A better brief gets you past mugs and candles.
You describe a real person.
The AI gives you a candle, mug, gift set, Bluetooth speaker and something with their name printed on it.
Nothing is completely absurd.
Nothing feels like them.
The fastest fix is not “give me 20 more.”
First identify the failure.
Failure 1: the ideas could fit almost anybody
Remove any idea that would work just as well for a random person with the same age or gender. For every remaining idea, name the exact detail from my description that caused you to suggest it.
If you cannot provide a specific clue yourself, add one before requesting another list.
Failure 2: one hobby becomes the whole person
You mention running once. Now you have six running accessories.
Keep running as one possible direction, but no more than one idea can come from that hobby. Use the rest of the person too: routine, relationship, current life, comfort, experiences and things they already use.
The goal is not to ban the hobby. It is to stop it swallowing the person.
Failure 3: the model invents traits
You said:
coworker who likes music
The answer starts calling them nostalgic, creative and introverted.
Plausible is not the same as provided.
Use only the facts I gave you. Do not infer personality, values, taste, living situation or needs. Mark assumptions I still need to verify. If the information is too thin, ask me up to three questions first.
A good response is allowed to need more information.
Failure 4: it treats a hard constraint like a suggestion
Budget
Treat $50 as a hard ceiling. If a strong direction normally costs more, change the category instead of recommending a poor substitute.
Deadline
The birthday is tomorrow. Remove anything that depends on normal shipping. Use only options that can work digitally, locally, through a booking, or as a concrete plan I can give now and do later.
A real constraint should change the recommendation space.
Failure 5: it sounds too confident about shopping facts
Watch for:
- exact prices without a current source
- delivery promises
- suspiciously specific product names
- “available near you” without live search
- stock claims that were never checked
Separate gift directions from live shopping facts. If you cannot verify current price, stock or delivery, tell me what I should search for instead of inventing a listing.
Then check the retailer yourself.
Failure 6: every refresh is the same list wearing different adjectives
Give me six directions that are structurally different. Do not repeat a category from the first answer unless the person details strongly justify it.
Or:
Give me three safe directions and three less obvious directions. Every one still needs a concrete reason from my description.
“Personalized” does not mean “print their name on it”
An engraved object can be a good gift.
The engraving is not the reason it fits.
Do not use names, initials, engraving or photo printing as the main reason an idea is personal. The fit has to come from the person, relationship, routine or current situation.
The best diagnostic question
If you are not sure what failed:
Looking at the person description I gave you, what are the three missing details that would most change your recommendations?
That turns the model into an interviewer instead of a slot machine.
Do not over-correct
A candle is not bad because candles are common.
If your sister uses one every night, has a favourite scent family and specifically asked for one, it may be the most grounded idea in the list.
The test is not novelty.
It is:
Why this, for this person, now?
If the missing clue existed months ago
Then the prompt may not be the real problem.
If you knew it in March and forgot it by October, use a system that keeps the clue.
That is where the Giftin app becomes relevant.

