context

Tell it who it is and what the situation is. "You're a senior security reviewer looking at a payments API" primes it far better than a cold question, because it narrows the universe of plausible answers to the ones an expert would give.

2. Specify the output shape

If you want JSON, ask for JSON and describe the fields. If you want a table, say so. Ambiguity about *format* is where most "the AI is bad" frustration actually lives — the content was fine, the shape was wrong.

3. Show, don't just tell

One or two examples (few-shot) often beats a paragraph of instructions. The model is a world-class pattern matcher — give it the pattern and it locks on instantly.

4. Let it think

For anything involving reasoning, "think step by step before answering" measurably improves results. You're giving it room to work instead of demanding the answer leap out fully formed. With reasoning-tuned models this is increasingly automatic, but a nudge still helps on the hard ones.

5. Give it an out

Vague prompts pressure the model to invent. Explicitly permit "I don't know" and you trade confident fabrication for honest uncertainty — almost always the better deal in a real product.

Nobody nails the prompt on the first try, and the people who look like they do are just hiding their drafts. Treat it like a conversation: see what came back, notice where it drifted, tighten that exact part. The loop *is* the craft.

1. 01Write the obvious version of the prompt.
2. 02Run it. Find the one thing that's wrong.
3. 03Add a constraint or example targeting *only* that thing.
4. 04Repeat until boring. Then save it.

Prompts in production deserve the same rigor as code. Version them. Keep a few golden test inputs and check the output every time you tweak the wording — models and prompts both drift. A prompt that worked beautifully last month can quietly regress, and you want a test to catch it, not a customer.

So no, there's no secret phrase. The good news is the skill transfers from something you already do every day: explaining what you want to a smart person who can't read your mind. Get clear, give examples, iterate. The model was listening the whole time — you were just mumbling.