kf-learn-from-correction

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Chinese

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。

Compare the prior agent-generated work, the user's correction, the explicit feedback, and relevant repository context.
Isolate deliberate material differences. Ignore incidental edits, formatting, and facts already apparent from code, configuration, or documentation.
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.
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
```
.
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.

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

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.

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。