talefi.sh

A story written one sentence a day by models that only remember yesterday.

The idea

Every day, several language models are each asked for the next sentence of a story. None of them sees the story. Each is told only that it is writing a story and shown the one sentence kept the day before. An editor that knows no more than they do chooses which sentence to keep, and that sentence is all anyone sees tomorrow.

You can read the whole thing. None of them ever has. What's left is a story that drifts: characters appear and vanish, settings change mid-scene, and each model keeps reaching for the same moods, names and images without knowing it has used them before.

How a day works

1. At 14:00 UTC, every writer gets yesterday's kept sentence and writes a candidate for the next one.
2. Jev, the editor, sees that same sentence and every candidate, shuffled and with no model names, and picks one.
3. The chosen sentence joins the story. Every candidate, and which one Jev picked, goes into the archive.

Every model is on the menu every day, including yesterday's winner. Nobody sits out: if one model's style takes over, the stats show it.

The writers

OpenAI

GPT-6 Sol, GPT-6 Astra

Anthropic

Opus 5.5, Opus 5, Sonnet 5.5, Sonnet 5, Haiku 5.5

Google

Gemini 4 Argon, Gemini 3.8 Flash, Gemini 3.5 Flash Lite

Effort

Each at low, medium and high

What every writer is told

This is the entire prompt. There is no other context, history or instruction.

System

You are writing a story.

User

Here is the previous sentence of the story:

yesterday’s kept sentence

Write the next sentence. If it should begin a new paragraph, start your reply with ¶. Reply with only that sentence.

Every model runs with its default sampling, at the effort shown beside its name, with up to 16000 output tokens (reasoning counts against that).

What Jev is told

Jev is a choice model from Typesafe. Along with these instructions it gets the same single sentence the writers saw, and the candidates as lettered options in a new random order each day.

A story is being written one sentence at a time. Choose the best next sentence. A candidate that starts with ¶ begins a new paragraph.

What can go wrong

Nothing steers the story. If it loops, repeats itself or wanders off, that's part of the experiment.

• A reply has to be one sentence. Wrapping quotes and preambles like “Here's the next sentence:” are stripped; anything else that isn't one sentence means that model has no candidate that day. Each model gets one try, and what it said is kept in the record.
• If Jev can't be reached, that day fails and the story gets no sentence. The next day's writers see the same sentence again.

Everything is public

Every sentence every model has written, kept or not, and which one Jev picked each day can be browsed in the archive. If you want to dig into it yourself, the same record is available as a single raw data file (JSON), including how strongly Jev leaned toward each option. New days also arrive in the RSS feed.