Start with the outcome
Most work prompts fail before the model sees them — because they do not describe the output. "Help me write an email about the delay" is not a work prompt; it is a wish. A work prompt names the output, the audience, the tone and the boundary of what the model should not touch.
Here is the honest difference between the two. A vague request makes the model guess your situation, your relationship with the reader and your standards, so the first draft is generic and you spend longer editing it than you would have writing it. A specific request pushes that work onto the model, where it belongs, and the first draft is close enough to send after a light pass.
The 4-part work prompt
Stick to four parts and the prompt almost always lands. Anything longer and you are writing documentation; anything shorter and you are guessing.
- Role — tell the AI who it is. A senior client-success manager writes a reply differently from a blunt project lead. One sentence is enough.
- Task — one clear action. "Rewrite this draft", "Summarize these notes", "Draft a proposal outline". If you list three actions, you get three half-done outputs.
- Context — the facts it needs: who the reader is, what happened, what the outcome should be. Two or three lines beat one line with a made-up scenario.
- Format — how it should deliver: length, structure, medium. "Three short paragraphs", "Six bullet points", "A subject line plus two paragraphs".
Role lines that earn their keep
Three role lines cover most work prompts: "Act as a senior email assistant" for client-facing messages, "Act as a copy editor" for rewriting existing drafts without changing the meaning, and "Act as a meeting facilitator" for agendas and recaps. Match the role to the output you want, not to your own job title — the assistant's tone follows whichever role you name. If the output comes back in the wrong voice, the fastest fix is usually a better role line, not a better task line.
Skip the flattery. "You are a world-class expert" does nothing in a work prompt; a defined role and real context do everything.
Example
Instead of "write me an email", try the four-part version:
"Act as a client-success manager. Write a 3-sentence reply to a client who is unhappy about a two-week delivery delay. Keep it warm, take ownership, do not promise a new date, and offer a status call. Format: three short sentences."
Every part is doing work there. The role sets the tone ceiling, the task limits it to a reply, the context gives the real situation and the boundary ("do not promise a new date") blocks the model from inventing the thing you actually have to decide.
A second mini-example, for a meeting agenda: "Act as a project lead. Draft a 30-minute agenda for a kickoff with a new client, covering roles, scope, timeline and open questions. We have a fixed budget of 40 hours and a deadline in three weeks. Format: five sections with 2-3 bullets each."
A third quick example, for a status report: "Act as a project lead. Condense this status report into five lines: progress, risks, decisions needed, next milestone. Format: five bullets, each under fifteen words." The word limit stops the model from narrating the whole report back to you — which is the most common failure of a summarising prompt.
A copy-paste template
Block out the framework once and reuse it. The structure below is deliberately boring, which is the point.
Role: [who the assistant is] Task: [one action] Context: [reader, situation, outcome, boundaries] Format: [length, structure, medium]
Fill the brackets, paste, and you have a work prompt. Keep this template in a notes app or a snippet launcher so you do not rebuild it from memory when you are tired — which is exactly when prompt quality collapses.
Refine until it's usable
The first output is draft zero, not the answer. Two short follow-ups beat one perfect prompt attempt. If the tone is wrong: "Make it more direct." If it is too long: "Cut it to half, keep the same points." If it missed the context: "Here is the missing fact — rewrite with it." Each follow-up is cheap and the model keeps the thread, so the refinement usually lands in one shot.
The one exception to the whole framework: if the task takes less than two minutes by hand, just do it by hand. Wiring a prompt for a tiny job is how people turn a productivity tool into a procrastination tool.
Honest expectation: AI output is a first draft with perfect grammar. Facts, numbers and commitments are your job to verify before anything goes to a client, team or stranger.
Once the framework clicks, the fastest upgrade is volume — you stop rebuilding prompts from scratch. The The Prompt Engineering Playbook holds 300+ proven work prompts organized by role, so you can adapt instead of architecting every time.
