Idea Name

/idea-name — naming that survives contact with the real world

House taste: crisp/modern over folksy — short evocative real words, no latin-startup-slop (Certus, Aurum), no "-ly/-ify" suffixes. If the user already has a family of named things, list them first and match their register.

The gate order (v2 — this is the whole point)

1. The partner test (gate #1, non-negotiable). Would the founder happily say this name out loud to their partner, or on a podcast, with zero explanation? Familiar spoken-English words only — words a normal human already says out loud. "Bakery" beats "bakehouse"; archaic/regional variants of common words fail here even when they're real words. A name the founder winces at gets under-said: hedged, explained, apologized for. The founder's wince outranks every other check.
2. Collision scan (gate #2). Same/adjacent-space players kill; see kill rules.
3. Domain availability (tiebreaker ONLY — never a generator). Exact-match .com hasn't mattered for dev tools in a decade (linear.app, notion.so, discord.gg). Availability breaks ties between names you already love; it never nominates candidates. Optimizing for unclaimed words silently optimizes for unfamiliar words — the exact failure mode this skill exists to prevent.

Step 1 — Extract the metaphor space

Write down: what the thing does (one line), the emotional register (trust? speed? patience-then-payoff?), and 3–5 metaphor families. The family gives each name a story, and the story is what makes a name feel inevitable later (badge language, design motifs, vocabulary all fall out of it). Generate per family.

Step 2 — Generate wide, expect a bloodbath

15–30 candidates, all gate-#1-compliant. Two structural truths (2026):

• Every crisp English noun is camped in dev/AI tooling — and since the agent gold rush, every familiar word or phrase adjacent to automation, agents, night-work, or working-unattended is camped too, often by multiple funded startups within months of each other. (A 2026-07 rename round for an agent-orchestration tool killed 10 of 12 familiar-word candidates to same-space agent startups.)
• Survivable names hide in territories another industry's software absorbs. Baking words stay free for dev tools because bakery-industry software soaks up all commercial use of them. Look for metaphor families where an unrelated vertical owns the search results — that's cover, not competition.

The compound-pair technique (uniqueness from the PAIR when single words are camped):

1. Common word + qualifier: "Agent Lockbox", "Proof AI" — formal/domain name is the unique pair; the spoken name stays short.
2. Noun+noun compound: Lockbox, Trailhead — one word, two common parts.
3. Coined-by-analogy: "Toolworthy" (like roadworthy) — unique by construction, instantly parseable.
4. Phrase-name: "Proving Ground", "Test Kitchen" — check dev collisions.

Take the founder's throwaway examples seriously — a name tossed off as an example has beaten two full rounds of generated candidates before.

Step 3 — Vet BEFORE falling in love

For each shortlisted candidate, in parallel:

• WebSearch "<Name>" developer tool OR AI startup — the semantic-neighbor check. This is the check that kills most candidates; run it before admiring anything.
• Domains via RDAP (more reliable than whois, free, no key). -sL is mandatory — rdap.org 302-redirects to the registry:

for d in name.com agentname.com name.dev name.sh name.tools; do
  code=$(curl -sL -o /dev/null -w "%{http_code}" "https://rdap.org/domain/$d")
  case $code in 404) echo "$d: AVAILABLE";; 200) echo "$d: taken";; *) echo "$d: unknown ($code)";; esac
done

• npm + GitHub by HTTP status (404 = free):

curl -s -o /dev/null -w "npm: %{http_code}\n" https://registry.npmjs.org/<name>
curl -s -o /dev/null -w "gh: %{http_code}\n" https://github.com/<name>

• Recommendation-grade confirmation: before any domain is recommended or bought, confirm with a raw whois read matching the registry's exact negative phrase (No match for on Verisign .com/.net) — registrar boilerplate false-positives on loose keyword matching (a 2026-07 lesson: two .com domains reported "available" by a registrar search were both taken). If a taken npm/GitHub name matters, inspect the holder (empty placeholder from years ago vs active project).

Kill rules:

• Hard kill: an active or funded player in the same or adjacent space (dev tooling, AI, agents). Worst case is the semantic near-miss (assay.tools was a competing MCP registry; potpie.ai is "AI agents for your codebase").
• Territory kill: if 2–3 candidates from one metaphor family die to same-space players, the whole family is camped — switch families, don't keep mining it. (Works-while-you-sleep died whole: Nightshift ×3, NightAgents, Ant Farm.)
• Soft kill: several small OSS tools share it — you'd never own search results. Also: a name whose search results are owned by a huge unrelated industry phrase (Ghost Kitchen → Otter/Toast forever).
• Connotation kill: borrowed terms carry borrowed fates — don't name things after business models that publicly flopped, slang with a second meaning, or sociology terms with baggage.
• Usually fine: big trademarks in unrelated industries/classes (Signet Jewelers, Anchor Brewing, literal bakeries). Famous-mark exceptions apply (don't name software "Crockpot" — famous marks reach across classes).
• A taken .dev alone kills nothing — name.sh, .tools, get<name>.com, or a subdomain of a domain you already own (<name>.yourdomain.com) all work.

Incumbents-are-evidence does NOT apply to names. For ideas, competition validates demand. For names, a same-space incumbent means you can never own the word — names must be ownable; ideas don't.

Step 4 — Judge and present

Score finalists, roughly in order:

1. Partner test — zero explanation, zero wince, spoken happily.
2. Story fit — does the metaphor match what the product actually does?
3. Brand language for free — vocabulary, badge verbs, newsletter name, CLI/MCP names fall out unforced. Bonus points for a double register (a "Bakeshop" where "the shop" = bakery AND shop floor).
4. Design motif potential — can you see the identity?
5. Collision status + an obtainable domain (tiebreaker).

Present: one recommended pick with an honest reason it wins, 3–5 runners-up (metaphor + collision status + available domain), and the graveyard with cause of death (append it to the graveyard file, see below — never relitigate). Cite collision sources as links.

Then offer, don't do: the deep-vet — GitHub org, npm bare + qualified package, X/Bluesky handles, trademark scan (USPTO Class 42), wider domain sweep.

Step 5 — Rename rounds (revisiting a shipped name)

• The cheapest moment to rename is now. Rename tax only grows — every webhook, board name, blog post adds to it. If the founder is second-guessing, resolve it before the next build phase, not after.
• A challenger must beat the incumbent on gate #1 or story fit, not just tie — ties go to the incumbent (rename tax is real even at its minimum).
• A name that survives the full vet but still draws a founder wince is the founder's call alone. Present the evidence, recommend honestly, let the founder decide. A "hold" can also mean "tabled" — never inflate a provisional hold into a settled decision in the records. (Precedent: an incumbent name survived a 12-candidate rename round; the founder tabled the decision, so the incumbent stayed as the temporary name and the round's graveyard stood either way.)
• Don't let a rhyme/family-pairing with another unsettled name cast the deciding vote.

Optional: the name radar

For a name being watched rather than decided: a tiny script sweeping a candidate/watchlist via the RDAP loop above (weekly cron/LaunchAgent, append verdicts to a notes file) costs nothing and catches expiry drops. Taste generates the watchlist; the radar only reports claimability. Build it on request — it's not part of a normal pass.

The graveyard (causes of death — do not relitigate)

Keep a running graveyard in a notes file the user names (default: NAMING.md in the project, or a naming note in their notes folder). Read it before generating, append to it after every round. One block per round, dated, with the outcome and each dead name's cause of death, e.g.:

**2026-07-10, <project> round (outcome: <pick / TABLED>):** Assay (assay.tools =
competing MCP registry) · Vouch (3 agent-identity startups) · Gantry ($28M
AI-trust co) · Kiln (kiln.tech, AI dev platform) · Oven (Bun's company, oven.sh)

Standing single-noun graveyard for dev/AI tooling (all camped as of 2026, skip them): Sentry, Waypoint, Checkpoint, Turnstile, Harbor, Atlas, Pulse, Lighthouse, Forge, Anvil, Litmus, Compass, Beacon, Ledger, Codex, Vega, Depot, Watchtower, Armory, Polaris, Lodestar, Crucible, Mint, Sterling, Karat, Drydock, Catalyst, Clockwork, Conveyor, Windmill, Foundry, Factory, Assembly, Crew, Autopilot, Flywheel, Greenlight.