One sentence, "Avoid candles, mugs and jewellery", did its job. Without it, a category we would have banned turned up in a third of answers. With it, none did.
What mattered more was the replacement. In most cases the model did not swap one cliché for another. It moved to an idea built on something we had said about the person.
What we tested
Advice about gift prompts usually focuses on what to add: "likes skincare, films and travel." Sometimes the faster move is to say what should go.
avoid candles, mugs and jewellery
The model should stop suggesting those things. That part is easy to check. The harder question is what happens next. Does it move toward the person, or just to the next safe default?
We wrote 16 fictional people, each with interests, a habit and something going on in their life. Every brief was run in two versions: once as written, once with a single extra sentence naming three categories to avoid. Nothing else changed.
Recipient: sister, early 30s. Birthday. Budget up to £70. She started ceramics classes this year and watches old films. She cycles to work every day, even in winter. She recently moved into a smaller flat. Avoid candles, mugs and jewellery. Suggest exactly four gift directions...
The categories were chosen per person, and they were things a model would plausibly suggest: socks and whisky for a cyclist brother, gift cards for a teenage nephew, mugs for a neighbour. Each version ran three times in a fresh session, in random order, for 96 answers and 384 gift ideas.
Did the model obey the avoid list?
Yes, every time. Not one of the 48 answers with an avoid line suggested a banned category.
Answers that included a category we wanted to avoid
- No avoid line33%
- With avoid line0%
The more useful number is the first one. Even with a detailed brief and a request to explain each idea, a third of answers reached for a default we did not want.
Those defaults were not spread evenly. They clustered around the thinner briefs: the teenage nephew got gift cards and revision guides, the neighbour got a travel mug in all three runs, the coworker got streaming gift cards, the sister-in-law got a bath set. Where there was less personal detail, the model filled the gap with what everyone buys.
What replaced the banned ideas?
This is where an avoid list is either useful or cosmetic. If "candle, mug, jewellery" just becomes "diffuser, bath set, tote bag", the prompt worked on paper and changed nothing.
That is mostly not what we saw.
Ideas per answer that were a sideways swap into another stock gift
- No avoid line0.31
- With avoid line0.10
- Compact watercolour travel set
- Crime novel or audiobook
Flagged: Luxury bath and body set- Flagged: Spill-proof travel mug
- Low-mess pocket watercolour set
- Short crime story anthology
- Audiobook subscription
- Luxury sofa throw
Two of the four ideas improved. One sideways swap survived: the sofa throw is the bath set's cousin. That pattern held across the study. The avoid line removed the worst defaults, and the gaps usually went to something tied to the person, not every time.
Two other measures moved slightly in the right direction. Ideas tied more clearly to a stated fact (recipient grounding up 0.11 on a 1 to 5 scale, 95% interval 0.03 to 0.20), and generic ideas fell slightly (down 0.09, but the interval runs from -0.20 to +0.02, so we would not lean on that one).
Why a short "no" can be useful
Positive preferences are broad. "Likes films" could justify hundreds of things. A short negative removes a whole branch at once:
- does not want decor
- hates scented products
- already owns too many mugs
- does not wear jewellery
That does not tell the model what to buy. It tells it where not to waste the four slots you asked for.
The danger: cornering the model
More exclusions are not automatically better. If you write "no mugs, candles, food, experiences, clothing, tech, decor, books, gift cards or hobby gear", the model is not discovering your friend. You have boxed it in.
We did not test long lists, so we cannot tell you where the line is. A good exclusion is one of three things:
- a category the person genuinely dislikes;
- something they already have plenty of;
- a risk you want to avoid, like guessing a size or a scent.
What to do with this
Add one sentence with two or three exclusions you are sure about, especially for someone you do not know well. That is where the defaults crept in.
Avoid candles because she dislikes scented things, and mugs because she has too many.
Giving the reason is a follow-up we have not tested yet. The idea is that "dislikes scented things" should also keep the model away from diffusers and perfume, which "no candles" alone would not.
Method
- Tested
- 5 October 2026
- Model
- Gemini 3.5 Flash via the Gemini API
- Settings
- Fresh call per run, no system instruction, no search, no memory, default temperature
- People
- 16 fictional recipients, UK, budgets £25 to £100
- Conditions
- Brief as written vs the same brief plus one "Avoid X, Y and Z." sentence
- Runs
- 3 per person per condition, random order, 96 answers, 384 ideas
- Scoring
- Each answer scored blind to its condition, by an AI rater (Claude) using a fixed rubric written before any output was read
- Analysis
- Differences compared within the same person (average of the three runs), with 95% bootstrap intervals over people
Raw data
Every prompt, every answer and every score is downloadable. Condition names were hidden from the rater and added back afterwards.
Free to reuse under CC BY 4.0: credit Giftin.ai and link to this page.