有的没的
Prompt Engineering: What Actually Works (Without the 8-Hour Hype)
I’ve seen people drop 8-hour-long videos on prompt engineering, and honestly, my reaction is 🤦♂️.
I won’t bore you with the obvious stuff or overcomplicate things. Instead, I want to share a few practical techniques that actually helped me write better prompts, some common sense, some hard-earned lessons. Most of what I’m sharing comes from the book Hands-On Large Language Models
So here’s what I’ve learned that actually works:
1. Specificity
This one seems obvious, but it’s also the most commonly missed.
A vague prompt gives you a vague answer. The more precise you are about your goal, format, and constraints, the better the result.
Bad Prompt:
Good Prompt:
See the difference? Specific inputs = Specific outputs.
2. Hallucination Guardrail
We all know that LLMs hallucinate, they confidently make stuff up.
A surprisingly simple trick: Tell it not to.
Try this prompt:
This becomes really important when you're designing apps or knowledge assistants. It helps reduce the risk of wrong answers.
3. Order Matters
This was a surprise to me and I learned it from the book.
Where you place your instruction in a long prompt matters. Either put it right at the start or at the end. LLMs often forget what’s in the middle (especially in long prompts).
Example:
Here's a paragraph. Also here's a use case. Here's some random info. Now summarize.
Simple shift, big difference.
Other Techniques That Help Me Daily
1. Persona:
Set the role clearly.
This changes the behavior completely.
2. Audience Awareness:
My favorite when I want to simplify things.
Works brilliantly for breaking down tough concepts.
3. Tone:
Underrated but essential.
Want a formal reply?
4. Instruction / Context:
Always useful.
Gives the model direction and expected output format.
5. Grammar Fixing:
As a non-native English speaker, this one’s gold for me.
It has helped me immensely in writing better content, emails, blogs, even this post :-)
These are the techniques I use regularly. If you have your own prompt engineering hacks, I’d love to hear them, drop them in the comments!
https://www.reddit.com/r/LocalLLaMA/s/5pIxRi5BU6
#ai
I’ve seen people drop 8-hour-long videos on prompt engineering, and honestly, my reaction is 🤦♂️.
I won’t bore you with the obvious stuff or overcomplicate things. Instead, I want to share a few practical techniques that actually helped me write better prompts, some common sense, some hard-earned lessons. Most of what I’m sharing comes from the book Hands-On Large Language Models
So here’s what I’ve learned that actually works:
1. Specificity
This one seems obvious, but it’s also the most commonly missed.
A vague prompt gives you a vague answer. The more precise you are about your goal, format, and constraints, the better the result.
Bad Prompt:
Write something about climate change.Good Prompt:
Write a 100-word summary on how climate change affects sea levels, using simple language for a high school audience.See the difference? Specific inputs = Specific outputs.
2. Hallucination Guardrail
We all know that LLMs hallucinate, they confidently make stuff up.
A surprisingly simple trick: Tell it not to.
Try this prompt:
If you don’t know the answer, respond with ‘I don’t know.’ Don’t make anything up.This becomes really important when you're designing apps or knowledge assistants. It helps reduce the risk of wrong answers.
3. Order Matters
This was a surprise to me and I learned it from the book.
Where you place your instruction in a long prompt matters. Either put it right at the start or at the end. LLMs often forget what’s in the middle (especially in long prompts).
Example:
Here's a paragraph. Also here's a use case. Here's some random info. Now summarize.
Summarize the following paragraph:" [then the content]Simple shift, big difference.
Other Techniques That Help Me Daily
1. Persona:
Set the role clearly.
You are an expert Python developer who writes clean code.This changes the behavior completely.
2. Audience Awareness:
My favorite when I want to simplify things.
Explain this like I’m five.Works brilliantly for breaking down tough concepts.
3. Tone:
Underrated but essential.
Want a formal reply?
Write this in a professional tone for a client. vs Make this sound like I’m texting a friend.4. Instruction / Context:
Always useful.
Summarize the following news article in bullet points.Gives the model direction and expected output format.
5. Grammar Fixing:
As a non-native English speaker, this one’s gold for me.
Fix the grammar and make it sound more natural.It has helped me immensely in writing better content, emails, blogs, even this post :-)
These are the techniques I use regularly. If you have your own prompt engineering hacks, I’d love to hear them, drop them in the comments!
https://www.reddit.com/r/LocalLLaMA/s/5pIxRi5BU6
#ai
最近一直想搞一个黑苹果
CPU：13100f 300
主板：h610 300
显卡：卡诺基rx590gme 后面要刷成 580 400
内存：32G：300
硬盘：1T + 512 免费
机箱：itx 80
电源： 100
但是有没有意义呢？谁来敲醒我
先冷却一阵。。
目前也不是电脑不够用，笔记 32g 10core 的 m1pro 现在正值壮年。
唯一比较烦的就是每次开电脑都得翻盖、连线。还是台式机方便点
#life
CPU：13100f 300
主板：h610 300
显卡：卡诺基rx590gme 后面要刷成 580 400
内存：32G：300
硬盘：1T + 512 免费
机箱：itx 80
电源： 100
但是有没有意义呢？谁来敲醒我
先冷却一阵。。
目前也不是电脑不够用，笔记 32g 10core 的 m1pro 现在正值壮年。
唯一比较烦的就是每次开电脑都得翻盖、连线。还是台式机方便点
#life
windows下因为缺少posix环境 命令行对python不友好
于是发现，win下用来新建虚拟环境啥的，竟然vsc最方便。。
新建一个文件夹，右键使用vsc打开，新建一个py文件，新建环境
一步一步走即可。
配置完成后，vsc内打开的terminal是默认激活了虚拟环境的，非常方便
#programming
于是发现，win下用来新建虚拟环境啥的，竟然vsc最方便。。
新建一个文件夹，右键使用vsc打开，新建一个py文件，新建环境
一步一步走即可。
配置完成后，vsc内打开的terminal是默认激活了虚拟环境的，非常方便
#programming
35 gOYZK-傻呵呵ywj发布了一篇小红书笔记，快来看吧！ 😆 bK80SHq6gzRSnom 😆 http://xhslink.com/m/8izrmclSz9R 复制本条信息，打开【小红书】App查看精彩内容！
44 随心发布了一篇小红书笔记，快来看吧！ 😆 0BSmH67Q7vHAXO0 😆 http://xhslink.com/m/Mrl1TeGijB 复制本条信息，打开【小红书】App查看精彩内容！
今天又学了一招～ sourcetree 里面快速对比两个 commit 版本之间的差别：
按住 cmd，然后在历史中选择两次 commit 就搞定了！
感觉自己好傻= = 这么直白的方法竟然没试过
#programming
按住 cmd，然后在历史中选择两次 commit 就搞定了！
感觉自己好傻= = 这么直白的方法竟然没试过
#programming
configuration for usteer:
https://openwrt.org/docs/guide-user/network/wifi/usteer
manual for wifi roaming on openwrt:
https://openwrt.org/docs/guide-user/network/wifi/roaming
搬家之后 如果 5ghz 存在死角，那么只能开启 2.4g 5g 切换了。这个时候可以使用 usteer 这个包
#life
https://openwrt.org/docs/guide-user/network/wifi/usteer
manual for wifi roaming on openwrt:
https://openwrt.org/docs/guide-user/network/wifi/roaming
搬家之后 如果 5ghz 存在死角，那么只能开启 2.4g 5g 切换了。这个时候可以使用 usteer 这个包
#life
今天对红米ax6000动手术！
参照https://www.right.com.cn/forum/thread-8265832-1-1.html 刷入uboot！
uboot选用hanwckf的mt7986_redmi_ax6000-fip-fixed-parts-multi-layout.bin！！下载地址https://github.com/hanwckf/bl-mt798x/releases/tag/20241115
然后刷入237的rom，需要选110m大分区！默认ip地址192.168.6.1
注意进uboot，上电之后要按住reset 15s！！看起来是有延时启动，所以需要按的时间长一点
ps: 被xiaomi原厂固件的ssh密码admin硬控5分钟。。。最后还是查我tabby的vault查出来的。。实在想不出来这个密码
#ax6000
参照https://www.right.com.cn/forum/thread-8265832-1-1.html 刷入uboot！
uboot选用hanwckf的mt7986_redmi_ax6000-fip-fixed-parts-multi-layout.bin！！下载地址https://github.com/hanwckf/bl-mt798x/releases/tag/20241115
然后刷入237的rom，需要选110m大分区！默认ip地址192.168.6.1
注意进uboot，上电之后要按住reset 15s！！看起来是有延时启动，所以需要按的时间长一点
ps: 被xiaomi原厂固件的ssh密码admin硬控5分钟。。。最后还是查我tabby的vault查出来的。。实在想不出来这个密码
#ax6000
面部做激光已经一周了，结痂还是没开始掉。
实在忍不了了，天天蓬头垢面的，倒不是在意形象，是实在太油了，那种油是妨碍你干任何事情，都觉得不得劲的感觉。
昨天晚上实在忍不了了，洗了一下头。清爽了很多
发现个好方法，实在不行就用淋浴喷嘴喷脸，用纸巾吸吸眼睛，然后再用吹风机热风吹干。也不错。
#life
实在忍不了了，天天蓬头垢面的，倒不是在意形象，是实在太油了，那种油是妨碍你干任何事情，都觉得不得劲的感觉。
昨天晚上实在忍不了了，洗了一下头。清爽了很多
发现个好方法，实在不行就用淋浴喷嘴喷脸，用纸巾吸吸眼睛，然后再用吹风机热风吹干。也不错。
#life
今天第一次用github action编译了libwrt。
当然用了zqking大佬的repo，github action都很完善了。
我就是 fork了一下，github action文件都不用动，只需要修改一下deconfig里面ipq60xx libwrt的编译选项。
在里面加入了zram，调整了一些原来自带的包。
zqking使用了一个第三方的包裹小仓库，主要是那个仓库里面能找到的，都可以在里面写上。
完事儿之后，可以通过github的对比功能看看与原repo的区别。检查一下修改之处。
—-
更：
经过了几天的研究，有了新的认识。
编译最终配置就是 deconfig 里的config 文件，里面可选的第三方包，得看 update.sh 里面拉取 feed 中包含的包。比如有一个 install_small8的函数，这个函数里面就是把这个第三方源的包配置安装到 makeconfig 里面。 （但这并不代表最后纳入到编译范围内，只不过是提供条件）
—-
action yaml 里面的变量 GITHUB_TOKEN 是一个自动生成的token，他的权限在 repo 的 action 一栏里面赋予。 不需要自己建这个 secret。
#programming
当然用了zqking大佬的repo，github action都很完善了。
我就是 fork了一下，github action文件都不用动，只需要修改一下deconfig里面ipq60xx libwrt的编译选项。
在里面加入了zram，调整了一些原来自带的包。
zqking使用了一个第三方的包裹小仓库，主要是那个仓库里面能找到的，都可以在里面写上。
完事儿之后，可以通过github的对比功能看看与原repo的区别。检查一下修改之处。
—-
更：
经过了几天的研究，有了新的认识。
编译最终配置就是 deconfig 里的config 文件，里面可选的第三方包，得看 update.sh 里面拉取 feed 中包含的包。比如有一个 install_small8的函数，这个函数里面就是把这个第三方源的包配置安装到 makeconfig 里面。 （但这并不代表最后纳入到编译范围内，只不过是提供条件）
—-
action yaml 里面的变量 GITHUB_TOKEN 是一个自动生成的token，他的权限在 repo 的 action 一栏里面赋予。 不需要自己建这个 secret。
#programming
AX6600 upgrade procedures
1. export current config
2. flash, unplug eth cable
3. overwrite network conf and
4. check if mac addr is changed
5. if ok, plug eth cable and opkg update
6. install openclash (which requires some deps)
7. export current conf
8. compare current conf with the old one
9. overwrite necessary config files to the new config
*If some configs are never used but changed, exclude them.*
10. overwrite the modded new configs to the emmc
* Modified configs mostly lie in:
* Outside of 'config' folder:
* Don't forget the network watchdog script in /root
* Disable packages in 启动项 instead of uninstalling
#ax6600
1. export current config
2. flash, unplug eth cable
3. overwrite network conf and
service network restart 4. check if mac addr is changed
5. if ok, plug eth cable and opkg update
6. install openclash (which requires some deps)
7. export current conf
8. compare current conf with the old one
9. overwrite necessary config files to the new config
*If some configs are never used but changed, exclude them.*
10. overwrite the modded new configs to the emmc
[use cp -r ] !!! * Modified configs mostly lie in:
athena-led, dhcp, dropbear, firewall, lucky, network, openclash, system, wireless, crontabs * Outside of 'config' folder:
crontabs, openclash, lucky * Don't forget the network watchdog script in /root
* Disable packages in 启动项 instead of uninstalling
#ax6600
