27:32The Next New Thing · Interview
How to Have Claude Run Your Company
A zero-employee founder rebuilds his LinkedIn-to-email funnel live, showing every tool he's wired straight into Claude Code.
September 1stThe argument in one line.
Most of this week's trending AI-coding repos are single-purpose skills that patch one narrow failure, but even OpenAI just retired its own standalone skills format in favor of plugins that bundle a skill with a real integration.
Read if. Skip if.
READ IF YOU ARE…
- You use Claude Code or Codex regularly and want to know which small add-ons other developers actually installed this week.
- You're deciding whether to package your own repeated AI instructions as a formal 'skill' or just keep typing them.
- You want a running pulse on GitHub repos and tools shaping AI-assisted coding, with real star and fork counts attached.
SKIP IF…
- You want a deep tutorial on any single tool here — each repo gets a few minutes of screen time, not a walkthrough.
- You don't use Claude Code, Codex, or GitHub-based AI coding tools at all.
The full version, fast.
Two hosts run through ten GitHub repos, mostly Claude Code and Codex skills, that each solve one narrow AI-coding annoyance: forcing terse answers instead of wordy ones, installing a full plan-test-build-review process, training the AI to reach for the simplest existing solution instead of writing new code, and others that manage context, strip AI writing tells, or connect an agent to thousands of outside tools over MCP. The throughline is that a skill packages a habit you'd otherwise repeat every prompt, but its real test is whether it beats just telling the AI what you want out loud. The hosts land on OpenAI's shift from standalone skills to bundled plugins as the more durable pattern, then close with four audience-submitted repos.
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.
Create a free account →Where the time goes.
00:00 – 01:40
01 · I Have ADHD
A 40.5k-star skill forces Claude to lead with the answer instead of wordy hedging; Hacker News debates whether this needed to be a skill at all versus one CLAUDE.md line.
01:40 – 03:02
02 · ECC: Give Your AI Coder an Engineering Process
ECC installs a plan-test-build-review-remember process once instead of re-explaining engineering discipline in every prompt.
03:02 – 05:20
03 · Ponytail: Strategic Laziness for Developers
Ponytail trains an AI to act like the veteran 'ponytail guy' who reaches for the simplest existing fix instead of the novice who overengineers everything.
05:20 – 05:55
04 · Zapier MCP
Zapier's MCP server and SDK both connect an agent or your own code to the same 8,000+ app catalog.
05:55 – 08:22
05 · HumanLayer Skills
Five small add-ons include one that rewrites CLAUDE.md so each rule only fires when it actually applies, fixing a quiet Claude Code compliance bug.
08:22 – 08:53
06 · Should We Skip Skills?
One host says he can usually just dictate instructions out loud instead of installing a packaged skill; the other argues workflows differ enough that skills still earn their keep for some people.
08:53 – 10:17
07 · OpenAI Plugins
OpenAI's Plugins repo works like an app store for Codex: 62 ready-made plugins bundling a tool connection (Figma, Notion, Stripe, Slack) with the best-practice skill for using it well.
10:17 – 11:36
08 · Archify
Archify turns a codebase or process description into an interactive diagram instead of the thousand-line text writeup AI defaults to.
11:36 – 13:23
09 · Context Mode
Context Mode claims to cut AI conversation memory clutter by 98% (or 99%, depending which part of its own page you read); the hosts flag the risk of trimming a detail the model later needs.
13:23 – 15:29
10 · Matt Pocock Skills
A 25-skill collection from a well-known TypeScript educator became one of the most-starred repos on GitHub; the standout is a five-line 'grill me' skill that interrogates the user's plan before acting.
15:29 – 16:07
11 · Why Skills Can Teach You New AI Workflows
You don't need to install 'grill me' to get its benefit, but seeing it packaged as a skill is what teaches the underlying habit worth copying.
16:07 – 17:59
12 · Humanizer
Humanizer rewrites AI-generated text to read like a person wrote it while keeping the same facts; one user ran the output through an AI-detector and got a 100% human-written score.
17:59 – 18:54
13 · OpenAI Skills Are Deprecated
OpenAI's standalone Codex skills repo has been deprecated in favor of the plugin repository covered earlier, a sign the standalone-skill format wasn't enough on its own.
18:54 – 19:28
14 · Audience Repo Submissions
The show is getting enough audience submissions that the hosts float building a dedicated submission site, something like Product Hunt for new repos.
19:28 – 20:05
15 · Reckoner
Reckoner consolidates fifteen open browser billing tabs across providers like OpenRouter, Anthropic, xAI, Railway, and AWS into one dashboard that warns before a job dies from a drained account.
20:05 – 20:36
16 · Tim Harris Skills
A 23-skill collection built for people who never read the code themselves, tied together by one orchestrator session.
20:36 – 21:34
17 · Clodex
Clodex keeps Claude Code's terminal but swaps in a GPT model underneath; the hosts note most people who prefer another model just switch to the Codex harness outright.
21:34 – 22:41
18 · AI Marketing Panel
AI Marketing Panel builds a synthetic focus group of researched personas that scores marketing copy before a real customer ever sees it.
22:41 – 22:57
19 · Wrap Up
Sign-off with the three-action CTA card: subscribe, like, comment.
Lines worth screenshotting.
- Wrapping every AI answer in throat-clearing paragraphs is a default failure mode of chatbots, and a single blunt instruction in a project's instructions file can fix it without installing anything.
- Most Claude Code users never edit their own system-wide CLAUDE.md file, even though it works like a permanent instruction layer applied to every project.
- AI's default instinct is to build first and test later, the reverse order an experienced engineer would default to.
- Given the same task, a novice tends to overengineer the solution while a veteran developer reaches for the simplest fix that already exists.
- Reaching for a browser's native date picker instead of installing a library, writing a wrapper, and adding a stylesheet is what strategic laziness looks like in practice.
- Every extra line of code you didn't need is a line you'll have to maintain the next time you add a feature.
- The excitement of a fresh AI-coded project commonly turns into frustration by day ten specifically because nobody was strategically lazy about scope early on.
- A skill can package a habit you'd otherwise have to retype into every single prompt.
- Dictating instructions to AI out loud instead of typing them removes much of the time savings a packaged skill is supposed to provide.
- A tool built to prove it strips 'AI tells' from writing displayed a 98%/99% inconsistency in its own marketing copy, undercutting its own pitch.
- Long AI conversations resend the full message history on every turn, so a memory-compression tool that trims away a needed detail forces an even more expensive full-context re-fetch.
- OpenAI discontinued its standalone Codex skills repo in favor of a plugin format that bundles a skill together with a pre-built integration and runnable scripts.
- A repo can hit the top of GitHub's trending list mainly because a reputable creator vouches for it, even when the underlying code is only a handful of short files.
Skills package a habit, plugins make it durable
WHAT TO LEARN
Every skill covered fixes one narrow annoyance in AI coding, but the real signal is that even OpenAI just merged its own skills into a bundled plugin format.
01I Have ADHD
- A 40.5k-star skill forces Claude to lead with the answer instead of wordy hedging, and most of what it does could also be one blunt line added to a project's own CLAUDE.md file.
- Most Claude Code users never edit their system-wide CLAUDE.md, even though it works as a permanent instruction layer applied to every project they touch.
- A 371-comment Hacker News thread split on whether a skill like this needed to exist at all versus just being a single rule in an instructions file.
02ECC: Give Your AI Coder an Engineering Process
- Installing a fixed plan-test-build-review-remember process compensates for AI's default instinct to build first and test later, the reverse of how an experienced engineer works.
- This kind of installed process particularly helps someone without a product or engineering background, since the AI otherwise won't surface that discipline on its own.
03Ponytail: Strategic Laziness for Developers
- Given the same task, a novice tends to overengineer the solution while an experienced developer reaches for the simplest fix that's already sitting there.
- The canonical example: asking for a date picker gets you an installed library, a wrapper component, a stylesheet, and a timezone debate, when the browser's native picker already does the job.
- This 'strategic laziness' isn't about avoiding work, it's picking the option that's cheapest to maintain later, because more code today is more code to break when you add features.
- The excitement of day-one vibe coding commonly turns into day-ten frustration specifically because nobody was strategically lazy about scope early on.
04Zapier MCP
- Zapier's MCP server and SDK both point at the same 8,000+ app catalog: MCP lets an agent call those tools directly, the SDK wires the same connections into your own code.
05HumanLayer Skills
- One HumanLayer add-on fixes a quiet failure where Claude Code treats a project's instructions file as optional and skips parts of it, by rewriting the file so each rule only fires when it actually applies.
- The broader value of a skill pack is packaging a prompt you'd otherwise retype every session, like 'use an HTML report' or 'make it visual,' into one reusable command.
06Should We Skip Skills?
- Skepticism about skills in general comes down to this: if you can just dictate instructions to AI out loud, a packaged skill often isn't saving you as much as it looks like it should.
- Different developers have different workflows, so a skill that feels redundant to one person can still be the missing piece for someone else.
07OpenAI Plugins
- OpenAI's Plugins repo bundles a tool connection with the accumulated best-practice knowledge of using that tool well, which is why a plugin can make an agent follow a proven workflow instead of improvising one.
- A plugin's value isn't just the connection to an outside tool, it's the framework for using that tool correctly that comes bundled with it.
08Archify
- Archify turns a codebase or process into an interactive diagram instead of the thousand-line text writeup AI defaults to producing, making it easier for both a human and a later AI session to spot and reuse the structure.
09Context Mode
- Context Mode claims to cut AI conversation clutter by 98% or 99%, depending which part of its own marketing page you read, which is itself evidence of how often AI-written copy loses internal consistency.
- The tradeoff with any context-trimming tool: if it strips out a detail the model later needs, the AI has to re-run the tool call and pull the full context back in, which can erase the savings in a long-running conversation.
10Matt Pocock Skills
- A 25-skill collection from a well-known educator became one of the most-starred repos on GitHub largely because of his existing audience, not because the underlying code was complex.
- The standout skill in that collection is roughly five lines long: it just tells the AI to interrogate the user's plan or decision before acting on it.
11Why Skills Can Teach You New AI Workflows
- You don't need to install a 'grill me' skill to get its benefit, you can tell your own AI to grill you and get the same result; what's worth studying is the mental model, not the code.
12Humanizer
- Humanizer rewrites AI-generated text to read like a person wrote it while keeping the same facts, and one user's test through a separate AI-detector came back scored as 100% human-written.
- The hosts expect tools like Humanizer to matter less over time as underlying models keep getting measurably better at natural-sounding writing.
13OpenAI Skills Are Deprecated
- OpenAI discontinuing its standalone Codex skills repo in favor of the plugin format is itself evidence that a skill alone wasn't a complete enough unit.
15Reckoner
- Reckoner consolidates fifteen open provider billing tabs into one dashboard that warns before a low balance kills a running job, built by someone burned by exactly that failure.
16Tim Harris Skills
- A 23-skill collection aimed at people who don't read code relies on one orchestrator session to tie the individual skills together into a single usable workflow.
17Clodex
- Clodex keeps Claude Code's terminal interface while swapping in a GPT model underneath, which mainly matters if you're attached to that specific interface rather than to Claude itself.
18AI Marketing Panel
- AI Marketing Panel scores marketing copy against a synthetic focus group of researched personas before a real customer ever sees it, and the more specific detail you give each persona, the sharper the feedback you get back.
Terms worth knowing.
- CLAUDE.md
- The project-level instructions file Claude Code reads before every session; several skills in this episode exist specifically to make it more reliably followed.
- MCP (Model Context Protocol)
- The connector standard that lets an AI agent call outside tools and services directly, instead of being limited to its own model output.
- Skill (Claude Code / Codex)
- A packaged instruction file that gets an AI assistant to repeat a specific workflow automatically, without the same directions being retyped every prompt.
- Plugin (Codex)
- A bundle that combines one or more skills with a pre-built integration and runnable scripts or commands for a specific outside tool, such as Notion or Slack.
- Strategic laziness
- Deliberately reusing the simplest existing solution instead of building something custom, to keep a codebase small and cheap to maintain later.
Things they pointed at.
07:46toolFluid Voice (local dictation tool)
Lines you could clip.
00:00
“If you act like you have ADHD, your AI will be a lot smarter.”
03:36
“Amateur will install Flatpicker, writes a wrapper component, adds a style sheet, and starts discussion about time zones. The ponytail guy would just use the one that's in the browser.”
04:35
“By day 10, they're frustrated because every change you request breaks something else.”
11:58
“98% is a lot of percents, Andrew... it says 99% in this image, and if you do a search on the page, it says 98%.”
14:25
“Grill the user relentlessly about a plan, decision, or idea.”
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.
metaphorstory
If you act like you have ADHD, your AI will be a lot smarter. You'll get a tool that shows you why and fixes it for you. You're going to get a tool that will turn your AI into a veteran developer like that.
OpenAI just killed their skills and replaced them with something better. You're going to get them. All that and so much more coming up.
Presented by Zapier, the AI automation company. All right, the number one repo of the week is I have ADHD. Listen, if you're using Claude, if you're using any AI, you know that it gets so freaking wordy.
And as you're trying to understand what is it saying, you get lost, you don't get it, and you blame yourself. Well, this says, you know what? What if it treats us like we have ADHD and it just gives us the facts and doesn't give us so much information that our mind wanders?
It has been popular before, Adam. It is popular again, number one. you think of this one?
I mean, I love it. Probably my most common prompt is, I'm not going to read that wall of text. Please tell it to me in a way that a human could understand.
And my understanding of this one is it's basically a skill that just does that for you automatically, which I would definitely appreciate. That's exactly it. And now there's controversy, of course, on Hacker News where people say, well, couldn't I just replace this with a line in CloudMD?
And so since you say this a lot, why don't you replace it with a line in your CloudMD? That's a great idea. I'm going to try that and see what happens.
I will point out, Andrew, I think fewer people are editing their system -wide CloudMD than they should be. People should be adding little rules there, updating things. They're like system -wide defaults.
Cloud can help you update that file. It'll also tell you what's in it, if anything's in it now. But yeah, I might add this line there and see what happens.
Next, ECC, give your AI coder a whole engineering process. Build, test. Plan, test, build, review.
Remember, my approach is always build, then test, and then move on from there. You install it once instead of re -explaining in every prompt. This is by someone here.
This tweet says it all. It's by a guy who won Anthropics Hackathon, and he gave away his entire Cloud Code playbook. This is kind of your model for building things, but you're not a super fan of this type of skill pack.
I like this for people that don't have the product building or engineering mindset, that don't have that background. It's nothing wrong with you, but we've figured out over a long time how to build things well. And if you don't have that, you're going to struggle to vibe code because the AI often won't bring it up.
It very happily will let you build first and then test later and then rebuild and then test and then rebuild. But if you just have something like this installed, it's going to encourage it to say, wait a minute. hey, before we build, Andrew, could we talk through some of the plan?
Could we write a specific plan doc? And that's going to help you get quality outputs and a lot less time and fewer tokens, which is money in a lot of cases, right? Like if you want to get to a better output faster, you're going to save money at the same time.
We've got this really good short here, but even though it's a short, I'm not going to hit play on it. Instead, I want to keep things moving because that's what you all are asking for. And I'll have a link to this whole site that I created to guide us and you can have it.
All right. Next is Ponytail. Really interesting for developers.
What it does is it says, look, you know that experienced veteran developer or the guy with the ponytail who's sitting in the back? If you give a novice a project, he's going to overthink it. But if you give it to , he's going to find a quick, easy way to get it done.
And here's the example, I think, right from the repo that I think nails it. And we've talked about it before. If you ask for a date picker, that...
Amateur will install Flatpicker, writes a wrapper component, adds a style sheet, and starts discussion about time zones. The ponytail guy would just use the one that's in the browser. Keep it simple, which saves you tokens, keeps the product that you're building light.
And what do you think? I think of this as good . You know, the ponytail guy like knows how he's been burned before and he's not so ambitious to write everything from scratch.
And he knows he's going to have to solve problems when it breaks later. So it's about strategic laziness. And this is another one where AI's default is to be gung ho and build everything new, build everything fresh.
And you need to if you want it to be strategically lazy, you need to bring that in that impulse. And then, yeah, you'll see here like AI will write a. big block of code.
When maybe it could have just changed one or two things in an existing line and that would have been enough and it would have gotten you there. And both of these solutions will work. But down the line, as you need to maintain it, as you add new features, the more code you have, the more complex the whole thing gets.
And this is why everybody's really excited the first day they're vibe coding something. By day 10, they're frustrated because every change you request breaks something else. And it's because you've got this massive bloat that's around.
You weren't strategically lazy. Adam, this has become a way of life for me ever since we've talked about this. Like my wife and I will talk about how do we find a really good place to go to eat dinner in this new Mexican city?
Where are we looking? What are we doing? I said, you know what?
What would the ponytail guy do? The ponytail approaches to reach out to my best friend who's a foodie and happens to live here and just ask him. And that's the easy way.
Just go wherever he sends you. You don't need to solve it from scratch. You don't need to open Google Maps and read all those reviews.
Yeah. Next is Zapier MCP. If you are building a tool that needs your users to connect to any of over 8 ,000 other tools, you need Zapier SDK, the one on the right over here, because that will allow you to build your tool and make it easy for your users to connect to the other stuff without having to rewrite this whole process.
It's available right now at zapier .com slash SDK. And of course, I've talked about the MCP, and that will allow your agent to connect to all those tools. Both of them are available for you right now.
And I don't have a unique URL. love you all for asking is there a unique url because you want to give me credit i don't have one just let them know that andrew sent you or the next new thing and uh let's move on now to skills by the way well let's talk about what this is five small add -ons from human layer team including one that fixes why claude code keeps ignoring your rules you keep uh You can use it to fix a quiet problem.
Cloud Code is told your project instructions file may or may not be relevant, so it ignores parts of it. One of these five add -ons rewrites the file, so each rule only fires when it applies. We've looked at this before a lot.
What is it, and what do you think of it? I think, you know, scroll down. I think there's one at the bottom that makes here, show me.
Yeah, the very last one, show me.
This is a prompt we've all probably written many times, which is, hey, explain this thing to me, whether maybe you're looking at a database or a bunch of customer reviews. Explain all of this to me. Summarize it.
And then you have to tell it every time. Use visual layouts, like have it be pretty. Use an HTML report.
Summarize things well. Don't be overly like all of that. You have to say over and over and over again.
So these skills basically take those workflows that you would have to paste every time or remind it every time. And they just package it into one thing that gets you the result in a heck of a lot less effort on your side. Adam, by the way, this week is very skills heavy.
A lot of these top 10 repos this week are skills. People know you as a guy who's not super enthused about talking about skills. Why not?
You know, the workflow that you have, Andrew, is probably different than the workflow that I have. And I want different things. And also, honestly, AI is pretty good.
And let's take that last one as an example. I could go and I could install this skill and it comes with other skills. Sometimes they'll be used.
Sometimes they won't. But also, if I just dictate my direction to AI, I can say everything out loud pretty quickly. And I'm not limited by how fast I can type.
And it's local AI. It's free. I use this thing called fluid voice.
And so I just don't often find that the skills replace much real work that you couldn't just. direct the AI to do to begin with. Or they have some wildly specific flow that was great for you, but isn't great for me.
I will ask viewers, do you think we should skip skills? Should this be a top 10 list that ignores skills and is just top 10 repos that are not skills? Let me know in the comments.
And I'd love it if you don't just say, yes, skills or no skills. So that's great. Let us know why you say that.
Next. Well, and tell us then, is there... a skill that you've downloaded that you're using all the time you know I've downloaded a bunch most of these I remove but are there ones where like no you've got to try these and I've tried this one before this is awesome so I'm very curious who's using them for real all right I'd like to hear that too This is OpenAI's plugins.
This is the App Store for Codex. Let's take a look at the 62 ready -made plugins from them. And I go into here.
I go into plugins. By the way, a lot of people don't even go into these folders, and I love that you taught me to go into them and what they do. It's not as scary as I thought it would be.
In fact, here, this is one for Notion. It will both connect into Notion, and it has skills for using Notion well, like meeting intelligence skill right here. Let's look at the skill.
Look at this.
What do you think? Yeah, these are nice because. You know, you could just connect Codex to Notion, right?
That's easy to do. But Codex doesn't necessarily know how to use Notion well. You know, using Notion well is it's best practices.
It's good organization of all of your documents. It's using links between documents correctly. There's a lot to think about, you know, when you're really setting up a Notion workspace or any of the things, linear, any of the things that are included in here.
And so it's not just the connection. So what these bring. are the framework and the method of how should Codex use Notion.
And now when you ask it to do meeting notes, like, oh, here's all of this great stuff, and it's going to not just create meeting notes for you fast, it's going to follow this exact path to do it really well. All right. There it is.
It has been growing, and we'll talk in a moment about why there's this sudden uptick in this, and it is from OpenAI. We'll have links to that below. Next.
Archify, we talked about this last week. It basically will take your code and turn it into a diagram like this. You've said before why this is useful.
Why don't you tell us again? And we'll see if anyone has actually used it since last week. You know, AI is pretty verbose, right?
And so if you ask AI to document something, it's going to write hundreds or thousands of lines of text about your application. But it's pretty hard to find the detail there. There's a lot of garbage.
It's hard to hide garbage. in charts like this. So when you force the AI into something visual and you give it a tool like Archify that makes creating the visuals easy, suddenly you have a really concise, detailed way to describe your application or your sales process or really anything you could chart, which is just about any process.
So when you do that, it makes it easier for you to understand it and make edits. And it makes it even easier down the line for the AI to understand again and say, oh, I see the world. workflow here.
You're asking me to add a step between steps C and D. Great. I'll go and do that versus, you know, trying to find it fresh in the code every single time.
All right. And it's really pretty. I like what they do.
Context mode is the seventh most popular repo this week. The one that's had the most growth, your AI assistance memory fills up with junk. This keeps it out 98 % less clutter and remembers where you left off.
You're a little skeptical about this one? Yeah, I, you know... 98 % is a lot of percents, Andrew.
So 99%. Basically saying nut. Right.
Which is, by the way, weird. It says 99 % in this image. And then if you do a search on the page, it says 98%.
Not to like. Yeah. I mean, a lot of these things, like, you know, not to knock it, we use AI for everything.
But a lot of this stuff ends up AI written and maybe you make edits, but AI kind of loses its case. Although now that I'm thinking about it, if their whole point is that this is a great memory layer to keep AI self -consistent, but then. It wasn't consistent in helping them create 99 versus 98.
It's a little questionable. But yeah, some of these can be useful, especially if you have a lot of short conversations. They tend to not work as well on very long conversations because there, every time you send a new message, the whole conversation history is billed back to you as tokens.
And if you run a tool call... and you compress away some little detail that AI actually needed, what happens is the AI has to run the tool call again and request the full context. And at this point, you have to send the whole message history back to AI.
And if we're talking 300 ,000, 400 ,000 tokens, that's every message you're going to send everything back. And so it's just very inefficient when it gets it wrong. When it gets it right, I'm sure that it saves tokens.
And I bet that's more likely to happen in shorter conversations. I've reached out to this creator now. Every time there's one that a repo that hits our video, I email them and a lot of them want to participate.
So if there's anything you want to hear from the people behind these repos, let me know. And they're they're now asking to come and do interviews. In fact, Matt Pocock has done an interview here.
It's done really well. He has created a collection of skills that is now one of the top 20 most popular repos, not just 20 most popular skills, 20 most popular repos on GitHub of all time with over 259 stars. That's because it actually is useful and it's kind of super simple.
Like if you go in and take a look at some of his skills, like the grill me skill. It's actually like five lines here. Grilling, is this it?
No, this is going to refer to another one. Here's a skill. Grill the user relentlessly about a plan, decision, or idea.
Use when the user wants to stress test their thinking or uses any grill trigger phrases. And you can go through and see how it works. Super simple.
And then there are other... in here that refer to this and back and forth. It's been incredibly popular.
People like even Theo, who I think is very skeptical of a lot of things, has used the Grill Me skill and gotten a lot of value out of it. Actually, I don't think I did a great job of explaining what the Grill Me skill is and why it's so popular, even though it's so simple. Yeah, I mean, it really is as simple as saying, like, just make sure that you've gotten all the information out of my head, right?
If I give the AI a plan, I say, go and build me this app. And then you run grill me. It's going to say, hold on, Adam.
Do you want it like this or like this? Why do you want this? Do you want this?
And it's going to grill me. It's going to put me on the hot seat. And I think what I do like from these very reputable creators on skills is that the skill is not very interesting.
It's a couple, you know, it's 20 lines long. You just saw it. It's simple.
But what's interesting is the concept. Oh, I should have AI grill me. And so when you're looking at these skill packs, part of what I'm looking at is like, what's the mental model that this person used to create them?
And sometimes you might just be able to ask AI, hey, grill me on this topic. And that might be enough. But it was still I learned that concept by looking at this package of skills.
I think it's a good knowledge, knowledge base to look at. Do you have to install this skill to get AI to grill you? Absolutely not.
Just go and ask it and see what happens. By the way, Theo gets over 400 ,000 views for something like this. I thought my video with Matt that got 30 ,000 views was killer.
Over 400 ,000 views. All right. You'll get there, Andrew.
I'll get there. I know our last video from last week is almost at 200 ,000 views. I can't believe it.
I really thought it would take me a few years to get to that. Now I'm like there's a part of me that goes, oh, man, Andrew, what if this is all going to go away tomorrow? What if you messed up?
What if, I don't know, this can't be. If you lucked into it, then. You don't have that craziness in your head.
That's my craziness. Let's go on to the next one. This is Humanizer.
Paste something that ChatGPT wrote, and it doesn't have to be ChatGPT, and it comes back sounding like a person wrote it. It takes the AI stuff and turns it into human. Here's actually this tweet that I think is the most helpful.
This is Mushtaq. He ran some AI writing through this, and then he ran it through... Through software that tests, where is this?
There it is. And it came back 100 % human written. And I don't think that's the best way of doing it and of seeing if it feels human, but it is one way of doing it.
And it gives you a sense of what it's about. It's about making your AI writing seem more human. What do you think?
Yeah, I think it's good. I think there's a lot of these that, you know, there's like anti -slop and no -slop. very similar sets of skills.
Dude, the models are getting better at writing. I think you've seen it. I know I've seen it.
The new models from OpenAI, they're just better and better at writing. Now, they still suck in a couple ways, but I don't think these are going to be necessary long term. I hope they're not necessary long term.
I just recorded a video about Astra from OpenAI, and the thing that came out from watching a bunch of videos to prepare for it was how good it is at writing. And so I think you're right. If these little collections exist as repos on GitHub, I think the models are going to learn from them and figure it out, and it won't be an issue for long.
Okay, but there it is, and it's available to you. Now, number 10, this is what I mentioned earlier. This is OpenAI's skills, and...
The thing is that it's no longer available. Yeah. Let me get on my skills soapbox here for a second, Andrew, which is this was deprecated and they've replaced it with the plugin repository.
So even here they're saying, oh, the skills aren't quite enough on their own. What makes a skill really interesting is when it's packaged into a plugin with a pre -built integration, with maybe some scripts or tools that are running, and then a package of skills all designed to work together. So like this build macOS apps, you're going to have a lot of skills in here.
You're also going to have these commands that can run. These are like workflows or even scripts on the computer that will run. And so this idea of plugins, if you're really excited about skills and you haven't tried plugins yet.
Go dip your toes in the water. I think that's that's where the ship is moving. All right.
I'm not going to say any more about this because it is deprecated. Instead, I'm going to go to first stars. This is from our audience.
Here's what they sent in. And I'm going to find another way of dealing with this because I want to give everyone the attention they need. And we're starting to get flooded with user submissions from the audience.
And I don't just like to I like to spend time with them. One idea, by the way, that I have is to create a site where people can submit it kind of like Product Hunt for new repos where people can say, look, I want someone to take a look at this the way Andrew does and give me some feedback and go easy on me and be useful.
OK, here's the first one. It's Reckoner. Fifteen billing tabs become one dashboard that warns you before a job dies.
I think your call to action on this is always to put images. And here's a great example of how much an image shows and says this is what. they created.
Useful, right? I mean, this is great. Yeah, I love it.
I like this. Anthropic, $1 .44 left. XAI from Grok, zero remaining.
You get a sense of what it's for. Did I give it a star yet? Let me make sure that I do.
By going up to the top. Adam, people who have had no stars come back to me with now 50, 60 stars. It is amazing.
So you see how many stars there are right now. It was four when I created this, and now it's five. And by the time you see it, let's see where it is.
Okay, next. This is skills from Tim Harris. Agent skills written for people who never read the code.
And everybody never reads the code. 23 skills in here. That's what I was going to ask.
Who reads the code, Andrew? All right. I think this makes a lot of sense.
I'll leave it there, and I'm going to make sure that I hit the star on that one right now. Number four star of the repo. Next, Claudex.
You like that name, don't you? I love it. Keep Claude Code's terminal.
Swap in a GPT model underneath. And your point on this one, it has one star. I'm going to give it a second star.
Your point on this one was maybe people don't need this because? Yeah, you know, I for a long time, Cloud Code was the only game in town. So we all learned it.
We all got familiar with it. And I've seen a lot of these ways to use other models with Cloud Code. But almost universally, Cloud Code is is not revered as the best harness.
People love the Codex harness. So if you like like the the, you know, GPT models, but but you haven't tried Codex yet and you're trying to make them work with Cloud Code. Hey, go try Codex.
See how that. See how that feels. Especially the desktop GUI is like, it's phenomenal.
I highly recommend you try it. But if you really, really, really like Cloud Code and want to use other models with it, then heck yeah, use a package like this. Yeah, I think it makes sense.
Let's go to the next one. AI Marketing Panel. You like this one a lot.
This is a focus group for your marketing copy before a single customer sees it. I like it too. Yeah.
Focus group is cool. I don't know if you remember, but there was a time at the beginning of AI where the only thing regular people used it for was, Ooh, I want to write this in the style of, you know, Michael Jackson or something like that. Like you'd like.
impersonate celebrities, basically. I haven't seen a lot of that lately, but it's a really useful way to use AI to say, hey, I've got a product. Pretend you're a pregnant mother who's, you know, reviewing this.
What's your opinion? Pretend you're a marathon runner. Review this.
And the more that you give, the more detail you give about these personas, the better the panel information you get back. And the models are so well trained now. You can get pretty deep, interesting insights just from the model.
pretending to be different people and especially when you have a package that can help you uh you know systematize that i like that this is from a consultant from long island near where i grew up in queens um all right next now that you've watched this obviously like subscribe we love it when you do give us feedback especially on the skills and if you like these we have another collection that we think you're really going to love and adam and i are going to see you in that one right now
The Hook
The bait, then the rug-pull.
The cold open sells the first repo on its bluntest possible pitch: pretend your AI has ADHD and it stops burying the answer in three paragraphs of hedging. From there the show runs down nine more repos on the same premise, one narrow annoyance, one narrow fix, until the hosts start asking whether stacking narrow fixes is even the right model going forward.
Named ideas worth stealing.
05:06list
Ponytail's Six Commands
- /ponytail lite | full | ultra | off
- /ponytail-audit
- /ponytail-review
- /ponytail-gain
- /ponytail-debt
A command set that sets how aggressively the AI should simplify, audits the codebase for over-engineering, reviews a diff for over-built code before merge, scores the measured impact (less code, lower cost, more speed), and tracks deferred shortcuts as a debt ledger instead of letting them silently become 'later.'
Steal forany AI-coding workflow where scope creep from an over-eager assistant is the recurring problem
01:45list
Plan, Test, Build, Review, Remember (ECC)
- Plan
- Test
- Build
- Review
- Remember
A fixed engineering-process order installed once so the AI proposes a plan and writes tests before building, rather than building first and testing after.
Steal foranyone whose AI coding sessions currently go build-then-test-then-rebuild
How they asked for the click.
VERBAL ASK
22:41subscribe
“If this was useful, do three quick things: Subscribe, Like, Comment.”
On-screen card lists three specific actions with icons; the spoken CTA restates all three by name instead of a vague 'like and subscribe' aside.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
