Agent审查
- 作者仓库星标 29
- 许可证 MIT
- 作者更新于 实时读取
- 作者仓库 ai-engineering
- 领域
- AI 智能 · meta · improvement · skills
- 兼容 Agent
-
- Claude Code
- Cursor
- Cline
- Codex
- Windsurf
- Gemini CLI
- +20
- 信任分
- 100 / 100 · 已通过审计
- 作者 / 版本 / 许可
- @arcasilesgroup · MIT
- Token 消耗评级
- 低消耗
- 接入复杂程度
- 即装即用
- 是否需要外部 API Key
- 不需要
- 兼容的系统
- 未声明(默认跨平台)
- 底层运行要求
- Python
- 文件与系统权限
-
- 只读
- 允许写入 / 修改
- 网络行为
- 仅限本地
- 安装命令数
- 26 条
档案由构建时根据 SKILL.md 与安装命令自动衍生,可能与作者实际意图存在差异。
---
name: ai-skill-improve
description: Improves an existing skill based on real project pain (prior eval corpora under .ai-engineering/…
category: AI 智能
runtime: Python
---
# ai-skill-improve 输出预览
## PART A: 任务判断
- 适用问题:提示词、Agent 工作流、模型评估或自动化推理。
- 输入要求:目标材料、限制条件、期望输出和验收方式。
- 证据边界:围绕“Quick start / Workflow / When to Use”读取原文规则,不把推断写成作者承诺。
## PART B: 执行结果
- **01** 任务判断:确认你的需求是否属于提示词、Agent 工作流、模型评估或自动化推理,并标出输入、限制和预期结果。
- **02** 执行计划:优先按“Quick start / Workflow / When to Use”拆成步骤,说明每一步会读取什么、修改什么、产出什么。
- **03** 交付结果:给出可复制的命令、文件改动、检查清单或内容草稿,并说明如何继续迭代。
- **04** 风险边界:结合 读取文件、写入/修改文件、主要在本地完成、通常不需要额外 API Key 给出执行前确认项。
## Running Rules
- 读取文件、写入/修改文件;主要在本地完成;通常不需要额外 API Key。
- 先小样例验证,再放大到真实任务。
- 交付时同时给结果、检查口径和下一步迭代建议。 原文出现了 `/ai-skill-improve`、`/ai-memory`、`/ai-scaffold`、`/ai-ide-audit`、`/ai-plan` 这类斜杠命令;如果你的 Agent 支持命令触发,优先用命令开场,再补充目标和边界。
告诉 Agent 目标文件或材料、期望结果、不可改范围、是否允许联网或执行命令。本 Skill 的权限画像是:读取文件、写入/修改文件。
先用一个小任务确认它会围绕“Quick start / Workflow / When to Use”工作;涉及文件或命令时,先看 diff、日志、预览或测试结果。
检查最终产物是否包含明确结果、必要证据和下一步动作;如果输出泛泛而谈,就补充输入、边界和验收标准后重跑。
---
name: ai-skill-improve
description: Improves an existing skill based on real project pain (prior eval corpora under .ai-engineering/…
category: AI 智能
source: arcasilesgroup/ai-engineering
---
# ai-skill-improve
## 什么时候使用
- 用于审阅代码、文档或方案并给出可执行反馈 适合处理AI Agent、提示词、模型评估与自动化推理,核心价值是把输入、判断、执行、验证和交付边界固定下来,避免 Agent 泛泛回答。 围绕 meta、improvement、skills…
- 面向提示词、Agent 工作流、模型评估或自动化推理,优先处理能明确输入、步骤和验收标准的工作。
## 需要提供什么
- 目标材料、目录范围、期望结果和不可改动内容。
- 是否允许联网、执行命令、读写文件或调用外部服务。
## 执行规则
- 围绕「Quick start / Workflow / When to Use」组织步骤,不把推断写成作者事实。
- 读取文件、写入/修改文件;主要在本地完成;通常不需要额外 API Key。
- 先跑小样例,确认结果可检查后再扩大任务范围。
## 输出要求
- 给出最终产物、关键证据、验证方式和下一步动作。
- 信息不足时标记 unknown,不编造命令、平台或依赖。 作者原文负责流程事实;仓库文件负责来源和命令;流狐只补充适用场景、限制和质量判断。
skill "ai-skill-improve" {
输入层 -> 用户目标 + 目标文件 + 禁止范围 + 验收标准
上下文层 -> Quick start / Workflow / When to Use
规则层 -> SKILL.md 触发条件 / 执行顺序 / 输出格式
运行层 -> Python | 读取文件、写入/修改文件 | 主要在本地完成
安全层 -> 通常不需要额外 API Key + 小任务验证 + diff / 日志复核
输出层 -> 可复制结果 + 检查清单 + 下一步迭代
} ai-skill-improve
Quick start
/ai-skill-improve ai-plan # evolve one skill
/ai-skill-improve all --dry-run # preview every skill
/ai-skill-improve all # batch evolve with evals
Workflow
Improve existing skills using evidence from real project pain (prior eval corpora under .ai-engineering/evals/, Engram cross-session observations via MemoryPort, LESSONS.md operator notes, decision-store, instincts, proposals). The skill owns pain diagnosis and rewrite strategy; it delegates the eval/grade/benchmark pipeline to Anthropic's skill-creator. Output is PR-comment only — never auto-merged (sub-007 M6).
- Phase 0.5 — load corpora (
.ai-engineering/evals/<skill>.jsonl), Engram observations (/ai-memoryMCP), andLESSONS.mdH3 sections that mention the target skill. - Phase 1 — load remaining pain context (decision-store, observations.yml, proposals.md).
- Phase 2 — analyze the target skill, score the 5 dimensions.
- Phase 3 — generate test prompts that exercise the failing pattern.
- Phase 4 — rewrite the skill (Start-Here, pain-injection, scope-gates, structured classification).
- Phase 5 — emit the proposed SKILL.md diff as a PR comment via
gh pr comment. Do not commit or push. Operator review is the merge gate. - Phase 6 — verify improvement on the operator's branch (pass-rate delta vs prior iteration).
Detail: see audit document skeleton, the six-phase protocol (load → analyze → generate → rewrite → eval → verify), batch mode for
all.
When to Use
- A skill keeps producing bad output despite correct instructions.
- You've accumulated corrections in LESSONS.md that a skill should already know.
- After a batch of sessions where the same skill pattern failed repeatedly.
- Periodic hygiene: evolve the top 10 skills once a month.
- NOT for creating new skills from scratch — use
/ai-scaffold. - NOT for platform audit — use
/ai-ide-audit.
Step 0 (load contexts): read .ai-engineering/manifest.yml providers.stacks; load .ai-engineering/overrides/<stack>/conventions.md for each stack and .ai-engineering/overrides/_shared/conventions.md; load .ai-engineering/team/*.md for team conventions.
Common Mistakes
- Rewriting before reading the pain profile.
- Skipping
--dry-runon batch (you'll burn rate limits). - Inventing test prompts that mirror the skill's own examples (no drift signal).
- Leaving Phase 5 evals unrun and declaring the skill "improved".
Examples
Example 1 — single-skill evolution from accumulated pain
User: "the /ai-plan skill keeps producing decomposition that ignores constraint X. Improve it."
/ai-skill-improve ai-plan
Loads pain context from LESSONS.md and proposals.md, scores ai-plan on 5 dimensions, generates 2-3 test prompts that exercise the failing pattern, rewrites SKILL.md, hands off to skill-creator for eval, reports the delta.
Example 2 — dry-run batch preview
User: "preview what improving every skill would change before I commit time to running evals"
/ai-skill-improve all --dry-run
Walks every skill in priority tier order, shows the proposed diff per skill, and stops short of running the eval pipeline.
Integration
Reads: decision-store.json, LESSONS.md, observations.yml, proposals.md, manifest.yml. Writes: target SKILL.md files. Calls: python scripts/sync_command_mirrors.py after rewrites. Delegates to: Anthropic skill-creator (eval/grade/benchmark, Phase 5). Feeds into: /ai-learn. See also: /ai-scaffold (new skills), /ai-ide-audit (cross-IDE).
$ARGUMENTS
先判断是否适合
作者设计意图
作者的方法与取舍
边界和复核