Prompting Fundamentals

← Previous: Tokens, Context Windows, and Why They Matter | Next: Embeddings and Vector Similarity →

We ended the last article with tokens costing you something real: money, speed, space on the model’s “whiteboard.” Naturally, that pushes toward being deliberate about what you actually write. Which is really the whole job of prompting: being deliberate with the input, because the output rides entirely on it.

There’s a persistent myth floating around that good prompting is some secret trick or magic phrase. I don’t buy it, and in practice I’ve never seen it hold up. Good prompting looks a lot more like good management: the exact skill you’d need to hand a task to a very capable but very literal new hire.

The Literal Intern Mental Model

Imagine you hired a brilliant intern. They’ve read almost everything ever written, they’re fast, and they’ll never complain. But they have one quirk: they will do exactly what you say, not what you meant.

Ask them “make this better” and they’ll genuinely try, but “better” could mean shorter, more formal, funnier, more technical, or completely rewritten. Without more direction, they’ll guess. Sometimes they’ll guess right. Often they won’t.

That’s an LLM. It’s not being difficult. It’s just working with exactly what you gave it, nothing more.

Zero-Shot vs Few-Shot: Two Ways to Direct the Intern

Zero-shot prompting means asking for something with no examples: just an instruction.

“Write a short product description for a stainless steel water bottle.”

The model has to guess your tone, length, and style purely from that one line.

Few-shot prompting means showing a couple of examples of what you want before asking for the new one.

“Here are two product descriptions in our brand voice: [example 1] [example 2]. Now write one for a stainless steel water bottle in the same style.”

flowchart TB
    A[Zero-shot: instruction only] --> B[Model guesses your style<br/>from scratch]
    C[Few-shot: instruction + examples] --> D[Model matches the<br/>pattern you showed it]

Few-shot prompts almost always produce more consistent, predictable results; you’re not describing the style you want, you’re showing it. That’s a much easier task for a next-token predictor than interpreting a vague adjective like “professional” or “punchy.”

Specificity Beats Cleverness

New prompt writers often hunt for a clever phrasing trick: some magic sentence that unlocks better answers. In practice, plain specificity beats cleverness almost every time.

Compare these two:

• ❌ “Explain data pipelines well.”
• ✅ “Explain what a data pipeline is to someone who has never worked in tech, in 3 short paragraphs, using a real-world analogy, with no jargon.”

The second prompt isn’t clever. It’s just specific about audience, length, structure, and constraints. That specificity does more work than any trick phrase ever could, because it removes the guessing the model would otherwise have to do.

A useful checklist when writing a prompt:

• Who is this for? (audience)
• What exactly do you want back? (format: a list, a paragraph, code, a table)
• How long should it be?
• What should it avoid? (jargon, fluff, a certain tone)
• Any examples you can show instead of describe?

Why Vague Prompts Fail Quietly

Here’s the part beginners miss: a vague prompt doesn’t usually fail loudly. It fails quietly: the model still gives you something, and it often sounds confident and polished. But it’s answering the version of your question it guessed at, not the one you actually had in mind. This is different from the model being “wrong”; it’s the model being unguided, and it will fill that gap with its best guess every time, because that’s exactly what a next-token predictor is built to do.

Mental Model

You’re not casting a spell: you’re delegating a task to a capable-but-literal intern who can’t read your mind and has no memory of your unstated preferences. The more clearly you brief them (audience, format, length, examples, what to avoid), the less they have to guess, and the closer the result lands to what you actually pictured.

Key Takeaways

• Prompting is really task delegation to a system that follows exactly what you say, not what you mean.
• Zero-shot prompts give an instruction only; few-shot prompts add examples, which usually produces more consistent results.
• Specificity beats cleverness: audience, format, length, and constraints matter more than a “magic” phrasing trick.
• Vague prompts don’t fail loudly: the model still answers confidently, just based on its best guess at what you meant.
• Good prompting sets up the next building block: once your inputs are clear, the model’s outputs become far more reliable, which matters even more once we bring in outside data, starting with embeddings next.

Next up: Embeddings and Vector Similarity, how text gets turned into numbers that capture meaning, and why that’s the foundation of search and RAG.