10 Trending AI Coding Repos: Open Code Review, God's Eye View, and WeKnora
Alibaba ships an automated code reviewer, a hobbyist wires public cameras and flight data into one live map, and a comment-section fight over AI 'skills' files splits 3-to-1 in their favor.
Posted
2 days ago
Duration
Format
Listicle
educational
Views
31.2K
420 likes
57 · 43
Big Idea
The argument in one line.
AI code review, disposable context sandboxes, and one-command git worktrees are becoming standard safety nets for developers who ship code they don't fully read line by line.
Who This Is For
Read if. Skip if.
READ IF YOU ARE…
You use Claude Code, Cursor, or Codex regularly and want to know which new tools other builders are actually adopting this week.
You're a solo or small-team developer looking for guardrails, like automated code review or test-driven skill packs, because you don't read every line an agent writes.
You run more than one AI agent against the same codebase and keep hitting merge conflicts between them.
You're on the fence about whether AI coding 'skills' files are worth the setup time.
SKIP IF…
You want a step-by-step tutorial. This is a rapid-fire roundup with links, not a deep walkthrough of any single tool.
You aren't using AI coding agents yet. Half of what's covered here (skill packs, worktree managers, context sandboxes) only makes sense once agents are already part of your workflow.
TL;DR
The full version, fast.
This week's GitHub countdown covers ten trending repos plus five audience submissions, and most of them solve the same underlying problem: people are shipping AI-written code they don't fully read. Alibaba's Open Code Review scans for security holes like plain-text passwords, ECC (261k stars) bolts a full spec-plan-code-test-review-ship workflow onto Claude Code, and Worktrunk lets several agents each work in their own git worktree without overwriting each other's files. A running debate over whether AI 'skills' files actually help splits the comments about 3-to-1 in favor of keeping them. Tencent's WeKnora and Context Mode both attack the same problem, agents losing track of files and context, from different angles.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Quick teaser cuts through the week's repos before the countdown starts.
00:28 – 02:07
02 · Open Code Review by Alibaba
An AI code scanner that caught a deliberately hidden plain-text password and can generate a one-command HTML report; can be wired into a Git pre-commit hook.
02:07 – 03:27
03 · I Have ADHD
A skill file (47.4k stars) that forces an AI coding agent into short, scannable, numbered-step responses instead of verbose prose.
03:27 – 05:05
04 · God's Eye View
An open-source map that layers public traffic data, CCTV cameras, and flight tracking into one interactive view; a hosted version exists for people who don't want to self-host.
05:05 – 06:26
05 · Monid AI (sponsor)
Sponsor segment: one API key and one bill for 1,800+ tools an AI agent can call, positioned as an OpenRouter for tools instead of models.
06:26 – 08:59
06 · ECC
A 261k-star engineering-workflow toolkit that enforces test-driven development on Claude Code: write the test first, then the feature.
08:59 – 10:17
07 · Should We Keep AI Skills?
Audience feedback recap: 32 commenters want to keep skills in the roundup, 9 want them dropped.
10:17 – 11:00
08 · Why Skills Matter for Less Capable Models
Skill packs add the most value to weaker or open-weight models; top-tier models often already know the right process.
11:00 – 13:08
09 · Context Mode
A 23,000-star tool that runs exploratory tool calls in a disposable side sandbox and returns only the final answer to the main session, protecting the context window.
13:08 – 14:48
10 · Tencent WeKnora
A 26.2k-star tool that auto-builds a searchable, browsable knowledge graph out of a folder of files, linking people, meetings, and projects.
14:48 – 15:08
11 · Zapier SDK & MCP (sponsor)
Sponsor segment: Zapier SDK connects a built product to users' apps, Zapier MCP gives an agent one connection into thousands of tools.
15:08 – 16:15
12 · OpenAI Plugins
A 62-plugin repo with install instructions and skill files but no code; Claude can convert or connect to a Codex-built plugin.
16:15 – 17:00
13 · Plugin Security
Plugins run their own scripts with access to files and secrets, a bigger trust decision than installing a skill file.
17:00 – 19:03
14 · Worktrunk & Git Worktrees
A CLI (7.9k stars) that turns git worktrees into a one-command way to give each of several AI agents its own copy of a project.
19:03 – 21:04
15 · Agent Skills
Addy Osmani's 95,000-star skill collection enforces a spec, plan, code, test, review, ship workflow instead of jumping straight to 'build it.'
21:04 – 22:25
16 · Claude Code on GitHub
Anthropic published an issues-only repo for Claude Code, no source code, just a place to file bugs, since Claude Code isn't open source.
22:25 – 22:49
17 · Audience Submissions
The show moves into 'First Stars,' repos submitted by viewers.
22:49 – 23:32
18 · Logic Loop
A project-centric workspace for switching between Claude, Codex, and Cursor without keeping each tool in its own window.
23:32 – 24:05
19 · Claude Dashboard
A local board that shows every Claude Code session on one screen and flags which one is stuck waiting on you.
24:05 – 24:43
20 · Deck Gauge
A task tracker pulled out of Jira and GitHub, built to show task status changes only when work is actually done.
24:43 – 25:31
21 · Campus AIOS
An 'AI department' framework for running a one-person or zero-person company, similar in spirit to Pulsia.
25:31 – 26:45
22 · Reckoner
A dashboard tracking remaining credit across OpenRouter, Anthropic, xAI, Railway, RunPod, and Neon on one screen, plus wrap-up.
Atomic Insights
Lines worth screenshotting.
Alibaba's Open Code Review found problems a developer deliberately hid in his own code, including a password left in plain text, and generated a full HTML report from a single command.
A Git pre-commit hook can force a security scan to run automatically before every commit, so the check happens whether or not the developer remembers to run it by hand.
The 'I Have ADHD' skill file (47.4k stars) forces an AI coding agent's replies into numbered steps with no filler, built for making increasingly verbose models like Claude readable at a glance.
God's Eye View pulls public traffic data, CCTV camera feeds, and flight tracking into one interactive map, and its creator ships a hosted version so viewers don't have to self-host it.
ECC (261k stars) enforces test-driven development on AI coding agents: it writes a test case, like '2 plus 2 should equal 4,' before letting the agent write the feature code.
In an audience poll about coding 'skills' files, 32 commenters wanted to keep them and 9 wanted them gone, with the pro-skills group giving detailed, specific reasons for how they use them.
Skill packs matter far more for smaller or open-weight local models than for top-tier models, because strong models often already know the right development process without being told.
Context Mode runs tool calls and file searches in a disposable side sandbox, then hands only the final answer back to the main coding session, so wasted searches never bloat the context window.
Tencent's WeKnora (26.2k stars) auto-builds a browsable knowledge graph from a folder of files, linking meeting notes to the people and projects mentioned in them like a personal Wikipedia.
OpenAI's plugins repo holds 62 plugin folders but no actual code, just install instructions and skill files. Claude Code can convert a Codex-built plugin to its own format or use a cross-platform connector.
Installing a plugin carries more risk than installing a skill: a plugin runs its own executable scripts with access to files and secret keys, while a skill only changes agent output.
Worktrunk turns Git worktrees, separate folders that each check out a different branch of the same repo, into a one-command tool so multiple agents can work on one project without fighting over the same files.
Addy Osmani's Agent Skills repo enforces a five-stage workflow (spec, plan, code, test, review, ship) because most people skip straight to 'build the login page' and miss everything else that has to change.
Anthropic posted an issues-only GitHub repo for Claude Code with no source code inside it, just a place to file bugs and track fixes, since Claude Code itself isn't open source.
One audience-submitted tool, Reckoner, tracks remaining spend across five LLM providers on one dashboard, because most people don't notice they're low on API credit until they run out mid-task.
Takeaway
The Tools Winning This Week Are Guardrails
WHAT TO WATCH
The GitHub repos climbing this week aren't flashy features, they're guardrails: automated code review, disposable context sandboxes, and multi-agent workflows built for people shipping code they didn't fully read.
02Open Code Review by Alibaba
If you're shipping code an AI agent wrote without reading every line, run an automated security scanner, like Alibaba's Open Code Review, before you deploy, since these tools catch things like plain-text passwords a quick visual check misses.
Wire the scan into a Git pre-commit hook so it runs automatically on every commit, rather than trusting yourself to remember to run it by hand.
03I Have ADHD
A single instruction file installed into your coding agent can force shorter, more scannable responses, numbered steps with no filler, which matters more as models get more verbose by default.
The same verbosity problem shows up outside chat replies too, in generated web copy and emails, so a tone-tightening instruction file is worth applying anywhere an agent writes text for you.
04God's Eye View
Public data, traffic feeds, CCTV cameras, flight tracking, is more available and more combinable than most people realize; one open-source project stitched it into a single browsable map.
If self-hosting a data project feels like too much setup, check whether the creator offers a hosted version first. It gets you the same result without the GitHub detour.
06ECC
Enforcing test-driven development on an AI agent, writing the test for '2 plus 2 equals 4' before the feature code, catches broken logic before it ships instead of after a user finds it.
A packaged set of engineering workflows can hand a solo builder decades of accumulated software-engineering practice they haven't personally earned yet. That's the actual value of a skill pack, not novelty.
07Should We Keep AI Skills?
Before deciding skill files aren't worth the setup time, know that opinions split hard: in one comment thread, roughly 3 in 4 people wanted to keep them and gave specific, detailed reasons why.
The people most skeptical of skill files tend to be the ones already running the most capable, expensive models, which produce good output without added instructions.
08Why Skills Matter for Less Capable Models
Skill packs deliver the most value on weaker or open-weight local models, since a strong model may already know the right process without being told, but a weaker one needs it spelled out.
If your agent's output already reads like a competent engineer, an added skill file may just be redundant weight. Judge it against the model you're actually running, not a blanket rule.
09Context Mode
Route expensive exploratory work, reading 15 files to find the 2 that matter, into a disposable side sandbox, and only pass the final answer back to your main session, so failed searches never eat your context window.
A shorter, cleaner main conversation isn't just tidier, it directly affects how much of your context budget stays available for the actual task instead of the side work of finding information.
10Tencent WeKnora
A tool that auto-links your files into a knowledge graph, this meeting relates to this project relates to this person, turns a folder nobody can search into something you browse like a wiki.
The automation only works because it runs retrieval-augmented generation against everything in the folder continuously. Expect this pattern to become standard for personal knowledge tools, not a one-off feature.
12OpenAI Plugins
A plugin repo can be useful as documentation alone. OpenAI's plugins repo holds no code, just install instructions and skill files describing how each plugin should behave inside a tool like Figma.
If you're on Claude and find a Codex-built plugin you want, expect Claude to either auto-convert it to its own format or route through a cross-platform connector. Check Claude's own plugin marketplace first for a version built for your platform.
13Plugin Security
A skill only changes what an agent says. A plugin can run its own scripts with access to your files and credentials, a materially bigger trust decision. Treat the two differently.
Install plugins directly from the platform that built them, OpenAI's own marketplace, Claude's own marketplace, rather than a random third-party source, since a malicious plugin's code executes with your access.
14Worktrunk & Git Worktrees
If you run more than one AI agent against the same codebase, give each one its own Git worktree, a separate checked-out folder for a different branch, instead of a shared branch, or they'll overwrite each other's edits.
A one-command wrapper around worktree creation, listing, merging, and deletion removes the main reason people avoid worktrees: the multi-step Git ceremony required to set one up by hand.
15Agent Skills
A five-stage workflow, spec, plan, code, test, review, ship, catches problems a straight 'build me a login page' request skips, because defining the full spec first surfaces the logout page and database table it also needs.
Testing and review are the two stages people skip most after building, and 'it worked on localhost' isn't proof it works in production. Validate the deployed version too, not just your local copy.
16Claude Code on GitHub
A company can publish an issues-only GitHub repo, no source code, just bug tracking, for a closed-source tool, purely because developers expect to file bugs on GitHub rather than a separate web form.
Don't assume a GitHub repo means open source. Check whether the code directory is actually populated before treating a repo's existence as license to fork or extend the tool yourself.
18Logic Loop
If you're juggling multiple AI coding tools across different projects, a project-centric workspace that lets you pick your agent per project beats keeping each tool in its own separate window.
19Claude Dashboard
Running several coding-agent sessions at once creates a real problem: knowing which one is idle waiting on you and which is still working. A single status board across all sessions solves it directly.
20Deck Gauge
Pulling tasks out of Jira or GitHub into your own tracker only makes sense if you need something those platforms don't do well, like moving a task's status only when it's actually done, not when someone drags it.
21Campus AIOS
The 'zero-person company' framing, you're the chairman of the board, AI agents run the departments, is becoming a recognizable product category, not a one-off experiment.
22Reckoner
If you're spread across multiple LLM providers, track remaining credit on all of them in one place. Running out of API credit mid-task is a common, avoidable failure most people only notice after it happens.
Glossary
Terms worth knowing.
Git worktree
A separate folder on disk that checks out a different branch of the same Git repository, letting you, or multiple AI agents, work on several branches of one project at the same time without switching branches back and forth.
Skill (AI coding)
A short instruction file installed into a coding agent, like Claude Code or Codex, that changes how it behaves, such as writing shorter responses or following a set process, without giving it new code-execution abilities.
Plugin (AI coding)
An installable extension for a coding agent that can include its own scripts and executable code, giving it broader access to files and credentials than a skill file, and carrying more security risk as a result.
Test-driven development (TDD)
A way of building software where you write the tests that define success, for example '2 plus 2 should equal 4,' before writing the code that makes them pass.
MCP (Model Context Protocol)
A standard connection method that lets an AI agent call outside tools and data sources through one consistent interface instead of a custom integration for every tool.
Context window
The limited amount of text, conversation history, files, and tool output, that an AI model can consider at once. Filling it with unnecessary file reads or failed tool calls leaves less room for the actual task.
“Did it make mistakes? Did some of the changes that you make open up doors that you didn't know were open?”
names the exact fear behind vibe coding in one breath→ TikTok hook↗ Tweet quote
07:50
“It's a way to get decades of experience from other software engineers and inject that into your coding sessions.”
reframes skill packs as borrowed expertise, not novelty→ IG reel cold open↗ Tweet quote
18:00
“Why not just give each of those agents its own little playground where it can have its own copy of the code, it can test it itself, it can edit, and then once the feature is done, you can merge all of them back together.”
plain-English explanation of worktrees anyone can follow→ newsletter pull-quote↗ Tweet quote
21:50
“It's like a normal open source repository minus the code.”
tight, quotable summary of a genuinely odd repo→ TikTok hook↗ Tweet quote
26:25
“You often don't find out about it until you run out of cash.”
punchy line about API spend blindness→ IG reel cold open↗ Tweet quote
The Script
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
17px
Alibaba is giving you a tester that will ensure everything you build with cloud code or other tools works great. There's a new tool that lets you spy on the world. I'm going to show you how pretty it is and how useful it can be.
Do you have a lot of files and you don't know what's where and where's what? Well, I've got a tool that will connect them all together, make them searchable, easy to find, and will work with your agents. All that and so much more.
Of course, we got links to everything and timestamps. Let's get to it. Presented by Zapier, the AI automation company.
Adam, the number one repo this week is OpenCodeReview from Alibaba. This is the thing that will let you check your code to find common or maybe even uncommon problems. I saw this one demo, which I'll link to up here, where someone said, I'm intentionally going to sneak stuff into deliberate problems into my code.
I want to see if it finds it. And sure enough, it did find those problems. Like he had a password that was just written in plain text.
And then he said, you know what? I want this to not give me just the red and green on my screen. I want to actually see a nice HTML report.
And all he did was fire off one command, and he got it. This makes sense, coming from Alibaba especially. Yeah, I mean, this is starting to be pretty standard practice where if you're out there, you're vibe coding stuff.
You're not reading all the lines of code. You don't exactly know how the integrations work. And you've asked Claude to build it in a secure way.
But did it? Did it make mistakes? Did some of the changes that you make open up doors that you didn't know were open?
So having something like this run and even better run automatically. is table stakes. I think everybody should be doing this.
Get this package. It's got a CLI. It's got some skills in there.
Super easy. You can even set these up so that there's this thing called get hooks where you could say every time I push to get or every time I commit, before I commit, run this other command first. And so now you know that something like this is watching.
It's got your back. And you can dismiss it if you want, but it's always going to run there and keep you a little bit safer. Yeah, this is the one that people are going to be giving their agents and using.
Let me know if you do. We're going to go on to number two. The second most popular.
Boy, this just keeps coming up, Adam. The second most popular repo is called I have ADHD. This is a simple file that all it does is it tells your coding agent to shut up and keep things really easy to follow.
The example is right here. I actually also found this video of a woman who decided what she wanted to do was use it in the in the. clawed web app and so she got her result and this is what it looked like before she invoked it.
Then she went in and she hit the slash command to invoke the skill. I wonder if people do invoke skills a lot from the web app. I don't.
And when she did, it got really short and it's more than short. It just makes so much more sense. It's scannable so you understand it.
I like this one and I could see why it continues to grow. Yeah, I mean especially like I've used Astra a lot.
And honestly, I'm switching back to Sol for a lot of my work because Astra is so verbose and it takes longer to run and it just generates more crap that I have to read through. And it's not just in these text responses. It's also the text that it writes for a web page or an email or whatever.
So I think these skills that help tune it down are getting more and more useful because the models are getting more and more verbose. 47 ,000 stars. Every week I feel like I email him and tell him that we've talked about him again, and we might even next week.
Next! This is God's eye view once again on the top 10 list. And what this does is it pulls together all the open data that's available online to make it available to you in a very, very, very, very, very, very pretty map that you can zoom in on.
I've got this video here for those of you who are downloading the full file. Actually, it's not downloading. There's a link to it.
And you can see he is seeing traffic data, which he. put in from TomTom, and then he has got all the cameras, the CCTV cameras on there. All he has to do is mouse over a camera and he gets to see a screenshot of it.
If he clicks into it, he gets to see live. And it's airplanes that you see overhead and so much data for it. It's kind of fun.
I could see this being useful. Yeah. I mean, it's just insane to watch this.
It feels unreal looking at it. I want to go and click around our city and see what I can see. One thing I will say about this is if you do not want to go through all this and set this up yourself on your system, he has got right here on Pinocchio.
You can go in and you can download it directly. And he created this guy here in the video. If you look for him, he has got his own version of it.
You'll see his face on it. That's the one that you can just click over and trust here. He's going to show it right now.
There it is. And so you can just go and install it and not have to go into GitHub if you'd rather have just a plain app. Okay, I'm very excited about this next thing.
This is a brand new sponsor. It is called Monid. And here's what they do.
You know how I'm actually going to click over to this. You know, you've got all these different tools, Adam, that you want your agent to have access to. Like, I give it Apollo.
I give it access to Reddit, scraping. I give it access to so many things beyond this. And then what you end up having is all these little bills from thousands or hundreds of different apps.
They connect thousands. What they do is they say, why don't you just connect to us? We'll bring all this over to you.
Make it easy for your agent to access all of them.
give you one bill without any subscriptions that's the beauty of what this is And since they paid me to read this, I'm going to read it exactly like they said it. Monit is the open router for agent tools.
Instead of a separate subscription for every API your agent touches, scraping, enrichment, scourge, social data, it's one key into 1 ,800 tools that you pay per call. Again, no subscription. Same layer open router built for models.
That's the one that Stripe just bought for $7 billion, except this is for tools that agents actually need. We've got a link below so you can go and sign up and start using it today. I like this one.
Yeah, it's awesome. I've got a lot of subscriptions that I would like to cancel and exchange for something that could just bill me per call. I would appreciate that.
Freaking Appify. That's the one for me. I had a couple of subscriptions and it's just...
Okay, ECC. I'm thinking about having this creator on. He and I have been talking a lot.
It's got 261 ,000 stars. It's a collection of skills. And we should talk about skills because people have big feelings about them.
And I'll show you what they are in a moment. But Adam, what exactly is this? When we were talking before, you said, Andrew, scroll down below to a section here to start explaining it.
What are we looking for here? What is this? So this is a set of workflows.
It's a lot of development software and engineering and development workflows. And each of those workflows has a process that is, Good.
We've built this up. We've learned over years and years the right way to develop. And not all the people that are out there building software today have that decade of experience or more in building software.
So something like this says, well, hey, when you want to build a new feature, this is the thing to start with. Type ECC plan, the described feature, and then you can run what's called TDD workflow. This is test -driven development.
What that means is that... This is going to, instead of just building the feature, which is what you would probably do by default, this is going to instruct Claude, hey, go and develop some test structures. Let's write the test first.
And you would say, well, the test for addition says if I was given 2 plus 2, it should equal 4. And then once you build a bunch of these tests out, you go and actually write the code. But by the time you're writing the code or your agent's writing the code, you already have all of these validation tests that can prove whether or not the thing works.
So it's a different way of thinking. And there's a lot of different ways of thinking that are just nuanced in a way where, again, a vibe coder is not necessarily going to learn all this stuff. It's a way to get decades of experience from other software engineers and inject that into your coding sessions.
By the way, I've got this video here. What I'm playing is him running the planning mode. And then this creator started responding to some of the planning mode questions.
And then before it built, he used the TDD. That's the acronym. And then.
It tested what he was supposed to get before it was fully built, made sure that it worked well, and then it gave him a Calendly link on his website. By the way, we've been saying that I've been showing these videos but not actually hitting play on them. There's a reason for it.
Let's first talk about last week. Last week, we asked people, do you want to keep the skills? Because every time there's skills, Adam feels like, oh, I don't know that I want them.
And here's the response from the comments. Nine people said kill the skills. I don't want them anymore.
Skip them. And 32 people said, yes, I want them. And they were super passionate about it.
I mean, if you take a look at this, not only is it like, yes, yes, yes, skills. I love it. Keep it.
And dude, Adam, all Saturday I was flipping through. They go, I don't know how to respond to all these because I like to respond to everybody, but I don't have like a line that I could give them. Some of them were really detailed about why they want it, how it definitely keep the skills, just explain in detail how to curate them to work best with them and so on.
I get that. Go back to that slide there. Number six.
This is actually a really important one. The one on the right from Black Logic. it helps less capable models become more capable if you're out there and you're using some of these local models or open weight models that aren't as good as astra or fable 5 .1 or whatever the skills actually are way more impactful there Because Fable's pretty good at all this stuff, right?
And that's like, yeah, it knows how to do test -driven development. You don't need a skill. But a lot of these other ones, you need to encourage it, and you might even need to explain how to do the development.
Now the skill packs make a lot more sense. I see. Okay, that makes a lot of sense.
I should also say there were people who very strongly felt that they didn't want the skills. And here are some of those. No to the skills.
Never adopt them. Skills are false. No fan of skills.
They seem overly specific to people. Just usually not my workflows. Okay.
There was so much heat over skills that what I ended up doing was creating a video this week with the top most installed skills, according to Vercel, which enables a lot of these skills to be installed. And if you want to watch it and go directly into the most installed, meaning actually use skills. you can see them there meanwhile this one is on our chart and we will for now keep the skills in the weekly roundup if they are uh trending and this one you can see is trending and continuing to grow let's look at number five which is context mode 23 000 stars today your ai coder forgets midway through the job this stretches the session we've talked about this before this is another tool that that how does it help us save tokens and what's the importance of memory and and then i'll show a video you know the word here is is probably sandbox if you scroll up just a little bit andrew into those you know top these four steps here so yeah stop right there how context mode solves it so so number one it's going to create this like kind of side sandbox where instead of your main session calling all the tools loading 15 files and figuring out that it only needs two it's going to do that in a little sandbox and then
Once that sandbox finds the answer to whatever the question was, it'll pass that data back to your main Claude session or Codex or whatever. And that... doesn't have to have gone through all of those files, all of the tool calls, failing once or twice, and then getting the right answer.
It's just delivered on a platter. Here's the answer you were looking for, and I've already vetted it. And that other conversation, that sandbox conversation can be thrown away.
It can be kept short. And so it makes the focus of your main conversation stay on the work instead of all the side work, like organizing and editing and getting context. All right, people have asked us to actually install this.
They said, okay, it's great that you're showing us all this information about how it's growing. We'd like to see and what it does. We'd like to see it used.
And I thought, using it will take a long time. I then clipped these videos where I thought I could take you directly to the part where somebody is using the tool and you could play it. And Adam and I went through it after I spent a couple of hours at least putting it together this week.
And we decided to even scrap playing the video because it does take a long time. It does feel tedious to hear all these different voices and get into each one of these. So I will say this.
Here you can watch him use it. He did actually analyze how much he saved in token use. There will be a link to this below so you can, if you want to see the videos, see them.
Let's go on to number six. This is Nora. It's from Tencent.
You know how you have all this data on your computer and it's hard to find? And God knows it's really difficult to create a system that will help you find them. Well, what Tencent did is they decided to create a tool that will make it really easy.
I'm going to play this just so you get to see what it looks like. And I should say. It is very much in Chinese, but the video on their site shows how it works and what you get back.
And essentially, how would you describe what this is? You had a really good description for this, Adam. Yeah, it seems like Obsidian plus RAG plus AI plus a graph database.
It's like all these things and it's meant to be. pretty automatic where it can automatically determine oh this project is related to this so i saw this in your downloads folder it should relate to this and and then it builds out these this graph and the wiki automatically so in theory you can drop in and and kind of just click through all these different documents who came to this meeting oh i could click on this person's name oh that shows me the projects i worked on with them and it shows me the other meetings i was in with them let me click on this meeting oh this is what we talked to like so you can follow it just like you would on on wikipedia you can click all the links and and because it has this engine built in and it's doing rag automatically on everything.
It's trying to build this one knowledge graph for everything that's on your computer, which is probably just about everything that's in your life for a lot of people. All right. Let's go on to the next one.
It is a sponsor, Zapier. If you're building anything, I will tell you, go use Zapier SDK because if you want to connect the software you build to the software your users want, Zapier SDK makes that easy. And of course, if you're just working with agents and you want them to have limited control to your tools, MCP lets you connect your tools but keeps it easy.
Regardless of which one you use, let them know Andrew sent you or the GitHub show. Next. This is OpenAI plugins.
It's 62 packs left open for you to read to see how they connect to Figma and also what are the skills that make it work in Figma. Can I just now take this and give this to Claude Code if that's what I'm using and saying now you connect to Figma and you use these best practices to create a canvas and so on? Yeah, you can.
Claude will do one of two things. These plugins are built for Codex and so Claude will either convert it into a format that works for claude and install it that way or there is a a plug -in connector that allows you to connect cross -platform but through an additional layer so claude will decide one of those two things there's also a similar plug -in marketplace for claude though as well so i'd be going to both of them and finding the ones that are custom designed for the platform i'm using or i'd ask it to you know make a custom version for me Okay.
But these are great, Andrew. My favorite part of these is security. You know, a skill doesn't have a whole lot of control.
The AI is still in charge. But when you start to install these plugins, it's like installing software on your computer and you're giving it access to all your data, your code, your secret keys. And plugins can have scripts.
They can have code that runs in them. And if you don't trust where you installed them, things can get pretty bad pretty quick. So I like installing plugins direct from OpenAI or from Cloud.
Okay. All right. Only, by the way, 6 ,000, 7 ,000 stars as of this moment.
It's not even that much. There's not that much passion for this compared to skills, for those of us who are wondering whether people care about skills, versus like a quarter million stars for the one we saw earlier, ECC. Okay.
Next, the number eight. trending repo for the week with 7 .9 thousand stars is called work trunk five agents five copies of your project one tool it's a command line tool forget work trees separate working copies of the same project that sits side by side on your disk work trunk makes creating listing merging and deleting them one short command each so several AI agents can each have their own copy I have like a very basic question here To the person who in the comments said Andrew doesn't know anything, you're right.
I don't. What is a work tree? Why do I need a work tree even?
What is this? Yeah, work trees, if you're not familiar with them or haven't used them, man, this is going to change your life. So listen up.
Work trees are like branches. A branch in Git is, hey, I've got my code over here, but I want to copy it and work on some changes. without those changes affecting the main version and you could have multiple branches that you're all working on at the same time a work tree takes that and says well instead of having multiple branches let's have all those branches but then also create individual folders on my computer so that i can check out all of those branches at the same time and now that you've got five copies of your code at the same time, each for a different feature.
You could put five agents on them, and one agent can work on one, one agent can work on the other. I'm sure you've seen, Andrew, given that you're probably only checking out one branch at a time, if you try to run multiple agents to edit the same thing, they start to run into each other. They start to edit the same files, and they'll say, oh, somebody else updated this.
Oh, somebody else updated this. And they start to argue about who gets to change it, or they'll break features. And so this WorkTrees is like, why not just give each of those agents its own little playground where it can have its own copy of the code, it can test it itself, it can edit, and then once the feature is done, you can merge all of them back together into your one main branch.
All right, I found a really good video. Walking through it, and I'll link over from Seth Fano. Really good use case if you all want to take a look at that.
Next, number nine, agent skills. 25 workflows that stop your agent from taking shortcuts. This is another skill pack.
And again, when we went through this before, you said, Andrew, scroll down to this and show this. What is this? Why is this the important thing?
This is it. Some of these that we look at, it's a bunch of random skills or it's a collection of workflows. This is one workflow.
You're meant to run these skills in this order. This is exactly how a professional software engineer would develop something, too. They'd say, well, first, I need to define what the thing is.
I need to make sure that the definition is good. I need to turn that definition of I want to add a login page and whatever the feature is. I want to add a login page.
Well, a plan. Okay, if we wanted to add a login page, we need the page. We need a logout page.
We need the button. We need a users table in the database. The plan takes your visual spec and turns it into something that could actually be built that's hopefully exhaustive.
Then you build it. Then you code, which is... Probably where most people start normally, right?
They skip spec, they skip plan, and they just say, build a login page for me right now, go. And then you miss out on all these other things that have to change too. Then, you know, test, review, and ship.
Test and review are the biggest parts that are skipped after you're building. Well, did it work? Was it complete?
Did you test it? Did the AI test it? And then for ship, you know, the...
The idea of testing doesn't stop when you've tested it on your computer. I've seen a lot of people send me something and say, well, it worked on localhost, but then when I deployed it to Railway or Netlify or whatever, it didn't work there, but I didn't know until users started complaining. Well, you've got to test there too.
So part of ship is this production validation. All right. We've got a link to it.
It's been doing really well. Let's take a look at how many stars. 95 ,000 stars.
And there's a video here walking you through it. Finally, the 10th trending repo for the week is Cloud Code. It's open source now, Andrew.
Is it open source? What? That's what we were looking for.
It's not. Why not? We were looking at why they were showing this, why we need this if it's not open source, if I can't just take it and create Andrew's cloud code.
Yeah. What are they doing here, though? I think the reason that they have posted this up is that given that cloud code is not open source, when people find bugs with it.
There's not a great obvious place to go and drop those. They could have made a web page somewhere, but people like to come to GitHub and create issues and track them there. It's also a great way to track it internally for their team that's making the changes.
And so I think they put all of this here. It's like a normal open source repository minus the code. A normal repository has a readme, install instructions, this issue list, pull request.
They have all that stuff. and documentation and code. This one is just missing the code.
It's everything else. But if you have a bug, this is a great place to go and report it on cloud code. I see.
Okay. So this is, I could see why this would be useful. I can see why they put it up.
It's not as exciting as I was hoping, but I get it. Let's go on to first stars. This is where people who watch us submit their repos, and I often give them the first star right here while we're talking.
And so the first one that we're going to look at is logic loop, five agent terminals running, and the thing that breaks is your memory, not theirs. Here's the point. When I built the site for today, I used Cloud Code Fable to design the pages.
I used... codex to find those videos that we decided not to use i use cursor to let me do some remote work on it and all of those were in different places what this wants to do is bring them all together into a single place where they're all one tab away and easy to follow that's the point right yeah instead of opening codex and cursor and all these things those are all centered around the agents this one is centered around projects And in one project, you might use, yeah, five different types of agents.
So it makes sense to open up a project workspace and then pick your agent, not the other way around. I gave it its fifth star. And I'm noticing people are getting lots of stars after this.
I want to make sure that we talk about as many as possible. All right. This one is called Cloud Dashboard.
Every cloud code session on your machine, one board that tells you which one is stuck. Here it is. I'm going to give it a second star.
And the screenshot, I think, is really helpful. I love when people are including screenshots. And we still need to do that.
Guidelines for creating a good readme file, which I will get on when I get a little time. Next, I've gone back and forth with him so many times with the creator of this. It's called Deck Gauge.
And I said, why do we need this? Why can't you put everything in Jira and have all those tasks come in? Why do you have to take...
tasks out of Jira, GitHub, etc., and put them into your own thing. And he made a good case for the way that he wanted it run. He wanted a few things that were available on other platforms, like the ability to naturally see where things are, to move tasks along when they're done, not necessarily when you slide them over.
This is what it looks like. I hope I'm doing a good job explaining how it's putting it all together. He's got a test site here somewhere.
Let me hit test. There it is. You can go to the demo page.
You can use his test email and URL, and you can try it. It's an internal tool that he decided, hey, you know what? I like this whole spirit of letting other people learn from each other.
I'm going to make it open source. Next, Campus AIOS. Some of you may have seen Pulsia.
It's a one -person company. A zero -person company, actually, because you are the person, but you're like a chairman of the board or something. And then the AIs run it.
I think that's the idea behind this. It's an AI department or set of departments that runs your company. Interesting idea.
I like this. Yeah, I like the idea of the compounding memory, too, as all these different agents talk together and learn over time. Over the last few days, the Grok team on X has been going live showing how you can use Grokbot to run a zero person company or how to build your business with it.
And so it's been interesting to see it. It's interesting to see that others are building it. Finally, Reckoner, we talked about this before.
And after I created it, he sent me some screenshots, and I thought I would show it to you. What this does is, you know, you're using lots of different LLMs, and you want to see, where am I? Am I running out of credits?
How much money have I spent? And you often don't find out about it until you run out of cash. What he did was he created a dashboard that does this, and I love that I can see from that bottom corner of his screenshots that he's doing this on an iPad, because I still do a lot of my work on an iPad, and I like that this is something that is usable.
remotely like that and i'm proud that did we give him a star last week i'm proud that we're getting to promote again let's see where he is since that look at this we did and he clearly was in the single digits now he is going bigger all right if you found this useful do a few things for us please subscribe like and comment the comment that i'd love to get is were the videos helpful when i just kind of talked over them was it useful for you to see a little bit more in depth what else do you want to see you can tell that we're clearly studying everything you send over and trying to figure out how to be helpful so send it over in the comments subscribe and if you think this top 10 list is good i've got another top 10 that you need to see and it's on your screen right here i'll see you in there
The Hook
The bait, then the rug-pull.
Ten GitHub repos climbed the charts this week, and the range says something about where AI coding tools are heading: a security scanner from Alibaba that catches a password left in plain text, a hobbyist's live map of public cameras and flight data, and a comment section split 3-to-1 over whether AI 'skills' files are actually worth the setup time.
Frameworks
Named ideas worth stealing.
06:26list
ECC's spec-to-ship pipeline
Write the test case
Build the feature
Run the test
Ship
ECC installs a full engineering process into Claude Code so a solo builder gets decades of accumulated software practice for free, starting with a test ('2 plus 2 should equal 4') before any feature code gets written.
Steal forany agent-built feature with a checkable success condition
19:03list
Agent Skills' five-stage workflow
Spec
Plan
Code
Test
Review
Ship
Defining a full spec first (a login page needs a logout page, a button, and a users table) catches the parts that get missed when someone just says 'build a login page, go.'
Steal forany feature request phrased as a one-line command to an agent
CTA Breakdown
How they asked for the click.
VERBAL ASK
26:25subscribe
“please subscribe, like and comment... if you think this top 10 list is good, I've got another top 10 that you need to see”
Standard end-of-video ask plus a specific engagement question (was talking over the clipped demo videos useful?), followed by a pointer to a related video on screen.
Add Modern Creator as a preferred source and Google shows you more of our breakdowns in Search, Top Stories, and AI Overviews. It only changes what you see, and you can undo it in your Google settings anytime.
Add to Preferred SourcesOpens your Google source preferences with us pre-loaded. Tick the box and you're done.
Two hosts run through ten GitHub repos and skills that fix one narrow AI-coding annoyance each, then argue about whether skills are already becoming obsolete.
A weekly roundup episode where the host and a Zapier guest run through ten open-source repos built to pair with a coding agent, from screenshot-to-code to a proxy that cuts terminal noise by up to 90 percent.
Two hosts count down the week's ten hottest GitHub repos, from a security-scanned Claude plugin index to a self-hosted finance tracker wired to 2,500 banks.
Andrew Warner and Adam run down this week's most-starred GitHub repos — three AI design tools, a parallel-agent "tokenmaxxing" rig, a video-comprehension skill — and pull the geophysicist who built the #1 trending AI job-search framework on camera to explain how it got him hired.
Andrew Warner and Peter Cooper run through 10+ GitHub repos that give AI agents cheaper web access, less token bloat, and a design taste system — all free and ownable.
A host runs eight creators' AI-agent experiments back to back, from a $60K app clone to a 150GB video edit, and scores which model actually wins each job.
00:20
00:44
01:02
01:30
01:50
02:30
02:50
03:10
04:10
04:22
04:44
05:31
05:51
06:11
06:51
07:11
07:31
07:51
08:06
08:31
08:51
09:11
09:31
09:51
10:31
11:12
11:32
12:12
12:36
12:52
13:32
13:52
14:12
14:32
15:12
15:32
15:52
16:33
17:13
17:33
17:53
18:13
18:33
18:53
19:33
19:53
20:13
20:33
20:53
21:34
22:14
23:13
23:34
24:14
25:12
25:40
25:54
26:38
22:57
13:57
18:45
37:04
32:25
28:03