← BACK TO THE HUB
✓ LEVEL CLEARED
THE FUNDAMENTALS · PILLAR 03

Prompting

Prompting is just asking well. It's also the single skill that separates people who get incredible results from AI from people who give up and say "it doesn't really work." Good news: it's learnable in an afternoon.

What a prompt actually is

A prompt is the text you hand the model. That's it. But a model can't read your mind, it only has the words you gave it. So the whole game is putting enough of the right context into those words that the model can't help but give you what you wanted. Vague in, vague out. Specific in, sharp out.

The anatomy of a good prompt

Almost every strong prompt has four parts. Miss one and quality drops:

ContextWho you are, who it's for, and any background the model needs. "I run a two-truck plumbing company in Reno" changes everything about the answer.
TaskThe one specific thing you want done. Not "help with marketing," but "write three subject lines for a furnace-tune-up email to past customers."
FormatWhat "good" looks like: length, structure, tone. "Give me a bulleted list, under 20 words each, plain and friendly."
ExamplesOne or two samples of the kind of output you want. Showing beats describing, every time.

Bad prompt vs good prompt

Same goal, wildly different results:

✗ WEAK
"Write me a social media post about my business."
✓ STRONG
"You're a friendly local expert. Write one Instagram caption for my Reno plumbing company promoting $89 furnace tune-ups before winter. Under 40 words, warm and a little funny, end with a clear call to book. Here's one I liked before: [example]."

The system prompt

A system prompt is a prompt that runs before every conversation, setting the ground rules once so you don't repeat yourself. When you give an agent a lasting identity, name, voice, job, rules, that system prompt becomes its soul file. Same mechanism, bigger job.

Give it a role

Telling the model who to be genuinely changes the output. "You are a patient teacher explaining to a beginner" and "You are a blunt senior editor" will answer the same question in completely different voices. Pick the role that fits the job before you ask the question.

Show, don't tell (few-shot)

The fastest way to get the format you want is to demonstrate it. Paste one or two examples of good output and say "more like these." This is called few-shot prompting, and it beats paragraphs of description because the model learns the pattern directly instead of interpreting your instructions.

The mistakes that quietly wreck results

Being vague, "make it better" gives the model nothing to aim at.
No context, the model doesn't know your business, audience, or goal unless you say so.
The kitchen sink, cramming ten unrelated asks into one prompt; do them one at a time.
No format, if you don't say what good looks like, you get the model's guess.
Giving up after one try, the first answer is a draft, not a verdict.

Iterating: fix an almost-right answer

The pros don't write one perfect prompt, they steer. When an answer is close, tell the model exactly what to change: "shorter, drop the second point, more casual." You're not starting over, you're nudging. Two or three nudges gets you further than any single mega-prompt.

Keep reading ▸
→ See how a prompt becomes an agent's soul file→ Turn great prompts into reusable skills→ Build your first agent, step by step