A practical AI workflow for a normal workday

Most AI advice fails on an ordinary afternoon. The demo assumes a clean problem, a patient operator, and an audience that wants to be impressed. A workday is none of those things. You have a review to finish, a note to send, a design that is almost right, and twenty minutes.

The workflow I actually keep is small. It is written up as a free starter kit, and it is still a practice, not a product. Four moves. One task at a time.

The job comes first

Before the prompt, write one sentence: what does done look like, and who has to use the result? If you cannot say that, the model will invent a task and complete it confidently. That is how you get a long answer to a question you did not ask.

Constraints belong in the same breath. Audience, length, what must not be invented, and the format you can edit. A model is better at filling a shape than at guessing which shape you meant.

Four moves

• Frame. One sentence for the job, plus the constraints and the non-goals.
• Draft. Ask for a first pass you expect to mark up, not a final artifact.
• Interrogate. What was assumed? What would a careful reviewer flag? What did this not solve?
• Keep. Save only the sentence, checklist, or prompt you would use again, rewritten in your voice.

The third move is the one people skip. A fluent draft feels finished. Fluency is not review. I ask the model to list assumptions and weak spots, and then I decide which of those are real. The interrogation is still my job.

A prompt shape that survives reuse

I do not collect prompts the way people collect bookmarks. A prompt earns a place when it has survived a second, different task. The shape is boring on purpose:

• Job — one sentence.
• Audience — who acts on this.
• Constraints — length, tone, sources, format.
• Do not — the inventions, claims, and flourishes you refuse.
• Output — something you can edit, not a performance.

If the prompt is more impressive than the edit you made afterward, you do not have a workflow yet.

What stays human

I do not automate a message that carries accountability, a review of work I have not read, or a decision that depends on context the model cannot see. Speed on those tasks is how trust leaks. The workflow is allowed to say no.

What it does well, for me, is the second draft of structure: an outline, a checklist, a comparison of two options, a plain-language summary I then correct. The correction is the work. The draft only buys the blank page.