// Strategic compression — Manual compression of contexts at logical boundaries rather than arbitrary automatic compression. Key decisions and constraints are preserved and intermediate exploration is discarded.
| name | strategic-compact |
| description | Strategic compression — Manual compression of contexts at logical boundaries rather than arbitrary automatic compression. Key decisions and constraints are preserved and intermediate exploration is discarded. |
In long sessions, the context window is limited. This skill guides when to compact, what to keep, and what to discard.
Core principle: **Compress at logical boundaries, not at any time. **
| Current stage | Next stage | Whether to compress | Reason |
|---|---|---|---|
| Research/Exploration | Planning | Yes | Exploration details do not need to be brought into planning |
| Plan | Implement | Yes | Keep the plan, discard the planning process |
| Implementation step N | Implementation step N+1 | No | Compression during implementation will lose context |
| Implementation complete | Validation | Optional | If the context is close to the upper limit |
| Verify | Submit | No | Verification results need to be submitted |
| Task A completed | Task B started | Yes | Compression between different tasks |
Must be retained:
Can be discarded:
When it feels like the context is running out:
Audit AGENTS.md / CLAUDE.md against the five high-leverage patterns (progressive disclosure, procedural workflows, decision tables, production code examples, domain rules with concrete alternatives). Reports per-pattern coverage, anti-patterns, and a prioritized fix list.
Shell 脚本中 awk 的 POSIX 兼容性指南。 Use when: 编写或审查包含 awk 的 shell 脚本， 尤其是需要 macOS + Linux 跨平台运行的场景。 触发词: awk, BSD awk, POSIX regex, [[:space:]], guard 脚本, 跨平台 shell
Assessment-driven development — Quantify code generation quality with pass@k / pass^k metrics, automatically scored by Grader.
Iterative retrieval — 4-stage loop (DISPATCH→EVALUATE→REFINE→LOOP) to pinpoint relevant information in the code base. Up to 3 rounds.
Post-hoc diagnosis of a failed agent trajectory. Classifies the first unrecoverable step into one of nine failure categories (plan adherence, hallucinated information, invalid tool call, misread tool output, intent–plan mismatch, under-specified intent, unsupported intent, guardrail trigger, system failure) and produces an evidence-backed root-cause report.
AI-assisted development of anti-hallucination specifications. Check out the seven-layer defense architecture, quantitative indicators, execution templates and practical cases. Used for code review, task startup inspection, and weekly review.