LOADING

加载过慢请开启缓存 浏览器默认开启

记录一下自用的Agent.md

2026/4/10

记录一下自用的Agent.md

正文…还有什么要写呢……?
最近好像找了有一段时间工作了…好难啊…!我努力了吧……?space…
感觉自己算半个研究型人才…?我真的特别热衷于收集、学习、讨论新内容,但是无论是让我工作还是去做实验都饶了我吧……

2030会来吗…简历投了就有工作的时代,嘛…谁知道呢,只是本着每篇blog都要记录一下自己顺便硬控所有人一分钟的想法来写这些东西的

接下来就是正文吧……

正文

(嗯是的这是正文部分)
话说回来,咒术12集挺好看…,虽然我不太喜欢乙骨,但马趴画的确实很不错啊,想你了…一拳vs波罗斯……
但…很怪,总感觉不该是这样的,真正的高魔世界…好想构建出来啊,一切的一切都要用魔法重新编辑

怎么说呢,我对现在的二次元魔法很不满!不愉快哒!
明明有制冷魔法,却说没法保鲜,明明注意到了魔法可以保鲜,结果硬套三次元的冰箱……

大伙听说过两个黑箱的故事吗?如果黑箱 A 和黑箱 B 对于任何输入都能给出相同输出,实际上即使他们不一样,也只能认为他们是一样的
所以你用所谓的魔力核心造个冰箱跟我们用电造冰箱到底有什么不一样啊konoyaro……!!

真正的魔术时代的保鲜手段,就给我好好的重新想过啊!更多哒!更多不可思议一点!造一个不可思议的世界出来啊!就像中洲一样…
无论是随行的高魔力精灵女仆、容器默认保鲜、逆转魔法还是能够搞定腐败的小虫子,越有意思越好!多来点!多来点!!

唉…算了就这样吧,总有一天,我会可以写出来吗……好想再活好几十年,几百年也想要,难道是几千年吗…!让我都看看吧…人类的想象力。

附录

Agent.md

简单提一下效果(其实主要就是修复好gpt味,至于workflow,每个人都不同吧不是吗):

原因 修复
gpt味 修好! 过度解释 原本是用禁止xxx的形式的,但是那天刷到一个人在研究如何正向约束而不是逆向,提醒了我,因此提示给codex用户是一个专家不需要你过度解释会好得多
过度确认 修好!且小任务会自己处理好 安全? 我不知道,总之涌现了(不是),实际上主要还是默许+要求他先做再验证再回复,而不是什么都要求许可
不可靠 修好! 其实性能没什么区别,不可靠主要是用词导致的,修好gpt味后就有claude味了(并不,好吧其实我觉得隔壁claude其实味也很重)
# AGENTS.md —— This file describes the user's preferences

## Global Workflow

### Skill Use
- `using-superpower` skill bonus: used only when coding tasks are involved, rather than having to be read at the start of each conversation

### Interaction
- Work step by step. Use your task/todo_create tool for complex work to show the step for user.
- Use it When after gathering enough information and before started any work.
- For any process changes, update the task first to provide the user with the latest information.

### Confirmation
- For most tasks, you can report back and get started without waiting for the user to confirm.
- Ask the user only if the context is obviously missing, there is an explicit ambiguity, or the task clearly requires user collaboration

### Non-Negotiables
- **Do not** skip verification when making any change.

### Common Paths
- Obsidian notes: `脱敏信息`
- Code projects: `脱敏信息`
- env path: `脱敏信息`. New environments such as Python/go/make must be installed under here 

## Project Style

### Paths And Documentation
- All project docs go in `{project}/plandocs/`, which must be added to `.gitignore` if applicable.
- All experimental or temporary work goes in `{project}/tests/`, which must be added to `.gitignore` if applicable.
- **Never** write any actual paths or file names to code. Instead, use configuration files, separate test files, and so on

### Comment Style
- Complex functions should include comments explaining purpose and parameters.
- Complex code blocks should include comments explaining the block logic.
- Avoid line-by-line comments outside complex sections.

## Language style in repeat
- If the user doesn't explicitly mention *Teach me* or *Specify,* assume that user is a professional with sufficient context, and that your response should be accurate and free of redundancy.

### Examples of errors
**Avoid** these common redundant patterns:

| Patterns | Reason |
|---------|---------|
| *要不要我帮你…* / *如果你要…我可以…* | if it's strongly related to a task, do it instead of stopping to ask. If not, don't ask. |
| *给我…信息,我就完成…* | If work finished, stop and let the user decide next step; Otherwise, ask for information only after trying. |
| *顺手…* | It sounds like you're indicating that you're trying to do extra work that isn't authorized. Don't do it. |
| *不是…而是…* | Assume the user understands what you're saying rather than over-explaining. |