context-engineering
bởi leoyessi10-tech
Smart Home & IoTTóm tắt
This skill should be used when the user asks.
Mô tả
Kỹ năng này làm gì
Covers strategies for compressing agent conversation history when sessions exceed context window limits. Three production-ready approaches are compared: anchored iterative summarization, opaque compression, and regenerative full summaries. The core metric is tokens-per-task rather than tokens-per-request.
Optimizing for tokens-per-task rather than tokens-per-request avoids the hidden cost of re-fetching lost context, which naive compression strategies miss entirely.
Khi nào nên dùng
- Preventing coding agents from forgetting modified files mid-session
- Handling codebases larger than the model context window
- Designing summarization pipelines for long-running agent workflows
- Evaluating compression quality with probe-based testing
- Tuning compression trigger thresholds for multi-phase tasks
Quy trình ví dụ
User asks: compress the current 90,000-token debugging session before continuing
- 1
Identify the structured summary sections: Session Intent, Files Modified, Decisions Made, Current State, Next Steps
- 2
Summarize only the newly-truncated span of conversation history into those sections
- 3
Merge the new section content into the existing persistent summary rather than regenerating from scratch
- 4
Verify artifact trail completeness by checking that all read and modified files are listed
- 5
Run probe-based checks asking factual questions to confirm critical details were retained
Compressed structured summary under 2,000 tokens, with full file tracking and decision history preserved, ready to continue the session
Cài đặt
clawhub install leoyessi10-tech/context-engineeringChạy lệnh CLI, hoặc dán prompt vào Claude Code, Cursor, Codex hay coding agent bạn đang dùng.