Prompt Engineering from First Principles

Introduction

Prompting works because of one fact about how language models operate: they generate text by predicting what plausibly comes next, given everything in the conversation so far. Your prompt is that "everything so far." A vague prompt leaves millions of plausible continuations, so you get a generic one. A specific prompt narrows the field until the most plausible continuation is the answer you actually wanted. Understand this, and every prompting tip stops being a trick and starts being obvious.

What you will build or practice

You will build a mental model — the prompt as a specification that narrows predictions — and use it to explain why real prompts succeed or fail before you even run them.

Before you begin

You need any AI chat assistant. It helps to have read Large Language Models Explained Simply, which covers what these models are; this lesson covers what that means for how you use them.

Key concept

Garbage in, garbage out — but for context. The model has no access to your intentions, your project, or your standards. It only has the words you provide plus patterns learned from training text. Everything you leave unsaid, the model fills in with the most statistically ordinary assumption. Vague output is rarely the model failing; it is the model accurately reflecting an underspecified request.

Step 1: See the model as a continuation machine

Ask an assistant simply "Tell me about bears." The reply will be an encyclopedia-style overview, because that is the most common way text about bears continues in the model's training data. Nothing went wrong — you gave it a context that fits thousands of continuations equally well, and it picked the middle of the road.

Step 2: Watch specificity narrow the predictions

Now add context and watch the continuation space shrink.

I run a campsite in Colorado. Write 4 plain-language rules for guests
about storing food so bears are not attracted to tents.

Every added detail — campsite, Colorado, guests, food storage, four rules — eliminates continuations that no longer fit. Specificity is not politeness or ritual; it is you deleting wrong answers before they are generated.

Step 3: Treat the prompt as a specification

Engineers do not ask a builder for "a good house"; they hand over a spec. Apply the same standard to prompts: if two reasonable people could read your prompt and picture different outputs, the model can too, and it will pick one you did not intend. Examples work for the same reason — one sample output is a compressed specification of tone, length, and structure that would take a paragraph to describe. Explicit output formats work because "give me a numbered list of five items" is a spec the model can match, while "keep it organized" is not.

Step 4: Use the model to predict failures

Before running a prompt, ask: what am I assuming the model knows? If your prompt says "make it better," better by whose standard? If it says "summarize this," for whom and at what length? Each unstated assumption is a slot the model will fill with an average. Filling those slots yourself, in advance, is the whole discipline — the element checklists in Prompt Engineering Basics are just organized lists of common slots.

Practice exercise

Take a prompt you have actually used that gave a mediocre result. Write down three assumptions it left unstated (audience, format, purpose, tone, length). Rewrite it with those slots filled and compare outputs. Note which single addition changed the result most.

Common mistakes

• Believing there are secret magic phrases, when clarity is what actually moves results.
• Blaming the model for vagueness the prompt permitted.
• Adding irrelevant context — narrowing only works when the details relate to the task.
• Forgetting the model cannot see your screen, files, or intentions unless you put them in the prompt.

Check your understanding

1. Why does a vague prompt produce generic output even from a capable model?
2. In what sense is a prompt a specification?
3. Why do examples steer output so effectively?

Key takeaways

• Models predict continuations from context; the prompt is the context you control.
• Specificity works by eliminating unwanted continuations, not by persuading the model.
• Every unstated assumption gets filled with a statistical average — write the spec yourself.

Next steps

Put the model to work with the hands-on loop in Prompt Engineering, or learn a structured checklist in The CLEAR Prompt Framework.

Related resources

• Generative AI Explained for Beginners
• Prompt Engineering Cheat Sheet
• Verify AI Answers

Newsletter or next lesson

Join the newsletter for a clear weekly lesson, or continue with the AI Beginner Roadmap.