namekit
Name a project against a convention rather than a blank page, and hand back only names that are actually free. Two explicit modes:
- . Turn a project description into a ranked shortlist of names built to the user's naming convention, then probe the two or three the user picks against the domain, npm, and GitHub namespaces and crown one.
- . Take names the user already has and report whether each one is free. No generation.
The reason this is a skill and not a one-line ask: left alone, a model returns ten spellings of one idea, invents a convention nobody uses, and never checks whether any candidate is taken. namekit fixes all three, in that order.
When this fires
- generate. "Name this project", "what should I call this", "give me name ideas", "come up with a name for X", "I need a name for my new SaaS".
- check. "Is taken", "check if this name is available", "did anyone grab this npm package".
An ask that supplies a description gets
. An ask that supplies names gets
. When the user supplies both, run
on their names first, then offer
, because a name they already like outranks anything you invent.
The probes (both modes)
Three namespaces, three commands. Each one is a public HTTP contract, so it is safe to pin.
sh
# domain registration: 404 = not registered, 200 = registered
curl -s -o /dev/null -w '%{http_code}' -L --max-time 10 https://rdap.org/domain/<name>.com
# npm package: 404 = free
curl -s -o /dev/null -w '%{http_code}' --max-time 10 https://registry.npmjs.org/<name>
# GitHub org or user handle: exit non-zero / 404 = free
gh api /users/<name>
Check
plus whichever second TLD the project's surface implies (
for a developer tool,
for a model product, the country TLD for a local audience). Skip a namespace the project will never occupy: a CLI-only tool does not need an npm name, and an internal tool needs no domain.
Probe two or three names at a time. RDAP answers a small batch and returns 429 to a burst, and a 429 is not a verdict. A bulk sweep of dozens of names throws away its own domain results.
Report each result as that probe's own claim. A 404 from RDAP means the domain is not registered. It does not mean the domain is for sale, is affordable, or is free of a trademark. Say "not registered", and let the user find out the price.
Degrade rather than block. With no shell, no network, or no
, say which probes were skipped and hand back the ranked shortlist on the four rubric criteria alone. A shortlist with a stated gap beats no shortlist.
Mode:
1. Restate, then interview
Reflect the project description back in one sentence before asking anything. A misread costs one correcting line here and a whole shortlist later.
Then interview. Six items make the pool; ask only the ones the description leaves open, cap the round at five questions, and ask them in a single batch:
- What it does. The domain and the outcome.
- Who it is for, and what language they read. A Bangla-reading audience opens roots an English-only audience closes.
- Tone. Plain, technical, or playful.
- Words to include or avoid. A founder usually has one of each.
- Where the name gets typed. Domain, package name, CLI command, org handle. This sets the length tolerance and decides which probes run.
- How permanent it is. A throwaway internal tool and a product you will defend for a decade deserve different effort.
Done when every unanswered pool item has an answer or an explicit "does not matter", and the restatement stands uncorrected.
2. Resolve the convention
Work down this ladder and stop at the first rung that answers:
- Stated. The user names the convention outright ("suffix everything with ").
- Derived from examples. The user cites their own names ("like , , "), and the shared tail is the convention.
- Read from their repos, on request. Only when the user points at them:
gh repo list <owner> --limit 100 --json name
- Ask. Offer suffix, prefix, portmanteau, or none.
State the resolved convention in the output, so a wrong read costs one correcting line rather than a discarded shortlist.
Done when the convention is written down as an affix plus its position, or the user has chosen to work without one.
3. Mine roots and build candidates
Draw roots from four separate sources, so the set is genuinely plural instead of one idea in ten hats:
- The domain noun. What the thing is about: , , .
- The outcome verb. What it does for someone: , , .
- The user or their material. Who holds it: , , .
- The metaphor, or a non-English root when the interview named a non-English audience.
A non-English root passes one gate: a single Latin transliteration dominates in common use. When a root has two spellings people genuinely both write, drop it rather than scoring it down, because a name the audience types three ways fails on the one job a name has.
Join each root to the affix with the
seam rule, score it against the
rubric, then cut. Generate
25 or more raw candidates, and rank the best
12 into the shortlist. The width is deliberate: the user probes two or three names per batch, so the list has to feed several batches before it needs regenerating.
Done when the ranked 12 draw roots from at least three of the four sources, and each carries a one-line rationale plus its rubric verdict.
4. Show the shortlist, then probe what the user picks
Print the ranked 12 before you probe anything. The names come first, in a table of name, root and source, seam case, and rubric verdict, with no availability column yet. The user reads the whole list, reacts to the ideas, and keeps the creative half of the job in front of them.
Then ask the user to pick
two or three names to probe. Use
when it is available, and a plain one-line ask when it is not.
Probe only the names the user picks. An unpicked name stays unprobed, however good its rubric score.
Run
the probes over the picks.
Any hit removes that name. Availability is a verdict here, not a score, because a name the user cannot have is not a candidate.
One exception: a namespace the user already owns passes, and reads yours. Resolve their owner from the git remote when a repo exists (
gh repo view --json owner
), from the repos that sourced the convention otherwise, and by asking once when neither answers. Without this exception a house convention rejects its own portfolio, which is how
fails a
filter.
When a batch returns nothing free, report each pick with the probe that killed it, reprint the candidates still unprobed, and ask for the next batch of two or three. Repeat the loop until one name is free or the user stops it. Never auto-select the next batch, because the pick is the user's.
Regenerate only when the list runs out. Once every one of the 12 is probed and taken, run
Mine roots and build candidates again with the taken roots excluded, and show the new list the same way.
Two generation passes is the cap.
Done when at least one probed name is free, or the user stops the loop, and every probed name carries a result for every probe that ran.
5. Crown, then search once
Crown the free name that ranks highest on the rubric. Then run one web search, on that name alone, for the existing product or live trademark the registries miss. Fetch the top hit when the search result is ambiguous about what the thing is.
A hit re-crowns the next free name and states the conflict in a line. When the batch left no runner-up, say so and send the user back to the pick loop. This is the only search in the run: the registries are exact-match and cheap, the web is fuzzy and expensive, so it earns one call at the point the answer changes a decision.
Done when the crowned name has been searched and either survives or has been replaced.
6. Hand off
Write every hand-off in this skill in the procedural register: one instruction per sentence, active voice, present tense, no metaphor.
Print the probed names as a ranked table before the three beats:
| Name | Root (source) | Seam | Rubric | | npm | GitHub |
|---|
| grow (outcome verb) | hard join | 4/4 | free | free | free |
What changed. Report the resolved convention, the crowned name and the one-sentence reason, how many batches the user probed, and whether a second generation pass ran. Name each probed name the registries took, with the probe that killed it. List the candidates that stayed unprobed, so the user can come back to them.
Where it landed. namekit writes no file by default. Write
docs/names/names-<slug>-YYYY-MM-DD.md
only when the user asks, where
is the project's short kebab-case name and the date is the file's creation date. Keep that date stable on a later edit. Follow the host repository's own artifact convention when it documents one. With no filesystem, print the document as a codeblock and give the filename.
Next. Crown one move. The name is chosen and nothing owns it yet, so the move is to claim the namespaces the project actually needs, in the order the interview's typed-surface answer implies. Name
repokit to set a new repo's About panel and topics when it is installed, otherwise
. When the name came out of an idea session, name
ideakit as the runner-up so the decision lands back in that idea's log. namekit never registers a domain, publishes a package, or creates an org.
The seam rule
The join is where bad names come from, and three cases cover every convention, because the rule is about the letters meeting at the seam rather than about any one affix.
- Hard join. The root's tail and the affix's head do not collide, so concatenate. + → . Prefix form: + → .
- Elision. They overlap, so drop the duplicate. + → , not . Prefix form: + → .
- Reject. The seam makes a vowel pileup or a syllable nobody says aloud, and no elision saves it. + elides to , which reads as "ideal-oy". Drop the root and pick another.
The rubric
Four criteria, scored pass, weak, or fail. Rank by the count of passes, and break a tie on semantic fit.
- Semantic fit. The root says what the thing does.
- Sound. Three to four syllables, and it reads correctly on first sight.
- Seam quality. A clean hard join or a clean elision, per the seam rule.
- Spell-on-hearing. Someone who hears the name spells it one way.
Availability is deliberately absent. It is a verdict in
Show the shortlist, then probe what the user picks, not a fifth criterion, because the rubric ranks names the user has not probed yet.
Mode:
Run
the probes over the names the user supplies. Generate nothing, rank nothing, and filter nothing, because there is no shortlist to cut down to.
Apply the same owner exception: a namespace resolving to the user's own owner reads yours, not taken. That is the common case here, since people check names they already half-own.
Hand off
What changed. Report one line per name with its result in each namespace, and mark the namespaces the user already owns. Name the probes that were skipped.
Where it landed. Nothing was written and nothing was registered. Say so plainly.
Next. Crown one move from the results. When a name is free, the move is to claim it, so name
repokit for a new repo's metadata when it is installed, otherwise
. When every name is taken, the move is
, which mines new roots around the same idea.
Notes
- namekit names the thing and stops. No logo, no tagline, no positioning, no brand identity, and no renaming of an existing codebase's identifiers.
- The naming convention belongs to . namekit works to whatever convention the user resolves and never teaches that one, so the rule stays in one place.
- No config file and no environment variable. The convention lives in the prompt or in names the user already has. State that is not visible in the conversation is state that goes stale.
- Social handles stay out. Headless checks against social platforms are rate-limited and return false negatives, and a wrong "taken" is worse than no answer.
- Prefer a probe's status code over scraping a page, because the status code is a stable contract and the page is not.