brainstorming-names

Brainstorming Names (multi-model tournament)

Overview

Tournament workflow: parallel subagents on randomized models generate name candidates each round, a fixed judge scores them against a stable rubric, and top names carry forward. Model randomization replaces the removed temperature knob — each model family has a distinct creative distribution ("house style"), so mixing models samples a wider name space than any single model can.

Core rule: randomize the generators, never the judge.

Phase 1 — Collect inputs (ask if missing)

1. What is being named: 1-2 sentence product description + audience.
2. If renaming: why the current name fails (collision, generic, legal).
3. Hard constraints — propose defaults, let user adjust: max words/syllables/characters, pronounceable in English, no trademark collisions (name the adjacent products to avoid), CLI-safe if relevant (lowercase, no spaces).
4. Budget: default 5 rounds × 3 agents × 10 names. Scale up only if user asks for thoroughness.

Phase 2 — Fix the rubric before round 1; never change it mid-run

Default weights: memorability & distinctiveness 40%, fit with purpose 30%, pronounceability/spellability 15%, brand risk (collisions, bad meanings in major languages) 15%. Score 1–10 per criterion.

Who judges: the orchestrator (you) scores inline with this rubric, on your own session model — no judge subagent needed. If a /score skill is available, use it with this rubric instead. Either way, the same judge and same rubric every round — the judge is never randomized, otherwise round-1 and round-5 scores are not comparable and carried winners get unfairly re-ranked.

Phase 3 — Tournament loop (per round)

1. Randomize generator models with bash — never pick "randomly" from your head (LLM self-selection is biased toward one model). Run via bash -c (zsh arrays are 1-indexed; this snippet assumes bash):

bash -c 'models=(fable opus sonnet haiku); for i in 1 2 3; do echo "${models[RANDOM % ${#models[@]}]}"; done'

Duplicates within a round are fine — the strategy still differs. Drop haiku from the pool if its candidates' average score is the lowest of all models in each of the first two rounds.

2. Spawn 3 parallel Agent calls, each with model: set from step 1 and ONE strategy from this ordered list: (1) invented/coined words, (2) real-word metaphors, (3) compound/portmanteau, (4) foreign-language roots, (5) mythology/nature/astronomy. Assign 3 distinct strategies per round, cycling through the list in order across rounds (round 1 → strategies 1-3, round 2 → 4, 5, 1, round 3 → 2-4, …). Each subagent prompt MUST contain — subagents see nothing of the conversation:

   • the full product context, verbatim
   • the hard constraints
   • its single assigned strategy
   • the full exclusion list (every name generated in ANY prior round; state "(none — round 1)" in round 1)
   • the current top 2 with scores and the instruction "beat these" ("(none — round 1)" in round 1)
   • output spec: exactly 10 candidates (or the per-agent count chosen in Phase 1) + 1-line rationale each, self-checked against constraints (drop failures and replace before returning)

3. Score new candidates only. Carried winners keep their original scores — re-scoring causes drift that kills good names.

4. Carry top 2 forward. Track per name: round, generating model, strategy, score.

5. Early stop: same #1 for 3 consecutive rounds.

Phase 4 — Deliverable

Final table: top 5 names × (per-criterion scores, total, round discovered, generating model, strategy) + one-paragraph recommendation for the single winner. Include which models produced the finalists — useful signal for weighting the pool next time.

Common mistakes

Mistake	Fix
Randomizing the judge	Scores stop being comparable; fix judge model + rubric for the whole run
Picking "random" models mentally	Biased; use the bash $RANDOM snippet
Re-scoring carried winners	Score drift; keep original scores
No exclusion list	Later rounds regenerate earlier names, wasting the budget
Thin subagent prompts	Subagents lack conversation context; include context block verbatim
Vague constraints ("short, memorable")	Unenforceable; convert to numbers before round 1