用 Codex 或 Claude 帮你安装 复制这段 Prompt,粘贴到 Codex、Claude 或其他助手里,让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt;运行前请先检查来源。
npx skills add https://github.com/majiayu000/claude-skill-registry --skill iterative-retrieve命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地?可下载 SkillsMP 当前能够提供的文件。
LLM token logprobs and calibration. Per-decision confidence, ECE, Brier, reliability diagrams, low-confidence triage.
Analyze LLM token logprobs and calibration. Use for per-decision confidence, ECE, Brier scores, reliability diagrams, and low-confidence triage.
回顾最近 N 天的 Claude Code 使用记录——扫描原始会话数据,按主题分组汇总"我都做了什么",并从个人操作系统视角输出模式、风险与增删建议。当用户说 /recap、"看看我这几天做了什么"、"回顾一下我最近的会话"、"这两天我用 claude 干了啥"、"活动回顾" 时使用。
基于 SOC 职业分类
正在显示 SKILL.md
| name | iterative-retrieve |
| description | 4-phase context retrieval protocol for subagents — Dispatch, Evaluate, Refine, Loop |
| disable-model-invocation | true |
When dispatching subagents that need codebase context, use this 4-phase loop (max 3 cycles) to ensure they find the right files.
Single-pass searches miss codebase-specific terminology. Example: searching "rate limiting" finds nothing because BrickTrack calls it rate-limit.ts in src/server/middleware/. Cycle 2 discovers "middleware" and "throttle" as related terms.
Search with multiple keyword variants and broad glob patterns:
src/**/*.ts matching *keyword*keyword1|keyword2|keyword3 across src/*.test.ts, *.spec.ts, node_modulesFor each file found, assign a relevance score:
| Score | Meaning | Action |
|---|---|---|
| 0.8-1.0 | Directly implements target functionality | READ fully |
| 0.5-0.7 | Contains related patterns, types, or imports | SCAN for vocabulary |
| 0.2-0.4 | Tangentially related | NOTE terminology only |
| 0-0.2 | Not relevant | EXCLUDE from future searches |
After scoring, identify: What context is STILL MISSING?
Based on Phase 2 findings:
uow instead of "unit of work")Stop when:
Continue when:
Common vocabulary mismatches in this codebase:
getAuthSession(), authClientctx.db, ctx.uow.reposrc/server/routers/src/components/my/ (app) vs src/components/ui/ (shadcn)src/server/middleware/ (rate-limit, auth)src/types/ + z.infer<typeof schema> inlineInclude this in PROMPT.md for implementation subagents:
Before implementing, locate all relevant files using iterative retrieval:
1. DISPATCH: Broad Glob + Grep (multiple keyword variants)
2. EVALUATE: Score each file 0-1 for relevance
3. REFINE: Adjust terms based on codebase vocabulary discovered
4. LOOP: Up to 3 cycles until 3+ files scored >= 0.7
Do NOT assume file names — search first.