The 4-part prompt that actually works
Module 02 · est. 20 min · You’ll build: a repeatable prompt structure you can apply to any task.
TL;DR: Most “bad AI output” is really a bad prompt. A reliable prompt has four parts: Role, Context, Task, Format. Miss one and you get vague, generic, or off-target results. Nail all four and Cowork feels like a sharp employee.
[Operator’s read] I don’t write clever prompts. I write complete ones. The difference between a useless answer and a great one is almost never the model — it’s whether I told it who to be, what it’s working with, exactly what to do, and what the output should look like.
The four parts
- Role — who Claude should be. “You’re a careful copy editor.” “You’re my operations assistant.” This sets tone and standards.
- Context — what it’s working with and why. The files, the audience, the goal, any constraints. This is the part people skip, and it’s the most important.
- Task — the exact action, ideally numbered. “1. Summarize. 2. List risks. 3. Suggest next steps.”
- Format — what the output should look like. “A markdown table.” “Under 150 words.” “Bullets, no preamble.”
Before / after
Before (vague):
Summarize this report.
You’ll get a generic wall of text.
After (4-part):
You’re my analyst (role). This is our Q2 marketing report; I’m presenting it to non-marketers on the leadership team (context). Do this: 1) summarize the 3 things that worked, 2) the 2 that didn’t, 3) one recommendation (task). Format: short bullets, plain language, no jargon, under 150 words (format).
Same model, completely different result.
The one habit that matters most
Give context before you give the task. “Here’s what this is, who it’s for, and what I care about” — then the instruction. Cowork can’t read your mind; it can only read your prompt. Front-load the context and the quality jumps.
Lab
Take the task you did in Module 01 and rewrite the prompt with all four parts explicitly labeled. Run it. Notice how much less you have to correct. Save the good version — in Module 05 you’ll collect these into a reusable library.