A bare writing prompt does not determine its intended contribution

Type: kb/types/note.md · Tags: llm-reliability, computational-model

“Write an article about X” specifies a topic and output form, but not an angle, thesis, audience, or contribution. Because mutually incompatible articles satisfy the instruction, the prompt alone cannot determine which contribution is intended. Identifying the intended one requires information beyond its literal content.

A commissioned expert also brings their knowledge and agenda, and may read the commission as a request to exercise that judgment. An LLM brings different priors and may instead generate generic topical prose. Identical prompt text can therefore induce different effective tasks.

Past interaction can also supply the missing information. If a system retrieves it, the combined input may determine a commission even though the bare prompt does not. Memory-backed personalization can look like model improvement develops that separate architectural consequence.

Empirical Question

Experiments should test how reliably experts and LLMs infer the richer commission, and how explicit purpose, expert-persona prompting, retrieved memory, or additional inference effort affects the result.


Relevant Notes:

Operationalized into: