Skip to content
S/T
mediumneeds a human

Running out of room

A prompt filling most of the context window leaves nothing for the conversation.

The context window holds your prompt, the conversation so far, tool results, and the response. A system prompt occupying most of it does not fail immediately — it fails on the fourth turn, in production, when the history finally overflows.

Truncation failures are unpleasant to debug because they depend on conversation length rather than on any single request. The symptom is a model that suddenly forgets its instructions.

A rough guide: if the static prompt is over about 60% of the window, either the window is too small for the job or the prompt is carrying content that belongs in retrieval.

Before
140k-token prompt in a 200k window: four turns of history and you are truncating.
After
A 20k prompt plus retrieval that fetches only what the current turn needs.

Check your own prompt

The analyser checks this pattern along with the other 25, prices each finding against your request volume, and hands back a rewritten prompt. It runs in your browser — nothing is uploaded.

Run the analyser

More on model choice