Turn a Weak AI Answer Into a Useful One

The first answer is often a sketch. It may be too long, too bland, or built on an assumption you never approved. That is normal. People who use AI well do not write a perfect opening prompt every time; they notice what is missing and steer the next turn.

This lesson gives you a small repair kit. You will practice asking for a tighter answer, a different perspective, and a visible list of assumptions. The aim is not endless chatting. Stop when the draft meets the job—or when you learn that the tool lacks the facts to finish it.

Weak follow-up

“Try again.”

The tool has to guess what went wrong.

Useful follow-up

“Remove the parking claim. Cut to 120 words. Keep the date blank.”

One precise correction makes the draft reviewable.

Edit the defect you can name; do not wait for a perfect first answer.

Name the defect, not the mood

“Try again” usually invites another guess. Say what failed. Was the answer too generic for your audience? Did it omit a key constraint? Did it invent a policy? Did it bury the decision? A single specific correction can save several vague follow-ups.

For the workshop email, imagine the assistant wrote 300 words and promised free parking even though parking has not been confirmed. A useful response is: “Cut this to 120 words. Remove the parking claim. Keep the hands-on benefit in the first two sentences. Mark the registration link as undecided.” That is editing, not prompt wizardry.

Try this

Copy a repair prompt

Use it after the first draft of the fictional workshop invitation.

Review your draft as an editor. Identify (1) any detail I did not provide, (2) any sentence a busy customer can skip, and (3) any question the invitation still leaves unanswered. Then revise it to no more than 120 words. Keep undecided details in square brackets. Show a three-line change note after the revision.

Why it works: A useful follow-up asks the model to inspect its own draft, but you must still check the result yourself.

Ask for a second angle when the choice is real

Sometimes the draft is fine, but the decision is not obvious. Ask for two plausible approaches and what each gives up. For instance: “Compare an invitation that emphasizes hands-on practice with one that emphasizes executive takeaways. What audience is each likely to suit?” The assistant can expose trade-offs; it cannot tell you which audience your actual customers are without evidence.

You can also ask it to argue against its own recommendation: “What would make this plan fail for a participant who is new to AI?” The answer might surface missing instructions, an inaccessible activity, or too little time for questions. Treat these as hypotheses to check, not findings from real participants.

Stop at a reviewable draft

Do not keep polishing until the text merely sounds impressive. Decide what “done” means: correct known facts, appropriate length, clear next action, and no unsupported promise. If a missing answer depends on a real person or policy, put it on the decision list. An assistant cannot settle an undecided price by writing confidently.

Save one follow-up that worked. Over time you will build a small library of repair moves: shorten, compare, flag assumptions, change audience, show sources, or turn a paragraph into a table. That is a more valuable skill than memorizing a single giant prompt.

Practice assignment

The three-turn edit

Start with an imperfect AI draft of an approved or fictional work item.

1Identify one factual problem, one audience problem, and one format problem.
2Give a precise correction for two of them in the same conversation.
3Ask for a change note, then check the revised answer yourself.
4Stop when the result is reviewable; list any unresolved fact separately.

What to save

A first draft, your follow-up, a revised draft, and a short “still needs a person” list.

You are ready when

The revision fixes the named problems without adding a new unsupported promise.