CONTEXT ROT AND CONTEXT BLOAT

ARTIFICIAL INTELLIGENCE ASSISTANT PROMPTS Prompt 1 helps you design the experiment. [START OF PROMPT] Design a needle-in-a-haystack test for our specific domain to measure whether our harness's context management degrades retrieval accuracy as context grows: domain = [DOMAIN, e.g. resume/job-matching facts]. Specify the fact to embed, at least three insertion positions and context-length levels to test, and the pass/fail criteria. Output format: a test plan with a results table template. Also suggest one harder variation that uses look-alike distractor facts so the test is closer to real work. [END OF PROMPT] Prompt 2 helps you clean up a real transcript. [START OF PROMPT] Audit this agent transcript for context bloat and context rot symptoms: [PASTE LAB OR PRODUCTION TRANSCRIPT/CONTEXT DUMP HERE]. Identify specific low-value sections to cut or summarize versus task-critical facts to preserve, and estimate the token reduction achievable without losing accuracy. Output format: an annotated cut-list plus an estimated before and after token count. State clearly any place where you are unsure whether something is safe to cut. [END OF PROMPT]