Prompts that get better name ideas from any LLM

The base prompt

You are helping name [a two-person design studio that mostly does packaging for food brands]. Give me 20 names. Mix four shapes: two real words joined (formhaus), an invented word that sounds like the field (quoin), a word split across the dot so the extension finishes it (kern.ing), and a real trade term on a fitting extension (bleed.agency). Use extensions that fit the business, not just .com. Do not tell me whether any name is available; you can’t know. One line each.

The last sentence matters. It stops the model decorating every name with “(available!)”, which is where the padding usually comes from. The four shapes matter too: without them you get twenty compounds.

Prompt 2: the trade-term mine

List 40 words that only people who work in [packaging design] would use, including tools, materials, measurements, mistakes and slang. Then pick the 10 that would work as a company name and pair each with the most fitting extension.

Trade terms are where the free names are. The model knows the jargon; it just doesn’t reach for it unless asked. This is the same reason trade-term compounds are still findable in .com when plain word pairs aren’t.

Prompt 3: the anti-brief

Here is the brief: [one sentence]. First list 10 names that would be wrong for this business and say why in five words each. Then give me 15 names that avoid every one of those problems.

Asking for the wrong names first makes the right ones sharper. It also surfaces what your brief is missing: if the model can’t say why a name is wrong, the sentence wasn’t specific enough.

Prompt 4: the narrowing round

These names came back taken: [paste list]. Notice which shapes keep failing. Stop writing those shapes. Give me 15 more that use a different shape or a different extension.

This is the loop. The model can’t check, but it can learn from what you tell it came back taken — if you tell it. A round that isn’t fed its own failures just repeats them.

Prompt 5: the say-it test

For each of these 10 names, write how a stranger would say it aloud after reading it once, and flag any that could be misread, misspelled or mistaken for another word.

Run this before you check availability. It’s cheaper to drop a name here than after you’ve bought it.

What none of these prompts do

They don’t check the registry. Every list you get is a list of ideas, some of which are taken, some of which are premium, and some of which are free. A model cannot tell you which is which, and a better prompt doesn’t change that — it only stops the model pretending otherwise.

Check them at the registry, or use a tool that does, before you get attached. That is the step this page can’t give you.

Plainly

Brief the model in one sentence, ask for specific name shapes and fitting extensions, and tell it not to claim availability. Then check every name at the registry, because the model can’t.

What Webault does about this

Webault runs this loop for you: one sentence in, four name shapes out, every name checked at the registry, taken shapes fed back into the next round.

Questions

Related reading

• Why AI makes up domain availability — why a model's availability claim is predicted text, not a lookup.
• How to write a one-sentence naming brief — what belongs in one sentence, with five keyword lists rewritten.
• How to find an available .com in 2026 — the shapes still free in .com, and how to check them properly.