NLP & Language · beginner · concept 117 of 311 · Updated Nov 4, 2026

🗜️Prompt and Context Compression

Shortening what goes into a model's context window without losing what the task needs, because every token costs money and time. Methods range from dropping low-information tokens with a small model (Microsoft's LLMLingua reports up to 20x compression with little loss) to packing an instruction into a few learned gist tokens, to agent habits such as summarizing a long transcript into notes. Claude Code's compaction is an everyday example.