How to Write Better AI Prompts: Define the Work Before You Tune the Prompt

Define the goal, usable context, constraints, and a done condition to turn a vague AI request into a first draft you can inspect.

If an AI response is vague, polished but wrong for the job, or impossible to review, the missing piece is often not a clever prompt. Define the goal, the material it may use, the constraints, and what makes a first draft usable. Then you can see what to correct instead of prompting at random.

A hand-drawn character turns one vague work note into four cards that define a task. — Turn a vague request into four checkable task fields.

Many people begin with a request like: “Write a professional email.” The model can return a fluent email, but the actual work is still undefined. Who will read it? Which facts are confirmed? What must not be promised? What would let you send the draft to the next reviewer?

My view is that an AI prompt is not a spell. It is a work brief that both you and the model need to understand. Before looking for a reusable template, make the brief inspectable.

A workable request has four fields

Field	Question to answer	What goes wrong when it is missing

Goal	What needs to be produced, and for whom?	The model has to guess whether you need a summary, a draft, or a final version.

Material and context	Which confirmed information may it use?	It may fill gaps with general knowledge, assumptions, or incomplete clues.

Constraints	What format, scope, tone, and decision limits apply?	A plausible response can still overstep your authority.

Done condition	What makes this first draft ready for its next step?	“Make it better” becomes the only feedback you have.

This is a teaching tool, not an official formula from an AI provider. Its value is diagnostic: when the output misses the mark, you can identify which part of the work was underspecified.

Turn a vague request into a brief someone else could follow

Here is a common request:

Write a professional announcement email for our event.

Here is the same job with its working conditions made visible:

Goal: Draft an event announcement for existing subscribers. It should explain
the topic, time, and registration step.

Material and context: Use only the confirmed information below: [paste source].

Constraints: Write in Traditional Chinese and keep it under 250 characters.
Do not add speakers, discounts, capacity, or performance promises. If a fact is
missing, list the gap instead of filling it in.

Done condition: Include a subject line, opening, three confirmed details, and
one call to action. Every factual detail must be traceable to the source.

The point is not that the second version is longer. It assigns the first draft to the AI while keeping facts, promises, and final judgment with the person who owns the work.

The field that matters changes with the job

For writing, pin down the reader, the confirmed facts, and any promise you cannot make. For analysis, pin down the data range, comparison fields, and inferences that are out of scope. For coordination, pin down decisions, actions, owners, and unresolved questions.

That is why I would not start with a universal workplace prompt. Keep the four-field structure if it helps, but rewrite the material, constraints, and done condition for each real task.

Try it on a small task you can judge yourself

Pick one low-risk task for which you know what a useful result should look like. Use public, synthetic, or authorized material. Fill in the card, ask for one draft, and compare it with the done condition.

[Task card]
Goal:
Material and context:
Constraints:
Done condition:

Data or decisions the AI must not receive or make:

If the result is off, do not immediately replace the whole prompt. Check whether you omitted material, a constraint, or a done condition, then change that one field. The reusable outcome is not a lucky paragraph of instructions. It is a clearer way to describe the work.

A clear task card does not grant permission

Defining the task does not make work data safe to share, and it does not make a draft ready to send. For customer data, personal information, contracts, budgets, or high-consequence decisions, the data path, authority, and human review still need their own check.

Where this method goes next

The next exercise applies this brief to a source-backed meeting record: turning existing text into decisions, actions, owners, and unknowns without inventing them. If you already have a complete-looking answer, use a verification record to decide which claims can be used.

Related reading: AI meeting notes, verify AI answers and data and authority checks.