“Make it better” often arrives when we are tired. After organizing the material, the hardest part remains: explaining what “better” means.
AI returns something complete. We spend longer explaining that it is too formal, insufficiently specific, or simply not what we meant.
Try a small task: a new-game description for a website. The facts and constraints below are invented for practice, not a claimed tool test. Use your usual conversational AI.
Say what the reader needs to do next
“Introduce the game” differs from “help a first-time mobile visitor decide whether to open it.” The second needs a goal, controls, and relevant conditions more than a pile of praise and lore.
| Item | Requirement for this exercise |
|---|---|
| Reader | A first-time mobile visitor |
| Needed information | Goal, touch controls, screen orientation |
| Evidence | The supplied, checked gameplay description |
| Completion | All three facts; no unsupported features; a short length target |
The original Chinese exercise sets 80–120 Chinese characters. Choose a suitable word target for an English task. Length is a local constraint; the essential part is knowing how you will check the result.
Leave unknowns visible
Provide these fictional facts: players move blocks to group matching colors; controls use dragging; portrait orientation is supported; no information is supplied about saving, multiplayer, or level count.
Draft only from this material. First list unresolved questions, then the description. Do not fill missing features with conventions from similar games. Finally connect every feature claim to its supporting source sentence.
“Save anytime,” “challenge friends,” and “hundreds of levels” would each fail this check. Remove unsupported features before polishing the sound.
Requesting evidence is still only a self-check prompt. Compare the explanation with the original material yourself.
Give a useful example and a bad one
Preferences can be easier to communicate through comparison than adjectives.
A useful sample: “Drag the blocks to bring matching colors together. Start in portrait mode when you feel like spending a few minutes on a small puzzle.” It is specific and avoids promises about unknown features.
A bad sample: “The year's unmissable masterpiece: a thousand players on screen and unlimited levels!” Explain why: exaggerated praise hides the controls, and multiplayer and level counts lack support.
Ask AI to explain the difference before drafting. You can discover mismatched meanings of “clear,” “natural,” or “appealing” early. Your own samples also help preserve your expression habits.
Repair one issue per pass
Check facts, then reader understanding, then rhythm. Simultaneously requesting shorter, more detailed, emotional, and professional text can make one improvement undermine another.
Try: “Keep the facts; change only the opening to state the action directly.” Then ask whether the revision introduced a new feature. A clear scope makes the cost of a change visible.
If two rounds miss the same requirement, improve the brief with an explicit constraint. Repeating “still wrong” leaves both sides circling the same uncertainty.
Save the small reusable part
Keep the brief, adopted prompt, and acceptance checks. Do not treat the whole conversation as a template. Next week you will probably need only the reader, evidence, and definition of done.
Use another card for minutes, an email, or comparing proposals. Begin with a small task whose answer you can check, and record where rework occurs.
A lucky guess feels good. Explaining the task well enough to give the next attempt a fair chance is a skill you can keep. Four clear lines today leave tomorrow a little less to untangle.