How many "Must-Install Skill" recommendations have you scrolled through?
You install a few, but when you open Claude, you still don't know which one to use.
It feels the same whether you have them installed or not.
Actually, the problem isn't the skills themselves.
Six skills without a division of labor are no different from six strangers stuffed into the same office; no one knows what they are responsible for, and more people just make it messier.
The root cause is simple: you have tools, but no workflow framework.
Your daily work consists of 4 steps: Collection → Digestion → Production → Reporting.
Think about it, isn't that the case?
What do you do when you open your computer every day?
You search for things, find materials, browse posts, and see what others are saying—that's Collection.
You gather a bunch of stuff, then you have to read it, organize it, and turn it into something you can use—that's Digestion.
Then you write something: an article, a proposal, a report—that's Production.
Once written, you send it out: post it to a public account, hand it to your boss, or drop it in a group—that's Reporting.
Collection, Digestion, Production, Reporting. You rotate through these 4 steps every day, right?
So today, we are talking about 4 skills to solve these problems plus 2 maintenance tools.
What you will learn today is a workflow framework, not just another trashy recommendation post.
Step 1: Collection
In the collection step, when you ask AI to search for things, it only finds public web pages.
But where are the discussions you actually want to see? Xiaohongshu notes, WeChat Official Account articles, Bilibili comment sections—they are all in places that require a login to view.
AI can't get in because it doesn't have your login status.
This isn't a prompt issue; it's a hard limitation of all AI tools.
The Web Access skill does one thing: it connects directly to your local Chrome via CDP.
Address: https://github.com/eze-is/web-access
Whatever you are logged into in Chrome, it can access.
Install with one command:
1npx skills add eze-is/web-access
One configuration step: Open chrome://inspect/#remote-debugging in the Chrome address bar and check "Discover network targets."
Configuration complete.
You give it an address to scrape a WeChat article, and it works perfectly, as shown below:
As you can see, it really gets in, retrieves the full text, and isn't blocked by anti-scraping measures.
It has more scraping methods, and after each successful trial and error, it remembers, so it won't make the same mistake next time.
Anywhere you can normally log in, it can enter for you. The collection step no longer requires manual browsing.
Step 2: Digestion
Collection is solved, but you've gathered a bunch of stuff—now what?
Normally, we have to organize it, take notes, and record useful things, or we'll forget in a couple of days.
But you know how it goes: you start taking notes, and then you stop.
To learn better, some people get more serious:
They build knowledge bases or "Second Brains" using Notion or Obsidian. They create a bunch of folders, organize for two or three months, and then stop opening them.
Whether it's note-taking or a second brain, they all fail for the same reason: the maintenance cost falls entirely on you, and it breaks as soon as you get busy.
But Karpathy's method is the opposite: you just throw the material in, and Claude is responsible for reading, summarizing, creating pages, and building bidirectional links.
The maintenance cost shifts from your time to a single command.
The knowledge base grows on its own instead of being maintained by you. Every entry is an accumulation, not a debt.
I wrote a full tutorial in this article: "Obsidian + Claude Code: Building a Truly Usable Second Brain Using AI Master Karpathy's Method"
https://x.com/lxfater/status/2042848343949480173
For those who want to use a ready-made Skill directly, the community already has three versions:
- AgriciDaniel/claude-obsidian, with /wiki /save /autoresearch Address: https://github.com/AgriciDaniel/claude-obsidian
- ekadetov/llm-wiki, a Claude Code plugin Address: https://github.com/ekadetov/llm-wiki
- vanillaflava/llm-wiki-claude-skills, GUI installation, no terminal needed Address: https://github.com/vanillaflava/llm-wiki-claude-skills
The digestion step no longer relies on your own brute-force understanding of input.
Step 3: Production
Now for production. You used Claude to write something, and the content is fine.
But do you dare to post it directly?
Nowadays, everyone has chatted with AI, and people know what AI-written content looks like.
It constantly uses phrases like "it is worth noting," ends with a mandatory "sublimation" paragraph, and every sentence is perfectly structured. Readers recognize these features instantly.
When you post it, the reader's first reaction is: "This was written by AI."
Your content and professionalism are ruined by the "AI flavor."
Here are two Skills:
humanize (https://github.com/blader/humanizer
Chinese version: (https://github.com/op7418/Humanizer-zh, made by a fellow user)
This command does something very specific: it scans a list of AI writing characteristics compiled from Wikipedia and clears them one by one.
This isn't a general polishing Skill; it's a Skill specifically targeting AI writing fingerprints.
Example image:
Before:
It is worth noting that this tool can not only significantly improve work efficiency but, more importantly, it fundamentally changes the way we work.
After:
Why is this tool so good? Work is faster, and the method has changed.
After the change, readers will think you wrote it.
This is what truly usable content looks like!!
Step 4: Reporting
Finally, reporting.
The content is written, but writing it doesn't mean it's ready to send.
You still have two hurdles.
The first is format. You wrote in Markdown, but others want Word, reports need PPT, and public accounts need layout. Each is a separate manual chore.
The solution is the four Skills released officially by Anthropic. Install them with four commands:
1npx skills add https://github.com/anthropics/skills --skill docx2npx skills add https://github.com/anthropics/skills --skill pptx3npx skills add https://github.com/anthropics/skills --skill xlsx4npx skills add https://github.com/anthropics/skills --skill pdf
Once installed, just tell Claude what format you want. It outputs professional documents with formatting that the recipient can use immediately.
The second is sending. Even with the format ready, you still have to manually push it to where it needs to go.
md2wechat is a one-command solution that automatically formats and pushes to your WeChat Official Account drafts. Just go in, change the title and cover, and post.
Installation steps:
1brew install geekjourneyx/tap/md2wechat2npx skills add https://github.com/geekjourneyx/md2wechat-skill --skill md2wechat
I have written a full tutorial on this before:
https://x.com/lxfater/status/2037047059384328315
It takes less than 5 minutes from writing to posting. The "last mile" has disappeared.
Two More Ops Tools
Collection, Digestion, Production, and Reporting—all four steps are in place.
But any system needs maintenance.
When problems arise, you need to find the root cause, not guess. When things run smoothly, you need to solidify good processes, not start over every time.
This system is equipped with two maintenance tools:
The first solves the issue of AI going in circles when things go wrong, leaving you clueless about where it's stuck.
Superpowers (https://github.com/obra/superpowers) is the most popular Skill suite in the Claude ecosystem, with 150k Stars on GitHub and 350k installations.
The /systematic-debug command inside forces a sequence of isolation, observation, diagnosis, repair, and verification. It prohibits proposing any repair plan until the root cause is found. It relies on process, not luck, to find the real reason.
How to install?
1npx skills add obra/superpowers
The second solves the problem of producing Skills.
These 6 Skills cover the main stages, but your work surely involves more.
You have your own processes and habits, and there are no ready-made Skills to do those for you.
Use the Skill below 👇
Skill-Creator (https://github.com/anthropics/skills, skills/skill-creator/) is an official Skill from Anthropic.
How to install?
1npx skills add https://github.com/anthropics/skills --skill skill-creator
Enter /skill-creator, describe your process, and it helps you generate a complete SKILL.md.
Your own way of working can also become a Skill!!
This system goes from being something you use to something that becomes more "you" the more you use it. With these two added, it is truly complete.
I admit the article is a bit long, so here is the solution for when you can't remember it all!!
Afraid you won't remember?
6 Skills, arranged by Collection, Digestion, Production, and Reporting, with problem-fixing and Skill generation as the foundation. This is your workflow framework.
Save this image:
Start installing from the first one, one at a time. Try it out after installing before moving to the next.
Find this useful? Forward it to a friend you think should try it, and follow me!!
