📖About

Carol per-turn inference is bottlenecked by the 43KB system prompt. Targets: dedup repeated guard rules in RESPONSE STYLE (~29 lines of similar never-expose-X rules), move CHOWPATTY block to chowpatty service prompt, consolidate MISSING CAPABILITIES + ERRORS + TOOL FAILURE TRANSPARENCY into one section, trim TOOL ROUTING (duplicates voice.py tool descriptions), shorten FACTS VS INFERENCES + FACTUAL ACCURACY + MESSAGE SEQUENCE AWARENESS. Target: <=20KB. Tests verify critical rules preserved (privacy, voice fallback, tier gating, response shape).