COMPACTION AND SUMMARIZATION

ARTIFICIAL INTELLIGENCE ASSISTANT PROMPTS Prompt 1 asks an assistant to compact a history and to be honest about what was dropped. [START OF PROMPT] Compact this agent run history into a short summary that preserves every decision, dead end, and open problem, and tell me exactly what you dropped as low-value: [PASTE LAB OR PRODUCTION RUN HISTORY / TRANSCRIPT HERE]. Output format: the compacted summary followed by a "dropped content" list with justification for each drop. Keep exact identifiers such as file names, numbers, and error codes word for word. [END OF PROMPT] Prompt 2 asks for a threshold design. [START OF PROMPT] Recommend a token-threshold trigger and compaction strategy for this agent given its typical task profile: context window = [SIZE], typical task length = [TYPICAL TOKEN USAGE / STEP COUNT], constraints = [CONSTRAINTS]. Justify the specific percentage threshold chosen, and specify exactly which recent messages stay uncompacted versus which get summarized. Output format: a short design spec with the threshold number, the rationale, and the keep and summarize boundary rule. Include the largest single step you assume and show how it affects the threshold. [END OF PROMPT]