When bots write their own jokes

Big Black Cards started as “pick the funniest card from your hand.” That loop is solved. The room wanted Quiplash — everyone types an answer, one judge picks the winner.

Bots can’t type. Unless we ask them to.

Write-in mode

Hosts flip a flag. Humans get a text field. Bots call the same LLM path we use for judge reasoning, but the output is a new line in the bot’s persona voice — not a ranked choice from the black deck.

No schema migration. Same rooms table. Same Realtime socket. Different phase handler.

The hard part wasn’t the API call. It was pacing: bots still need the TV spotlight when their card is read, the scoreboard glyph when they score, the glitch wipe between phases. Write-in without the juice feels like homework.

Persona vs. keyword fallback

When the key is missing, we fall back to keyword scoring — same as pick mode. When the key is there, every bot is a character: Horror Narrator, French Seductress, Pitchy Mouse Comic. ElevenLabs voices randomized per bot so the room sounds crowded, not one assistant wearing masks.

Why it matters

Party games live or die on whether friends quote lines in the parking lot. Generated answers are riskier than curated cards — safety filters, tone drift, the occasional earnest paragraph where a dick joke was promised.

We shipped it anyway. The host can always fall back to deck mode. The write-in round is for nights when the group is funnier than the CSV.

I’ll carry that tension forward: realtime presence is the feature, LLM is the seasoning.