TL;DR - A strong prompt has four parts: context, instruction, input, and output format. Add them and unreliable results become repeatable ones.
Why it matters
The exact same model can hand you junk or gold depending only on how you ask. "Rewrite this" gives a random paragraph; a structured prompt gives the email you'd actually send. Because every other AI skill - writing, research, automation - runs on prompts, this is the highest-leverage thing you can learn. In studies, structured prompts have cut task time dramatically and raised output quality across the board.
The four parts
- Context - who the AI is and the situation. "You are a support agent for a calm, premium brand."
- Instruction - the one clear task. "Reply to this unhappy customer."
- Input - the raw material to work from.
- Format - the exact shape of the answer. "Under 80 words. Apologize, give the fix, give a timeline."
Worked example
Weak: write about onboarding
You get a generic 400-word essay on why onboarding matters. Useless.
Strong - paste this and you'll see the difference:
You are an HR lead at a 30-person startup (context).
Write a welcome email for a new designer starting Monday (instruction).
Details: their buddy is Sam, kickoff is 10am, bring a laptop and ID (input).
Format: warm but brief, under 120 words, end with one clear next step (format).You get a near-final email that needs a 5-second tweak, not a rewrite.
Steal this template
Reuse this for almost anything - just fill the brackets:
You are [role / expertise], helping with [situation].
Task: [the one thing you want].
Input: [paste your material].
Format: [length] | [structure] | [tone].Common mistakes (and the fix)
- Burying the instruction in a wall of context. Fix: put the task on its own line, near the top.
- No format. "Make it good" isn't a format. Fix: name the length, structure, and tone.
- One giant prompt. Fix: get a draft, then refine with one small nudge - "shorter", "warmer".
Good to know - this works everywhere
The four-part structure is model-agnostic: it works the same in ChatGPT, Claude, Gemini, and Copilot. A few notes: Claude and ChatGPT both reward a clearly stated role; for long inputs, paste the material last and restate the instruction at the end. You learn prompting once and carry it across every tool.