Peter Yang 教程:5 步给 Claude Skills 加上自评分与记忆

Peter Yang 教程:5 步给 Claude Skills 加上自评分与记忆

Source: <https://x.com/petergyang/status/2062181445746192497&gt;

Author: @petergyang

Original Date: 2026-06-04

Quality Score: 5

Fetched: 2026-06-05 20:24:48

Summary / 摘要

EN: Peter Yang releases a 5-step video tutorial for Claude Skills that adds a self-evaluation loop and memory so the skill improves with each use. Targets Claude Code users who maintain personal knowledge bases, fixed styles, or custom workflows in Claude Code.

中文: Peter Yang 发布 The Only Claude Skills Tutorial You Need 视频教程, 5 步教用户为 Claude Skills 加上评估循环 (让 AI 自我修正错误) 与记忆功能 (让技能随使用越多越准). 教程重点不是 skill 本身, 而是把会用升级到会自改进: 用一次, skill 保留反馈, 下次调用时按更新后的版本执行. 适合在 Claude Code 里维护个人知识库固定风格定制工作流的人. 原帖强调技能能精确编码个人知识与品味, 是当前 Claude Code 体系里被低估的能力.

Original Tweet / 原始推文

@petergyang (Wed Jun 03 14:35:44 +0000 2026):

My AI skills now grade themselves and get better the more I use them.

My new tutorial walks through exactly how to build skills with:

→ An eval loop to have AI fix its own mistakes

→ Memory so the skill improves over time

Skills are honestly incredible for encoding your knowledge and taste, and I can't get enough of them.

📌 Watch now: https://t.co/u434ytNSZd

*Engagement: 65 likes, 6 retweets*

中文翻译 / Chinese Translation

Peter Yang 发布 The Only Claude Skills Tutorial You Need 视频教程, 5 步教用户为 Claude Skills 加上评估循环 (让 AI 自我修正错误) 与记忆功能 (让技能随使用越多越准). 教程重点不是 skill 本身, 而是把会用升级到会自改进: 用一次, skill 保留反馈, 下次调用时按更新后的版本执行. 适合在 Claude Code 里维护个人知识库固定风格定制工作流的人. 原帖强调技能能精确编码个人知识与品味, 是当前 Claude Code 体系里被低估的能力.

Notable Replies / 热门回复

@petergyang (❤️ 2):

Also available as a written guide: https://t.co/qYNLCYn1UF

@bettercallsalva (❤️ 1):

@petergyang self-grading skills work until the eval becomes the weak link. if the model writes its own rubric and scores against it, the skill can look like it's improving while drifting from what you wanted. how are you anchoring the eval to ground truth?

@delvynokrstudio (❤️ 0):

@petergyang This pattern is the missing piece for most AI tooling. We tried it for evaluating OKR quality at OKR Studio - skill grades a key result, user accepts or rejects, skill learns the team's bar over time. Took about 3 weeks before it actually became useful.

@adelbucetta (❤️ 0):

@petergyang that's just iterating on meta-learning, not a real breakthrough most self-improving models still rely on hand-coded objectives

@damoosmann (❤️ 0):

@petergyang Skills are my favorite part of the workflow. I build new ones constantly.

Self-grading is the piece I keep wanting. The skills I use daily get stale fast and I rewrite them by hand.

*Content fetched via opencli twitter thread command. Entry ID: dba5fe77*