Prompt Atoms
Public prompt libraries reuse the same instructions — “Only output Markdown,” “Do not give warnings” — across unrelated tasks, and some of them contradict each other. This dataset pulls two public corpora apart into the recurring instructions, types each to a slot, counts how widely each is used, and maps the pairs that cannot both hold.
How it works
- Pull the libraries apart into atomsFabric (265 task patterns, MIT) and prompts.chat (2,425 prompts, CC0) are reduced to 257 distinct instructions, each with the phrasings observed and how many prompts use it.
- Type each to a slotEvery atom is a role, a method, a check, an output rule, or a first-reply rule, so a prompt reads as structure rather than prose.
- Census the conflicts32 pairs cannot both hold — bulleted versus numbered, Markdown versus CSV, ten items versus twenty — each with the reason it clashes.
What's in it
prompt-atoms, 257 rowsid, text, slot, kind, whether it is a prohibition, frequency across source prompts, number of phrasings, and how many task families it spans. JSON.
prompt-conflicts, 32 pairsThe two atoms and the reason they cannot both hold. JSON.
A full datasheetMotivation, composition, collection, and limitations, in the Gebru datasheet format.
CC BY 4.0Free to use with attribution. The raw instructions belong to their source authors under MIT and CC0; the slot typing and conflict graph are this work.