Prompt Compression

Reduces prompt token count without losing effectiveness. Covers semantic deduplication, example pruning, instruction consolidation, and format directive compression. Critically: distinguishes between safe cuts and cuts that silently degrade output.

Phase 1 — Audit Before Cutting

Read the full prompt. Map it into buckets before touching anything:

• Role/persona block — who the model is
• Behavioral constraints — what it must/must not do
• Task instructions — what to actually produce
• Format directives — how to structure output
• Examples — few-shot demonstrations
• Context/background — grounding information

Compression failures almost always come from cutting across buckets without noticing. A constraint buried in an example isn't redundant with the same constraint stated in instructions — it's reinforcement. Flag these before cutting.