How to write better AI prompts for work

Choose one task with a checkable result

Begin with an email outline, a summary of notes you supply, or a comparison of options you already understand. A broad request such as “improve my business” leaves the model to guess the problem, audience, and evidence.

Choose work where you can explain what a good result looks like. For a meeting summary, that might mean every action has an owner explicitly named in the notes. For an email, it might mean a single question and a clear next step.

Use four fields instead of a magic phrase

The following template is a TechlyUp practice method. It is a convenient way to make your request explicit, not a guarantee that a model will follow every instruction.

1. Task: say exactly what you want produced, using one main action.
2. Context: supply the intended reader, relevant background and source material. Use only information you are permitted to share.
3. Format: specify an email, table, checklist, or another output you can review.
4. Constraints: state word limits, facts that must remain unchanged, and how to handle missing information.

TASK: Draft a project status update.
CONTEXT: Synthetic practice notes: API review finished Monday. Load testing has not started. No release date is agreed. Reader: internal project lead.
FORMAT: Three bullets: completed, next, decision needed.
CONSTRAINTS: Use only these notes. Do not invent owners or dates. State “not agreed” for the release date.

Compare the draft with the input

The expected result should distinguish a finished review from an unstarted test and should not announce a release date. If the draft says “on track for Friday,” reject that unsupported addition rather than merely polishing its wording.

Change one instruction at a time when a result misses the target. Save the prompt together with the checks that made the output acceptable. This gives colleagues something more useful than an isolated screenshot.

What this method cannot guarantee

A detailed prompt cannot make unknown facts true or guarantee confidentiality. Tool settings, organization policy and model behavior still matter. For information outside your supplied source, verify against an appropriate primary source.