Prompt
The text you hand to a generative model to shape its output — instructions, context, worked examples, and the user's question.
Last updated
A prompt is how a generative model is “programmed” without code. It bundles instructions, context, and the user’s question into a single block of text that is sent to the model. Prompts are how two users of the same model can get very different answers.
Why it matters for your research. When a colleague shows you something impressive an LLM produced, the prompt matters as much as the model — often more. Replicable work requires capturing the prompt alongside the output; subtly worded prompts can bias the answer in ways the reader cannot see.
In Archēglyph. Not user-facing. We don’t accept free-form prompts over your corpus; the analysis pipeline runs deterministic plugins with recorded inputs and outputs. Where our internal code prompts a model (e.g. a layout-assessment VLM), the prompt template is part of the plugin and versioned with it.
Not to be confused with. A prompt is not a question to a database — the model is not consulting your corpus unless a retrieval step is bolted on first. See Retrieval-augmented generation.