People often blame the AI for “weak, generic answers” — but the real culprit is usually the prompt. The vaguer your instruction, the more it can only give you confident filler.
The good news: writing great prompts isn’t magic. One formula covers most cases — Role + Context + Task + Format. Fill in those four and the same model gives dramatically better answers. This is the foundation for every other AI tip, so it’s worth getting right.
When to use it
Any time the answer disappoints — too generic, off-topic, wrong format — check whether these four parts were actually spelled out.
How to do it
- Role: have it play an expert (“senior HR / primary-school teacher”) so it pulls the right knowledge and tone
- Context: a line or two — who you are, who it’s for, any constraints
- Task: replace “write something” with one clear action
- Format: say what the output looks like — table / bullets / email / word count
- Not happy? Follow up with “redo part X / more casual / half as long” to converge
Weak vs strong
Same request — the right version fills in role, context, task and format, and the output is usable as-is.
Copy-paste prompt
You are a【senior role】. Context:【who you are, who it’s for, any constraints】. Please【the exact task】. Requirements:【length / tone / audience】, output as【table / bullets / email】.
Worked examples
You get：A send-ready email: clear subject, respectful tone, a graceful out and a deadline.
You get：An explanation a child actually understands — not a wall of jargon.
Level up
- For accuracy: add one example in the context (“match this style: …”) — imitation is its fastest path
- For complex tasks: ask it to “outline first, don’t write yet,” confirm, then expand
- For reuse: save your common role setups in the app’s memory/custom instructions
Common mistakes
- Cramming many tasks into one prompt — do one thing per prompt
- Instructions with no context — it can only guess and stay generic
- Giving up after one try — great output comes from iterating, not the first shot
FAQ
Do I always need all four parts?
Does this work on DeepSeek, Doubao and ChatGPT?
Pro tip：Bookmark this as a template — run any prompt through these four parts first.