Meta Astryx, Modern Web Guidance and 3 More Trending GitHub Repos
Prefer to watch? Watch on YouTube · Subscribe for new builds · Read the transcript
In this video I count down five repos blowing up on GitHub Trending this week: Plannotator, Anthropic's skills, Alibaba's Open Code Review, Chrome's Modern Web Guidance and Meta's Astryx. I installed every one in a sandbox, with no Claude login and no API keys. If you build with Claude Code or Codex and want to know which ones are worth installing and what the catch is on each, this is for you.
What I cover in this video
- 0:00 The five repos and why every one has a catch
- 0:45 Plannotator: marking up a Claude Code plan in your browser
- 3:02 Anthropic's Office skills move to Apache 2.0
- 5:18 Alibaba's Open Code Review and how it works
- 6:50 The rest of Alibaba's benchmark chart
- 8:23 Chrome's Modern Web Guidance and the 52% to 92% claim
- 10:40 Meta's Astryx and why it's trending now
- 12:13 Astryx's beta caveats, then which of the five I'd install
Plannotator and Anthropic's skills
Plannotator: mark up a plan in your browser
Plannotator has 13,162 stars, 2,926 of them this week, and it's built almost entirely by one developer, Michael Ramos in New York. In Claude Code's plan mode, your choices are pretty much yes, no, or typing a paragraph about the step you don't like. Plannotator adds a hook that fires when Claude tries to leave plan mode and opens the plan in your browser. You highlight a line, comment on it, delete it or label it, and your notes go back to the agent with line numbers.
I installed it with the minimal flag and scripted the browser clicks, without my real Claude Code login, so this was the review screen and not a live session.
The catch: the default install is a curl pipe to bash that wires hooks into every coding agent it finds, so use the minimal flag and add only the Claude Code plugin yourself. It also sends one anonymous ping a day for each of its tools you use, which NO_ANON_PING=1 turns off.
Anthropic's Office skills are now Apache 2.0
Anthropic's skills repo has 196,854 stars, 4,847 this week. Until November 6, the docx, pdf, pptx and xlsx skills came with a license that said "All rights reserved." Then one commit moved all four to Apache 2.0. A company can now fork these skills, change them, build them into a product and even sell it, as long as the license and notices stay.
I installed only the PDF skill and ran its scripts on the real IRS W-9. They found 23 fields. I didn't give it my login, so I typed five sample values by hand, and the scripts filled the form and rendered all six pages.
The catch: the doc-coauthoring folder has no license file, so don't assume the whole repo is Apache. Skills also run code with your agent's permissions, so read them first.
Open Code Review: read the rest of the chart
Open Code Review has 53,894 stars, 3,675 this week. It started as Alibaba's internal code reviewer and is now a command line tool that reads your git diff and leaves comments line by line.
On Alibaba's own benchmark with Claude 4.6 Opus, it scored an F1 of 25.1% against 11.6% for Claude Code. But Claude Code found more of the real issues, 28.9% versus 20%. Alibaba's tool wins that score because it cries wolf a lot less. This is Alibaba's own test on its own data set, with older versions of both.
In my sandbox it refused to run without a key: "No valid LLM endpoint configured." So I ran the prep step of delegation mode, where your own Claude Code or Codex does the reviewing. I didn't attach an agent, so there were no AI comments, but it found both changed files and handed over its security rules. The catch: native mode runs on your API bill, and it will reuse Anthropic settings already in your environment or shell profile.
Modern Web Guidance and Astryx
Modern Web Guidance: a skill for current web code
Modern Web Guidance from the Google Chrome team has 8,531 stars, and 7,015 came this week. Version 1.0 came out November 10. It installs as a skill from the GoogleChrome/modern-web-guidance repo and tells your agent to run it first on any front-end job, pull the matching guide and check its code against it. Google's own dashboard says agents passed 52% of its best-practice checks without it and 92% with it, across 1,114 tasks on Claude Code, Codex and Antigravity. That's Google grading Google.
With telemetry off, I pointed its legacy scan at an old Bootstrap admin template and it found 12 things, including nine progress bars built out of divs. The catch: telemetry is on by default and sends your agent's name, the guides it pulls and its search queries. DISABLE_TELEMETRY=1 turns that off. Running it first also means more tool calls. In one of Google's example runs, tokens went from 82,000 to 112,000.
Astryx: Meta's design system for agent-built UI
Astryx has 30,270 stars and 8,700 this week. Meta says it grew inside the company for 8 years and powers more than 13,000 apps. What's on GitHub is the open source rebuild for React 19, in public beta since June. It's trending now because it got a Claude Code and Codex plugin on November 10, a San Francisco meetup on the 12th, and landed inside v0 on the 14th.
I started a fresh Vite app, ran init, added a theme and copied the dashboard template in. Then one command each switched the same app between the neutral, Y2K and Gothic themes. The catch: it's still beta with over 700 open pull requests, it needs React 19 or newer, and init edits your agent files.
Which ones I'd actually install
- Modern Web Guidance: on any web project today, with telemetry turned off.
- Plannotator: the minimal install plus its Claude Code plugin, if you use plan mode.
- Astryx: on a side project, because it's still beta.
- Anthropic's Office skills: bookmark them, and fork them if you're a company.
- Open Code Review: only if your team has a budget and a pile of pull requests.
Key takeaways
- Every one of these has a catch, and almost all were in the README, the license or the docs. Read those first.
- Two of the five send data by default. Use
NO_ANON_PING=1for Plannotator andDISABLE_TELEMETRY=1for Modern Web Guidance. - The benchmark numbers from Alibaba and Google are each company grading its own tool.
FAQ
What is Meta's Astryx?
Astryx is Meta's design system, released under the MIT license as an open source rebuild for React 19. Meta says it grew inside the company for 8 years and powers more than 13,000 apps. It has been in public beta since June and needs React 19 or newer.
Are Anthropic's Claude skills open source now?
On November 6, the docx, pdf, pptx and xlsx skills moved from All rights reserved to Apache 2.0 in one commit, so you can fork, change and sell them if you keep the license and notices. The doc-coauthoring folder has no license file, so don't assume the whole repo is Apache.
Is Alibaba's Open Code Review better than Claude Code at code review?
On Alibaba's own benchmark with Claude 4.6 Opus, Open Code Review scored an F1 of 25.1% against 11.6% for Claude Code. Claude Code found more of the real issues, 28.9% versus 20%, but flagged far more false ones. It is Alibaba's own test on its own data set with older versions of both.
How do I turn off telemetry in Modern Web Guidance and Plannotator?
Modern Web Guidance has telemetry on by default, and setting DISABLE_TELEMETRY=1 turns it off. Plannotator sends one anonymous ping a day for each of its tools you use, and NO_ANON_PING=1 turns that off.
Chapters
Transcript
Read the full transcript
0:00Meta's design system just pulled in 8,700 GitHub stars in one week. Google says its new skill took agents from 52 to 92% on its own checks. And earlier this month, Anthropic moved Claude's office skills to open source. But the stars aren't the story. Four come from big companies. One is almost all one developer in New York. And all five have a catch. So, in this video, I'm counting down five repos blowing up on GitHub trending this week, from an indie developer plan reviewer up to meta and which ones I'd actually install. I ran everyone in a sandbox first. Let's jump right into it. So, number five is Planotator. With over 13,000 stars and
0:45over 2900 of them this week, it's built almost entirely by one developer, Michael Ramos in New York, on the same trending page as Meta and Google. You put Claude code in plan mode. It writes you a plan and your choices are pretty much yes, no, or typing out a whole paragraph about which step you don't like. Planotator turns that plan into something you mark up like a Google doc. So, how does it actually work? Step one, the installer downloads one program and checks its fingerprint. Step two, it adds a hook to Claude code that fires right when Claude tries to leave plan mode. Step three, it starts a little server on your own machine and opens the plan in your browser. Step four, you highlight any line and comment on it, delete it or label it. And step five,
1:30you hit send feedback and your notes go back to the agent as its next instruction, line numbers and all. So, I installed it in the sandbox with the minimal flag, which puts just that one program on your machine. I gave it a madeup dark mode plan where step four touches about 140 files and I scripted the browser clicks so you can see each one highlight step four comment too much for one PR do the header and buttons first send and there it is in the terminal line 10 the way the agent would get it I didn't hook it into my real claude code login so this is the review screen not a live claude session now the fine print the default install is a curlpipe to bash that wires hooks into every coding agent it finds. So use the minimal flag and add just the claude
2:16code plugin yourself. It sends one anonymous ping a day for each of its tools you use, which no anon ping equals 1 turns off. And it ships 10 releases in 8 days, so things move under your feet. So if you live in plan mode, this one's worth a look. But number four comes from the company that makes Claude, and it's a license change most people scrolled right past. $99 a month. That's every course in shipping school with a new one every day. Two or three live calls a week, a Q&A with me every Saturday, and now the app that makes your videos. Nightcrew Studio is built from the machine that runs my channels. And in the last 7 days, that machine got me 711,000 views, every look, unlimited videos on
3:02your Mac, on your own Claude plan. Join free, upgrade inside, and your key lands in your inbox. So number four is Anthropic's skills repo with over 196,000 stars and over 4,800 this week. This is where the skills behind Claude making Word docs, PDFs, PowerPoints, and Excel files live. And up until November 6th, those four office skills came with a license that said all rights reserved. You couldn't make derivative works, and you couldn't hand them to anybody else. Anthropic's own readme called them source available, not open- source. Then in one commit, an anthropic engineer moved all four to Apache 2.0 and the readme now says these are also open- source. I couldn't find a blog post or
3:48an announcement anywhere. It was just the commit. So what does that mean for you? Apache means a company can fork these skills, change them, build them into a product, and even sell it as long as you keep the license and the notices. Earlier this month, you were only allowed to read them. So what's inside a skill? I use the PDF one because the Word, Excel, and PowerPoint scripts need Libre Office. Step one, a skill is just a folder in your Claude skills folder. Step two, Claude only loads it when a PDF job shows up. Step three, the instructions say in capital letters, you must complete these steps in order. Step four, scripts check for fillable fields, pull out every field, and turn the pages into images Claude can read. Step five,
4:33Claude decides which value goes in which box. And step six, one last script checks Claude's field names and values, then fills in the form. So, I installed just the PDF skill into one project in the sandbox and ran its scripts on the real IRSW9. It found 23 fields. Claude normally writes the values, but I didn't give it my login, so I typed in five by hand. Jane example@amplest studio, and the scripts filled the form and rendered all six pages. The license has one gap though. One folder in this repo, doc co-authoring, has no license file at all, so don't assume the whole repo is Apache. Skills also run code with your agents permissions. So read them first. And if you only use Claude in the app,
5:18you already have these built in. For anyone building their own skills, this is the best reference out there. And speaking of Claude, the next company says its tool scores better than Claude Code on its own benchmark. So, I read the rest of their chart. So, number three is Open Code Review from Alibaba with over 53,000 stars and over 3600 this week. It started as Alibaba's internal code reviewer. And Alibaba says it served tens of thousands of their developers. Now, it's a command line tool that reads your git diff and leaves comments line by line. So, how does it decide what to look at? Step one, plain code picks which files get reviewed, not the model. Step two, it bundles related files and every bundle gets its own sub
6:04agent. Step three, it matches rules to each file type. For JavaScript, using var is strictly prohibited, which fair enough. Step four, the AI reads and searches with tools that can only review, never edit. And step five, it rechecks every comment against the actual diff before you see it. And then there's their chart on Alibaba's own benchmark with the same model, Claude 4.6 6 Opus Open Code review scored an F1 of about 25% and Claude Code scored 11 12 on average a minute 23 per poll request versus 13 minutes about 385,000 tokens versus 5.7 million. So it beats Claude Code, right? Well, look at the rest of that chart. Claude Code actually
6:50found more of the real bugs about 29% versus 20. Alibaba's tool wins that score because it cries wolf a lot less. Claude Code flagged almost 6,000 issues and only 435 were real. And this is Alibaba's own test on its own data set with older versions of both. Even the winner catches about 1 in five known issues. So take that for what you will. So I installed it in the sandbox and gave it a branch with a SQL injection and a cross-sight scripting bug planted in it. And it refused to run. No valid LLM endpoint configured. That's on purpose because I gave it no key. So I ran delegation mode's prep step where your own claude code or codeex would do the reviewing. I didn't attach an agent. So there are no AI comments here. But it
7:36found both changed files and handed over the security rules it would give the agent. Cross- site scripting inner HTML eval. No SQL rule in that list. By the way, native mode runs on your own API bill. And if Claude Codes anthropic settings are already in your environment or your shell profile, it will reuse them, and that's your bill, too. Delegation skips that setup, but it still uses your agents own usage. So, if your team is buried in pull requests, test it. If you're solo and happy with how Claude Code reviews your code, I'd skip it. And Alibaba's tool catches mistakes after you write them. Number two stops your agent from writing old code in the first place. So number two is modern web guidance from the Google Chrome team with over 8,500 stars and
8:23about 7,000 of those came this week. So most of this repo's stars came in the last 7 days. If you've ever had an agent build you a web page, you've seen this. It writes code like it's 2019. jQuery, a progress bar made out of divs, a modal built from scratch. The show HN title said it straight. Get your coding agent off 2019 JavaScript. That post hit 448 points. Version 1.0 came out November 10th and Firefox and Edge are joining Chrome on it. So, how does it work? Step one, you install it as a skill and its description tells your agent mandatory execute first on any front-end job. Step two, the agent searches it with a few words like animate a dialogue. Step three, it pulls the matching guide with
9:08browser support data. Step four, the agent writes the code, checks it against the guide, and has to say which rule it used. And step five, a legacy scan goes through an existing codebase looking for old patterns. Google's own dashboard says agents passed 52% of its best practice checks without it and 92% with it across 1,114 tasks on Claude Code, Codeex, and anti-gravity. That's Google grading Google, so take it with a grain of salt, but that's a 40point jump. So, in the sandbox, I ran it with the telemetry turned off. I searched animate a dialogue modal backdrop, and it came back with 10 guides. Then, I pointed the legacy scan at a real old Bootstrap admin template, and it found 12 things. Password fields, missing autocomplete,
9:54and nine progress bars built out of divs it says could be the native progress element. Two things before you install it. Telemetry is on by default, and it sends your agents name, the guides it pulls, and the search queries your agent writes. Disable telemetry equals 1 turns that off. And execute first means more tool calls. In one of Google's own example runs, tokens went from 82,000 to 112,000. So if you ship frontends with an agent, I'd install this one today. Which brings us to number one. The most stars on the whole page this week from a company that says it's used this thing inside for 8 years. So, number one is Asterisk from Meta with over 30,000 stars and 8,700 this week, the most on
10:40GitHub's weekly trending page. Meta says this design system grew inside the company for 8 years and powers more than 13,000 apps. What's on GitHub is the open- source rebuild made for React 19. Now, people on X were posting it like Meta just dropped it. It didn't. It went public in beta back in June. So, why now? On November 10th, it got a Claude code and Codeex plugin. On the 12th, a meetup in San Francisco. On the 14th, it landed inside Vzero. And then the Exposts did the rest. And you know the AI slop look. Every app Claude makes looks like every other app Claude makes. Asterisk is built so your agent builds UI from a real design system instead of guessing. So what does it actually hand your agent? Step one, you install the
11:27core and a theme with no build plugin. Step two, asterisk init writes instructions into your agents files. And in my run, that was agents.mmd and gemini.md. Step three, instead of guessing, the agent asks the asterisk command line tool or its MCP server for the right component and its docs. Step four, asterisk template dashboard copies a whole page into your app. And step five, the themes are just CSS variables. So one command restyles everything. So, in the sandbox, I started a fresh VIT app, installed it, ran in it, added a theme, copied the dashboard template, installed the two packages it asked for, and wired up one file by hand. And there it is, running a whole dashboard with the template sample data. Then, one
12:13command each, neutral, Y2K, and Gothic. Same app, three totally different looks. I'm not a designer, but that looks better than anything I would have made. It's still beta though with five releases this week and over 700 open pull requests. It needs React 19 or newer. So if you're on Vue or Spelt, this isn't for you. In it edits your agent files. And in my barebones setup, Gothic needed dark mode turned on or it was pretty much unreadable. So if you're a React builder letting Claude or Codeex make your UI, this is the most fun one on the list. So which ones would I actually install? Modern web guidance goes on any web project today with the telemetry turned off. Planotator with the minimal install plus just its claude code plugin if you use plan mode.
12:59Asterisk on a side project because it's still beta. Anthropic office skills bookmark them and if you're a company fork them now that you're allowed to. And open code review only if your team has a budget and a pile of pull requests. And like I said at the start, the stars aren't the story. Every one of these had a catch. And almost all of them were sitting right there in the readme, the license or the docs. So read those before you install anything. I'll put the links to all five down below. So which one are you letting into your repo first, the indie developers or one of the big companies? Tell me down in the comments. And if you want part six next Tuesday, a like is how I know. Have a blessed day.
Ready to start building with AI?
Join the Shipping Skool community free. Paid members get live bootcamp calls every week and weekly-updated courses.
Join Shipping SkoolKeep reading
Claude Code GitHub Repos: Manim, Remotion, Blender MCP and More
Five free repos take Claude Code past writing code. Manim and Remotion Skills make videos, Lightpanda is a browser built for agents, MCP for Blender builds 3D scenes, and agent-desktop lets an agent use your whole computer. I'd install MCP for Blender first, and I'd only run agent-desktop in a VM or a separate user account.
Trending GitHub Repos for Claude Code: T3 Code, claude-mem, ECC
The five repos that gained the most GitHub stars last week were T3 Code, claude-mem, ECC, Superpowers and TradingAgents. I'd install Superpowers first, try T3 Code if you pay for two or more agents, take only a few pieces of ECC, hold off on claude-mem, and read TradingAgents' code without trading with it.
Cloudflare Security Audit Skill, Graphify and 3 More Trending Repos
The five repos are Cloudflare's security-audit-skill (+15,072 stars), Graphify, ComfyUI, AI Memory and Ruflo (formerly Claude Flow), ranked by stars gained last week. I'd start with the security skill in quick mode on one folder, and add Graphify if your project is big.
