The prompt decides more than the model
The same model, on the same question, answers with garbage or dead on target — the difference is in the prompt. Prompting remains the main skill for working with AI: all the advanced topics of the course (Projects, Claude Code, agents) stand on the ability to frame a task.
In its official guide, Anthropic gives the golden rule: show your prompt to a colleague who has no context for the task. If the colleague can't tell what to do, the model won't either. The model doesn't read minds, doesn't know your situation, and doesn't see your screen. It sees exactly the text you wrote.
Four elements of a strong prompt
1. The task — specific and direct
Weak: "Tell me about typography." Strong: "Pick a font pairing for a law firm's website: headings and body text. Give three options with a one-sentence rationale each."
Specifics — what exactly to do, at what length, with what result at the end.
2. Context — why and for whom
The model tunes its answer to the situation if you describe the situation: "I'm writing a cold email to a small-business owner who isn't tech-savvy; I have one shot at grabbing attention." Without context the model produces an average answer for an average person.
Worth stating: audience, goal, constraints, what you've already tried, what you're unhappy with.
3. Format — how the answer should look
Say it explicitly: a list or continuous text, length, language, tone, structure. "Answer with a table: columns 'option,' 'pros,' 'cons'" — and you get a table, not an essay. Don't specify a format and the model picks one itself, most likely longer than you need.
4. Role — from what position to answer
"You are a business editor with twenty years of experience. Edit the email below" — a role shifts the style, depth, and evaluation criteria. It works more gently than people tend to think (the model doesn't get smarter from the word "expert"), but it noticeably affects the angle and tone.
Order within a prompt: context and materials first, then the task and format. Put a long document at the start and its instructions at the end. That way the model reads the instructions already knowing the material.
A worked example
Weak prompt:
Write a post about our product.
Strong prompt:
You are a social media editor. Write a post for a Telegram channel about the launch of the auto-export-reports feature in our analytics service.
Audience — online-store owners; they value their time and dislike marketing hype.
Format: up to 800 characters, start with the pain (manual report exports every Monday), end with a link to the guide. No emoji in every line — two at most for the whole post.
The second prompt is five times longer — and saves three rounds of edits. The time invested in framing the task comes back as the quality of the first answer.
Common mistakes
- A question with no task. "What do you think of my text?" → the model praises it and points out a couple of trifles. "Find the weak spots in the argument, be picky" → you get real work.
- Several tasks in one message. "Edit, translate, and suggest headlines" blends into mush. Big tasks — one at a time.
- Hidden expectations. You expect a formal tone and the model writes casually — because there wasn't a word about tone in the prompt.
Key takeaways
- The prompt test: would an outsider with none of your context understand the task.
- Four elements: a specific task, context (why and for whom), the answer format, a role.
- Materials go at the start of the prompt, instructions after them.
- One message — one big task.