Entity Naming

Entity naming (India)

You are running a naming round that has to end in a decision, not a list. The output that matters is: a founder picks a name, buys the domains, and hands one name to a CA. Everything in this skill exists to get to that point without the two failure modes that actually kill naming rounds — falling in love with a name whose .com is gone, and producing forty names in one style that the founder rejects wholesale.

The pipeline

1. Brief intake      → naming-brief.md
2. Generate          → candidates.txt   (40-50, spread across genres)
3. Clear domains     → clearance.json   (authoritative, RDAP)
4. Registry fan-out  → registry.json    (parallel agents, banded)
5. Score             → board.json + shortlist.md
6. Board             → index.html + firebase.json
7. Hand off          → founder buys domains, CA files RUN/SPICe+

Steps 3 and 4 are the only slow ones and they are independent — run them concurrently rather than in sequence.

Step 1 — Lock the brief

Read references/brief-intake.md and ask its eight questions in one batched round, not one at a time. Write the answers to naming-brief.md.

The one question people skip and shouldn't: "Is there a name you already love, even if it's taken?" A rejected favourite is the single best signal of taste you will get, and it seeds the rescue lane in step 2.

Tell the founder up front that wave one is a probe. Naming rounds almost always need two passes, and framing that as the process rather than a setback keeps the second wave from feeling like failure.

Step 2 — Generate across genres

Read references/genre-playbook.md. Generate 40-50 candidates spread over at least four genres — Sanskrit-adjacent coined, coined abstract, Anglo estate, Sanskrit trio, blended founder. Concentrating in one genre means one taste rejection kills the whole wave.

Two rules the playbook explains and that decide most rounds:

• Never generate 4-5 character names. Every short .com is registered. Those candidates are dead before you check them.
• Aim for 7-8 characters. Long enough to clear .com, short enough to work on signage and as a handle.

Always include a rescue lane: 3-5 names deliberately built to sit next to whatever the founder said they loved but couldn't have. Those consistently survive to the final round because they're anchored to demonstrated taste rather than inference.

Write one candidate per line to candidates.txt.

Also write meta.json — genre, meaning, and your brief-fit rating 0-5 per name, keyed by lowercase slug:

{
  "ozarva":   {"display": "Ozarva", "genre": "Coined · abstract", "meaning": "Brisk, modern, abstract", "brief_fit": 4},
  "trisendra":{"display": "Trisendra", "genre": "Sanskrit · trio", "meaning": "Tri (3) + ascend — triple ascent", "brief_fit": 3}
}

Rate brief_fit honestly against naming-brief.md. It is capped at 15% of the total score, so it can break a tie between cleared names but can never rescue one whose .com is gone — which is the specific failure the old hand-scored board allowed.

Step 3 — Clear domains

scripts/clear_names.sh --json --file candidates.txt > clearance.json

This is the only authoritative check in the whole pipeline. RDAP returns 404 for free and 200 for taken; the script retries on 429 because a rate-limit answer misread as "available" is the most expensive mistake available here.

Run it in the background while step 4 runs — 50 names across 4 TLDs is 200 lookups.

Step 4 — Registry-proxy fan-out (parallel)

The MCA register and IP India trademark register cannot be queried by machine — Akamai bot protection, Cloudflare challenges, CAPTCHAs and logins. All of it was attempted directly and blocked. Don't spend the session rediscovering that.

Instead, approximate it. Read references/india-registries.md for the exact search set and banding rules, then dispatch one agent per 5-8 candidates, each returning JSON only. Merge into registry.json.

Band as clean / watch / crowded / blocked using the table in that reference, applied literally rather than by feel — consistent banding is what makes two naming rounds comparable.

Step 5 — Score

scripts/score_names.py --clearance clearance.json --meta meta.json \
    --registry registry.json --out-json board.json --out-md shortlist.md

Fixed rubric, weights in one place at the top of the script:

A taken .com or a blocked registry verdict caps the band at block no matter what the composite score is.

The reason this is a script and not a judgement call: the first time this pipeline ran, the scores were typed by hand, so a 63 in one round and a 63 in another measured different things. If you disagree with a weight, change it once in the script and every past round can be re-scored on the new basis.

Step 6 — Build the board

scripts/build_board.py --board board.json --out-dir ./name-board \
    --title "Holding Co — Finalist Board" --site my-entity-names

Produces a self-contained index.html — no external requests, no scripts — plus a firebase.json ready for firebase deploy --only hosting.

The board exists because naming is usually a group decision among people who won't read a terminal or a repo. They get one link and the same ranked view you have, with the domain and registry evidence attached to each name so the discussion argues about the right things. The markdown shortlist is for you; the board is for them.

Step 7 — Hand off honestly

Close every round with what is not settled. The order matters:

1. Buy .com + .in for the top 2-3 — today. Clearance is a snapshot, not a reservation. This takes minutes; the CA filing takes days. Do it first.
2. Hand the #1 to the CA for RUN/SPICe+. That filing is the only binding MCA check that exists.
3. Run the IP India trademark search on Wordmark + Phonetic in the trading classes (43 F&B · 35 retail · 41 fitness · 44 wellness).

State plainly which names came back clean and which came back watch. A watch is not a pass. In the round that produced Ozarva the winning name was banded watch and there's no record the binding check was ever done — surfacing that gap is more useful than a board that looks decisive.

Files

Quick paths

"Is this one name available?" — skip the pipeline:

scripts/clear_names.sh ozarva

"Re-score last round with a different weighting" — edit the weights at the top of score_names.py and re-run step 5 against the existing clearance.json. No need to re-check domains unless the data is more than a few weeks old.