Kelly Mears

Prompt

The text supplied to a language model to elicit a response.

Agents & Language Models1 min read183 words10 out · 9 in
also calledPrompting

A prompt is the input given to a Large Language Model. In modern systems it is not a single string but an assembled structure: a System Prompt establishing role and rules, a sequence of conversational turns, the definitions of any available tools, and whatever documents or state have been injected for the task.

Two properties are worth internalising. First, everything in that assembly is text the model reads — including tool descriptions, which are frequently more influential on behaviour than the instructions written for the purpose. Second, everything competes for the same Context Window, so adding instruction has a cost, not just a benefit.

Prompts are also the least type-checked part of most systems. Nothing errors when a prompt describes a capability that was removed, states a rule the runtime does not enforce, or duplicates a fact that has since changed elsewhere. That makes them a prime site for Documentation Rot, and it is the argument for keeping a single source of truth: a tool's description ships beside its implementation and cannot drift, while a catalogue of tools in a separate document will.

See also5

Hand-picked in the note itself — the neighbours worth reading next.

Related2

Nearby in the graph rather than deliberately chosen. Looser, sometimes surprising.

Linked from9

Notes elsewhere in the wiki that reach for this one.