I sold a roll of wallpaper as a pet bed, so I built a validator that can veto my own product names
A dropshipping supplier's product title looks like this:
Pet Dog Cat Bed Warm Soft Plush Comfortable Sleeping Mat Puppy
Kennel Washable Nest Cushion Pad Four Seasons Universal 2024 NewThat is not a name. It's every search term the seller could think of, stapled together and run through a translator. You cannot put it on a storefront.
The obvious fix is to have a model rewrite it into something readable. I did that. And a while later my store was selling a roll of wallpaper, listed as a pet bed.
How that happens
Nobody hallucinated a lie on purpose. The supplier title mentioned a mat, a pattern and a room; the rewrite reached for the most product-shaped noun in the pile and picked wrong. Once it's written down as a clean, confident, four-word product name, nothing about it looks generated. It looks like a decision someone made.
That's the real hazard. A messy title is obviously messy. A tidy name is trusted, and if it's wrong, the trust is what does the damage — the customer doesn't find out until a roll of wallpaper arrives.
The rule
So the display name is not allowed to be creative. It's allowed to be an edit.
Every word in the display name must already appear in the supplier's real title. Reorder, trim, re-case. Never add.
That's it. It rules out the entire class of error above, because a word that isn't in the source can't appear in the output, so there's nothing for a wrong guess to be made of.
Making it a gate, not a guideline
A rule that lives in a prompt is a suggestion. This one runs as code, after generation, and any name that fails is thrown away and replaced by the mechanical cleanup:
def honest_name(candidate: str, source: str) -> bool:
"""True iff every word of candidate already appears in source."""
vocab = _honest_vocab(source)
for chunk in (candidate or "").split():
for w in chunk.split("-"):
key = _honest_key(w)
if key and key not in vocab and key not in GLUE_WORDS:
return False
return TrueTwo details did most of the work in practice:
Normalisation. Comparing raw words rejects far too much. Women's should satisfy female; Stones should satisfy stone. So each word is folded first — lowercased, stripped of punctuation and possessives, simple plurals removed, and a tiny equivalence table maps female/woman to women and male/man to men. Without that fold the validator rejects good names and you learn to ignore it, which is worse than not having it.
Glue words. and, with, for, the, of are exempt. They carry no product claim, and forbidding them produces names that read like a telegram.
What it changed
I ran all 120 display names in the catalogue through it. The validator is the reason I can state plainly, on a storefront that rebuilds itself every three days with no human in the loop, that no product name says anything the supplier didn't. Not because I check them — because a name that says something new cannot survive the build.
There are eight exceptions in the current catalogue, and I'd rather write them down than let you discover them: eight items got one noun added that I confirmed from the supplier's own product photo. Each one was a judgment call I made by looking at the picture, and they're the only names in the shop that a machine didn't have to approve.
The part I'd tell anyone building the same thing
The temptation with AI in a pipeline is to make the prompt better. Better instructions, more examples, sterner warnings. It genuinely helps, and it is not a guarantee, because the failure you care about is the confident one — and a confident wrong answer is exactly what a well-written prompt produces more of.
The thing that actually holds is a check that runs afterwards and doesn't care how good the output sounded. Give it a rule narrow enough to be mechanical — "no new words" is mechanical; "be accurate" is not — and let it throw work away. Mine throws away perfectly nice product names on a regular basis. That's the feature.