product-naming

Find a name that people remember, that feels like the brand, and that the user can actually own: the domain, the trademark, the handles and the search results.

Names are cheap to generate and expensive to own. In the run this skill was built from, about 870 candidates were tested. Nearly every single word, obvious compound and "hey/meet + name" was taken as a .com. Of the few that were free, many then died on a language or search filter. So generate wide, let the scripts do the cheap checks in bulk, and spend judgment only on the survivors.

Talk to the user and write the report in the user's language.

What a good name is here

The name does not need to say what the product does. Descriptive names (PostScheduler, Agência do Bairro) are almost never free. They make weak trademarks: offices refuse or narrow them. They drown in generic search results, and the company outgrows them when the product changes. Prefer brandable names, unless the user asks for a descriptive one:

• arbitrary words: Apple, Slack, Notion
• evocative words: Nike, Patagonia
• invented words: Kodak, Spotify
• unexpected combinations: Mailchimp, Dropbox

What the name needs instead:

• Fit (aderência). It feels like the brand. Its sound, rhythm and associations match the personality and the promise: calm or energetic, premium or playful, expert or friendly. The audience would say it out loud without embarrassment. Fit comes from the feeling the name gives, not from a description of features.
• Stickiness. Heard once, it is remembered the next day and typed correctly.
• Marketing potential. You can build a story, a visual identity, a tone of voice and campaigns around it. It can become a verb or a name for the community. It does not box the company into one feature or one market.

references/criteria.md turns these qualities into filters and a scorecard.

Working folder

Keep the run in naming/<slug>/ in the current directory. If that directory is a code repo the user would not want touched, use a temp directory instead.

naming/<slug>/
├── brief.md               # answers from step 2
├── candidates/NN-<strategy>.txt  # one name per line
├── results/NN-<strategy>.tsv     # check_domains.sh TSV output only
├── shortlist.md           # survivors with scores and notes
├── vetting/               # step 7 output: domains.txt, inspect.txt, vet.txt
└── report.md              # the final ranking

The files make the run resumable, keep the same name from being checked twice, and let you measure the yield of each strategy. Keep results/ for the plain TSV output, because the dedupe command in references/generation.md reads it. Anything else, such as --table output, goes in vetting/.

Workflow

1. Understand the product

Read what exists: the landing page, README, pitch, docs. Find out what the product does, for whom, how it charges, and what it promises.

Do not infer markets or languages from details like the payment method, the currency or the language of the copy. The run behind this skill assumed the product was Brazil-only because it had Pix and prices in reais. It was international, so a whole round of Portuguese names was wasted. Ask instead.

2. Brief: ask before generating

Ask everything the material does not answer in one message, with a suggested default for each question:

1. Markets and languages. Where does it sell now, and where in two or three years? Every one of these languages, current or future, becomes a meaning and sound filter: the name will travel before the product does.
2. Domain. Which extension is a must: .com, .com.br, .ai, .app? Is a modified .com acceptable (get-, try-, -hq)? Is there a budget to buy a taken domain? Aftermarket names cost from a few hundred dollars to tens of thousands.
3. Personality. Three adjectives, and the feeling the name should leave. Brands they admire, from any category.
4. Freedom. Brandable (the default), or should the name hint at the category? Are there languages whose words must be avoided, such as no Portuguese words for a global product?
5. Constraints. Maximum length, sounds to avoid, names they already like or rejected (and why), and competitors to stay away from.
6. Handles that matter. Instagram and TikTok for consumer and social products; GitHub and npm for developer tools.

Save the answers to brief.md. If the user says "just go", use the defaults and state them in one line.

3. Generate in batches by strategy

references/generation.md has the strategies, how to work each one, and the yields measured in the reference run. Read it before the first batch.

• Batch size. Write 40–80 names per batch, one strategy per batch, to candidates/NN-<strategy>.txt. Use one name per line, ASCII lowercase. Put the intended spelling in a comment if it differs: postencia # Postência.
• Start where the yield is. Unusual combinations of two known words left about 40% of .com domains free; most other strategies left almost none. If a non-.com extension is acceptable, invented and arbitrary words come back into play.
• Check each batch before generating the next, and record its yield. When a strategy drops below about 5% free, switch to another. Switching groups was the most useful move in the reference run.
• Mine around winners. When a name looks promising, try the same pattern with other words.
• Drop the obvious failures as you write. Cut names that are unpronounceable, that are a word in a forbidden language, or that sound like a competitor. Don't deliberate over names before knowing they are free: the domain check costs seconds, judgment costs much more.

After the first one or two batches, show the user a sample of 8–12 free names across strategies and ask which directions feel right. It is the cheapest way to calibrate taste before generating hundreds more. Skip this only if they asked for an end-to-end run.

Stop generating when about 15–30 free names have passed the filters in step 5. A typical run tests 300–900 candidates.

4. Check domains in bulk

bash <skill-dir>/scripts/check_domains.sh -t com < candidates/03-combos.txt > results/03-combos.tsv
bash <skill-dir>/scripts/check_domains.sh -t com,com.br,ai --table plazabuzz kioskbuzz

• Output. One line per domain: domain<TAB>free|taken|unknown|invalid<TAB>how. With --table, one row per name. -v adds the get-, try-, use-, -hq and -app variants.
• Unknown is never free. It means a timeout, a rate limit or a failed self-test. Run those names again later.
• Free means not registered. Premium or reserved names also look free; the registrar shows the real price.
• Self-test. Before each extension, the script checks a domain that is surely registered and a random one. If either answer is wrong, that extension's results come out unknown. In the reference run, a hand-written .io check reported registered domains as free. In testing this script, a wrong RDAP server called google.co free.
• Speed. The script asks the registries directly and throttles itself for each one. Rough speeds: .com 150 names in about 12 s, .com.br about 4 per second, .app/.dev about 1 per second, whois-only extensions about 1 per second.
• Order. Check the must-have extension on every candidate first, and the other extensions only on the survivors.

5. Filter the survivors

Run every free name through the filters in references/criteria.md:

• a bad meaning in any market language
• autocorrect pulling it toward another word
• a word from a forbidden language
• the radio test (people who hear it can't spell it)
• an expression too common to stand out
• sounding like a competitor

Each of these eliminated at least one real name in the reference run. Two of them are cheap to run on the whole list, so do them here rather than leaving them to vetting:

• Autocorrect: bash <skill-dir>/scripts/vet_name.sh --quick -l <locales> <names> gives one line per name and market.
• Competitors by root: take the words that repeat across the survivors (kite, plaza, chirp) and run one web search per root with the category: "<root>" social media app, "<root>" AI marketing. In the end-to-end test of this skill, the finalist Chirpbell died at vetting because Chirpy, an AI social media tool for small businesses, already existed. A search per root would have dropped every chirp- name before scoring.

6. Score and shortlist

Give the 8–12 most promising survivors the full scorecard from references/criteria.md, with one line of justification per criterion. The rest get a single line saying why they rank lower. Write it all to shortlist.md, and take the best 5–8 to vetting.

7. Vet the finalists

For each finalist, saving the outputs in vetting/:

1. Every relevant extension: check_domains.sh --table -t com,com.br,ai,io,app <names>. Add -v only for names whose plain .com is taken: it checks six times as many domains, and .app runs at one per second.
2. Who holds a taken .com: inspect_domain.sh name.com. It shows the registration date, registrar, nameservers and where the site lands, then gives a verdict: FOR SALE (and on which marketplace), PARKED, NO RESPONSE, REDIRECTS, ACTIVE SITE or UNCLEAR. An active site in the same category is a red flag. PostKite.com, for example, was registered in 2025 for an AI tool to grow on Twitter: a direct competitor for a social media product.
3. Search, trademarks and handles: vet_name.sh -l pt-BR,en-US <names>, with one locale per current market plus one for each future region. It reports:
   • autocomplete that steers away from the name (for example "postencia" → "potência"); references/criteria.md says how much each case weighs
   • similar USPTO marks, live or dead, with their classes
   • INPI counts for exact and radical searches
   • GitHub and npm availability
   • links for everything that must be checked by eye
4. Search conflicts. Search the name in quotes for each market. Look for a company in the same category, autocorrect to another word, or generic results that would outrank the brand. Agents' web search tools often run from the US: good for spotting conflicts, not for judging competition in another country. If you have a browser tool, open the Google links from vet_name.sh, which set the country.
5. Trademark pre-screen and handles. references/clearance.md explains which classes matter, how to read the hits, when a hit is a blocker, and how to check Instagram, TikTok and other handles.

If something the user said matters can't be checked, say so up front instead of burying it. This happens, for example, with Instagram and TikTok when no browser tool is available. The report then opens with a "Pending checks" line, and the ranking is provisional until those checks are done.

references/clearance.md also covers search volume and the SEO logic for brandable names.

8. Report

Write report.md and show it in the conversation. If you can't write files (some subagent setups block it), put the full report in your reply instead.

# Names for <product>
 
**Brief:** markets · required domain · personality · style
**Search:** N candidates in K strategies → F names with .<tld> free (+G only as get-/try-) → S passed the filters → V vetted
**Pending checks:** (only if any) e.g. Instagram/TikTok handles not verified; ranking provisional until then
 
## Ranking
 
### 1. Name (how to say it)
- **Domains:** name.com free · name.com.br free · name.ai taken (for sale on Atom) · getname.com free
- **Why it fits:** the association or story, and how it connects to the brand's promise
- **Marketing potential:** memorability, visual and tone ideas, verb or community use, search (distinct, no autocorrect)
- **Risks:** trademark hits in the relevant classes, similar names, pronunciation in market X, handles taken
- **Score:** fit 5 · memory 4 · sound 5 · distinct 4 · story 4 · range 5 · trademark 4 · availability 5 = 36/40
- **Next steps:** what to register or buy today, which trademark classes to file in, which handles to claim
 
### 2. ...
 
## Also considered
| Name | Why it was dropped |
|---|---|
 
## Before you register
- Availability changes fast: run check_domains.sh again right before buying.
- This is a trademark pre-screen, not clearance. For the final pick, have a trademark attorney or agent search the target classes before investing in branding.
- Register the main domain, the defensive ones (.com.br, an obvious typo) and the handles on the same day.

Rank by the scorecard, but explain the trade-offs. "Best name, but the .com costs US$ 4,000" and "free everywhere, but less memorable" are different recommendations, and the user should see both.

If the user already has a name

Go straight to step 7 for that name and its close variations. If it fails, explain why and offer to search for alternatives in the same spirit, starting the brief from what they liked about it.

Pitfalls

• Untested checks. Never trust an availability check you haven't seen fail correctly. The scripts test themselves. If you write an ad-hoc check, try it on a domain you know is registered and on a random string first.
• Raw whois output. Don't parse it by eye. Some registries print "Domain Name: x" even for names that don't exist (.so), and some never answer (.es). check_domains.sh handles both.
• zsh loops. zsh does not split an unquoted $list into words, so for n in $names runs once with every name. Feed names from a file or stdin, or run loops with bash.
• Availability changes. Re-check before registering. Never register, buy or contact an owner on the user's behalf without asking.
• Legal advice. A pre-screen finds obvious conflicts; it does not clear a mark. Say so in the report.