Back to Details

kf-learn-from-correction

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

Learn From Correction

从修正中学习

Extract reusable lessons without accumulating uncontrolled instructions.
提取可复用的经验，避免积累不受控的指令。

Workflow

工作流程

  1. Compare the prior agent-generated work, the user's correction, the explicit feedback, and relevant repository context.
  2. Isolate deliberate material differences. Ignore incidental edits, formatting, and facts already apparent from code, configuration, or documentation.
  3. Decide whether persistence is justified by explicit durable user direction, repeated corrections, or a repeated repository pattern. Treat weaker evidence as a one-off requirement.
  4. Choose the narrowest useful destination: global or project 
    AGENTS.md
    , an existing K Fleet skill, project docs, a canonical example, or nowhere. Avoid direct duplicates; route broader cleanup of existing guidance to 
    kf-maintain-guidance
    .
  5. If persistence was not authorized, present the focused proposal for review. If the user already authorized the relevant update, apply it without requesting redundant confirmation.
  1. 对比之前Agent生成的内容、用户的修正、明确的反馈以及相关的仓库上下文。
  2. 分离出刻意的实质性差异。忽略偶然的编辑、格式调整以及从代码、配置或文档中已能明显看出的事实。
  3. 根据用户明确的持久化指示、反复的修正或仓库中的重复模式，判断是否有必要持久化此类经验。证据不足的则视为一次性需求。
  4. 选择最窄范围的合适存储位置：全局或项目级的
    AGENTS.md
    、现有的K Fleet skill、项目文档、标准示例，或者不存储。避免直接重复；将对现有指导的更广泛清理工作交由
    kf-maintain-guidance
    处理。
  5. 如果未获得持久化授权，则提交聚焦的建议供审核。如果用户已批准相关更新，则直接应用，无需请求重复确认。

Report each proposed lesson

报告每个拟议的经验

text
Decision: persist or do not persist
Lesson
Evidence
Destination
Suggested wording
Confidence
text
Decision: persist or do not persist
Lesson
Evidence
Destination
Suggested wording
Confidence

Constraints

约束条件

  • Do not turn every manual edit, isolated preference, or one-off requirement into a permanent rule.
  • Use behavioral guidance only when it will improve future decisions. Prefer project docs or canonical examples for architecture and product facts.
  • Keep wording concise, imperative, and narrowly scoped. Do not duplicate guidance already clear from the repository.
  • Base confidence on evidence strength, not wording specificity.
The goal is controlled learning from real usage, not automatic background learning.
  • 不要将每一次手动编辑、孤立偏好或一次性需求都转化为永久规则。
  • 仅当行为指导能改善未来决策时才使用。对于架构和产品相关事实，优先使用项目文档或标准示例。
  • 措辞要简洁、命令式且范围明确。不要重复仓库中已明确的指导内容。
  • 信心等级基于证据的强度，而非措辞的具体性。
目标是从实际使用中进行可控的学习，而非自动的后台学习。