Ponytail 5 é ABSURDO! O Claude Code escreve metade do código e gasta MUITO menos

Author: Matheus Battisti - Hora de Codar · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Ponytail 5 é ABSURDO! O Claude Code escreve metade do código e gasta MUITO menos · Channel: Matheus Battisti - Hora de Codar · 22:19 · Watch on YouTube

Topics: AI agents, Claude Code, Software engineering with AI, Model inference and optimization

Summary

The video compares building the same small-market credit ledger app with and without Ponytail 5 using Claude Code. In the test, the Ponytail-assisted version finished faster and used fewer tokens, while both versions delivered broadly similar functionality; the main visible differences were in layout and readability. The presenter reports 42% lower cost and 38% less development time with Ponytail, and demonstrates its commands for documenting technical debt and auditing an existing project. He concludes that the savings are promising, but simplified implementations need to be reviewed as a project grows.

Key points

Ponytail 5 is presented as a tool that encourages simpler code to reduce token use and development time.
The presenter compares two versions of a credit-ledger app for a local grocery store, built with the same prompts.
The Ponytail version took about 10 minutes for each development round, compared with about 15 and 18 minutes without it.
Both apps supported the requested workflows, but the presenter found the Ponytail version’s notebook-inspired layout harder to read.
The token-counter results showed 42% lower cost and 38% less development time with Ponytail in this test.
The comparison also showed fewer lines of code and one fewer dependency in the Ponytail project.
The ponytail dep command identifies and documents deliberate simplifications that may need revision as a project grows.
The ponytail audit command can review an existing project and suggest improvements, including a potential server-timezone issue.
The presenter cautions that aggressive simplification can create technical debt, so projects should be monitored and improved over time.

Chapters

Start	Chapter	Summary
00:01	What Ponytail 5 promises	The presenter introduces Ponytail 5 and its goal of reducing code and token usage, then outlines the comparison.
02:07	New features and test setup	He describes version 5’s live cost display, smaller rules, project map, and audit features before preparing two test projects.
05:24	Building the credit-ledger app	The presenter runs matching prompts to build a local grocery store’s customer-credit app with and without Ponytail.
08:14	Comparing the first app versions	He checks both apps and discusses their similar functionality and differences in layout and readability.
10:59	Adding SaaS features	A second prompt expands the app with a product page, authentication, cashier management, WhatsApp collection notices, and month-end closing.
13:10	Testing the expanded apps	The presenter reports the second-round completion times and verifies the requested features in both versions.
15:21	Token results and technical debt	He reviews the token-counter comparison, reports the savings, and demonstrates commands for documenting simplifications and auditing a project.
20:13	Conclusion and cautions	The presenter recommends monitoring technical debt as simplified projects grow and closes with resources and a summary of the test.

Quotes

The idea is to be lazy with the solution and never with the care. — 01:36

The portail one took 10 minutes, and the non-portail one took 15 minutes — 08:14

So we had a 42% lower cost here with Ponytail, right? — 16:23

Some things are kept very simple for the project scope — 17:28

you have to watch the technical debt so things don’t get left behind — 21:48

Concepts and entities

Ponytail, Claude Code, Codex, Appmax, Next.js, TypeScript, Tailwind, SDD, débito técnico

Full transcript (original language)

00:01 Garapony version 5, for those who don’t know, is one of the packages that saves the most tokens for tools like Cloud Code, Codex, and similar, and it has now reached version 5. The main idea of this package is to simplify code, and that’s where the savings come from. In this video, we’re going to compare a project without it to one with it, to see if version 5 has improved as much as they say. Several features have been added; we’ll go through everything here and test it in practice. So come on in, let’s get started. Well guys,

00:33 those who have been with the channel for a while know about Pil, right? We covered it back when he launched a package that is blowing up on GitHub. Over 160,000 stars, and it’s open source, right? It’s open source and works with various agents. I’ll test it here in Cloud Code, but you’ll have the same experience with all of them, because it has a setting that makes the AI write simpler code, which results in savings. Look at that, a skill that makes us write only the code that the task actually requires. Version 5 was rebuilt from scratch and promises half the code, even less than the last one,

01:05 right? Faster and cheaper. Today I’m testing if it’s worth it in a real app. So let’s go. Look at this, here is a comparison, with and without it, right? They obviously improved the performance of the pointer from the last version, but the difference here is striking— the number of lines the AI writes with and without the pointer, from 335 down to 10 in this scenario of adding a date picker to a form. Well, before writing, it goes up a ladder of seven questions, right? So, this is its normal operation ; it’s not a significant change in

01:36 version 5, folks. But here’s what happens: it asks if the code needs to exist, if it’s already ready in the project to be reused, or if the language already handles it without needing external installs. And if the browser or platform already has it, meaning, if the browser has a native feature for it. So, as it gets negative answers, it moves up, and when it finds the exact portion, the exact point, it stops and works with that solution, okay? The idea is to be lazy with the solution and never with the care. So it

02:07 will have security, validation in the things it does, it will work to prevent data loss and ensure accessibility, which is a very important point in web projects. Alright guys, look at what changed in version five. So, now about version five, rebuilt from scratch and it now shows the bill while it works. It’s combining, let’s say, with cloud code mods, showing the savings we’re getting in real time. But even so, I’ll use a little program that will calculate the session cost, and we’ll see if it’s not lying, right? Comparing the costs also with projects that don’t have Pony Tail, okay? So

02:42 here we’ll have this counter, which is an upgrade in version five, okay? Besides that, we have rules that are half the size, meaning they slimmed down the rule files, a project map to write, like a graph plugin that will structure the project so it understands better where it’s working. It tells you what it skipped, so it becomes more explicit, right? Look, Codex working there, it’s clearer what it didn’t do , right? For example, here I skipped exporting to PDF, risk of time zone in maturity. So it will give a reason so

03:14 we don’t end up with vague reviews, but with real audits and audit reviews. So it will have new skills here, last commands to work on this as well. What are we going to do here in today’s test then? Basically, we’ll create two apps, one with pure cloud code, one with Pony Tail, and at the end we’ll see what it shows us, because we’ll have a token counter tracking both sessions. Deal? Let’s go. So guys, before we start, don’t forget to download Fáb Sistemas. It’s my newest e-book, completely free for you. The link will be right

03:48 here in the video description. In it, you’ll find a method to create projects with AI using SDD, which is a methodology I use and is well-validated . And I’ll explain to you how I think , how I structure things to create this project we’re going to build together . And at the end, I’ll still show you ways to get your first users and clients for your app. Deal? Well, download it, don’t forget, okay? Important links. We have Pony Tail here , okay? Right? So, it has updated quite a bit, see, the number of stars there, it’s already growing. We have here, right, the entire installation method

04:20 for it. It’s talking about version five, but you can follow how to install it here. It’s just a few little commands, it’s very simple, okay? A lot of ads here, but look, here are the installation commands. Just scroll down a little and you’ll find it, okay? It can be installed in cloud code, Codex, and other agents. So, repository, I’ll leave the link here in the description. This here was the version release, okay ? So, when they announced it, they basically said what I mentioned on the slide, right? But here it is explained by the creator himself. And here are the review and audit commands, right,

04:52 that we can also use to check the project and the results, okay? That said, let’s head to cloud code. And here I have the structure set up, okay? One folder with components, another one without. I’ll do a demonstration here for you to see. I’m going to start both . Let’s rename the two terminals here as well, just so we’re clear. I’ll even change things here. The one with will be green and the one without will be red. And here’s what happens, look

05:24 , this bar here is the pointer. This full 0.0 seconds is its metric, right? The one without has nothing. So it is installed at the project level only on this one, okay? I’ve prepared some prompts here, okay? which I can also share if you want, just ask in the comments. The prompts I’ll be using will be these here, okay? So here is the installation, right, of Pil, the first prompt, the second, right, which I’ll use to direct this video, guys. So , moving on. Let’s go with the first one, so I’ll put it here, I’ll put it

05:59 here too and read it for you. So it’s like this, create a new project, a folder called “caderneta,” completely separate from the rest, use Next.js with TypeScript, Tailwind, and a database, so it will have a database. Then I explain that the “caderneta,” right? is the tab notebook for the local grocery store, on the counter computer. The owner records what the customer took on credit, receives partial payments, and sees who has been owing the longest. The customer receives a link to check their own balance without installing anything. It replaces the spiral notebook on the counter that goes missing, right, gets wet, and causes arguments. Basically a

06:32 way for the guy to keep track of who owes what, right? So both are running now, and we’ll see how it unfolds, and later I’ll run the script that will give us the actual cost data for each session. So here, guys, what’s happening is it’s not counting the costs. That might be because I’m on a subscription, okay? I’ll check at the end when it actually delivers the results, okay? But it can happen. Even so, my script will work because it has a token estimation, right? I used this in another video where I compared Opus

07:06 versus Opus with several Haiku sub-agents, right? And there I was able to show that there’s savings in using several Haiku sub-agents and the result is still very similar, okay? So, if you want, take a look at that video— editor, put the “Haiku Army” thumbnail there; it’s quite interesting for those who want to improve token usage in Claude Code using only Claude’s own features, without installing anything. But I’ll keep waiting here, guys. It should be almost finished. I’ll be back in 10 minutes. I’ll get back to you guys shortly. Well, guys, the portail one finished, so we have that—it’s a

07:42 classic characteristic of it, right? Because it finishes a bit sooner, since it develops in a simpler way, right? Not that it’s bad or missing things, but it uses native code without installing as many things, so everything ends up simplified, right? Including the tests. So it finished a bit before the non-portail one, which must be near the end too, right? With both here, I’ll start the projects, continue for a bit, and then we’ll see if it saved money or not. Oh, it just

08:14 went through, guys. The portail one took 10 minutes, and the non-portail one took 15 minutes, okay? I’ll ask both to start the project. It’s here too. Let’s see if both delivered basically the same thing, right? And here, guys, I thought the values would be compared for the version with and without, but no. It’s only showing how much the session cost in total and how much the last prompt cost. So, I think my counter will be more effective for this comparison at the end, okay? It’s going to start here. Port 3000 was the

08:48 one without, and port 3002 was with. So let’s go. Opened 3000. Opened 3002. Whoops. See, here we can see there’s a difference in layout, right? This component went more for a notebook vibe . It even used a slanted, quirkier font . And so on, while this one didn’t; it used a more professional font. In my view, I prefer this scenario here because the guy is actually building a

09:20 real system that someone will use. In this scenario here, a person would have to squint to read it, right? So I didn’t like it as much, but a prompt adjustment would fix it, okay? So here I’m entering a debt as well, look. It stays professional, right? Since it uses the more elaborate font, let’s say , only in a few places, right? Where clarity is needed, it’s using the right font. And here it changed the font a bit, right? Now it’s going back to a more normal font, let’s say. And another thing I noticed is that here it

09:53 uses a white background so the notebook lines don’t interfere, while here it’s using it over the lines, which is also hard to read. So there are some negative points with the other one, right, with the same prompt that should be adjusted. And maybe this adjustment doesn’t necessarily cost more than just the cloud version, but it’s something to observe—details matter, okay? So here, register customer. I’ll put a quick one in here so we can see a validation, right? Put it in there,

10:28 look. There’s my customer. Let’s check this one. It registered as well. So the test is here. Alright. So both are working well, right? There are just these layout differences which are minimal and, as I said, the prompt would fix them. Now I’m going to evolve the app a bit to give the test more substance, and then finally see how much each cost in my token counter,

10:59 right? So I’ll come here and send the prompt, and here too. So what did I ask for now? The notebook will turn into a product for various small markets. Create a product presentation page that can be the first screen for those who arrive. So it’s as if we’re evolving the product into a SaaS, right? It will show what the notebook is. There will be a registration now with login, with email and password. So, there will be authentication. The owner can register the market’s cashier, right, to be able to manage the customers. For each overdue customer, a collect button that will open and send a notice via

11:31 WhatsApp and a month-end closing screen . So, more features for us to push a little harder here and let’s see what happens in the end, folks. While we’re not creating it, a very important notice, talking about payments for the store, right? Appmax is the payment solution I recommend for you to integrate into your system, okay? What happens? They have zero fees on Pix via the link here in the description and in the pinned comment of this video. This will help you receive the full value of the purchase. So, if the guy paid 100,

12:02 you receive 100. In all other payment gateways, you will have a fixed fee or a percentage. And here you have this special condition of zero fees on Pix, which is very interesting, since, for example, at Loja Codar, my platform, it’s basically half Pix/half card. Or even more Pix than card, right? We never have more card than Pix. So it’s a preferred payment method for customers on our platform currently, and it probably is on others too, so you’ll be able to have more cash flow, okay? So just integrate it via API and you’ll explore the documentation. Here is a quickstart to help you integrate

12:36 faster. Put it into the AI, hit Ctrl + A, Ctrl + C here and send it over, and it will be able to do the perfect integration for you. You’ll just have to sign up here at Appmax and put in your token, right? Obviously, but it’s very simple, man. They have everything, right? They have standard payments with various payment methods, recurring billing to charge every month, splits to divide between accounts if you have a partner, anti-fraud, tokens, webhooks , and so on. So, it’s a complete solution for your system’s payments, and if any questions remain, send them a message, they’ll respond right away

13:10 to assist you and instruct you on how to set up payments in your system. Got it? So, let’s go back to our Ponytail. And let’s see what happens in this final round, right? Well, guys, we’ve reached the end of this second round. Look, here in Ponytail it took almost 10 minutes again and without it took 18 , okay? So, almost double the time, right, to work without the pool. Let’s see then if this result is better in the Ponytail version or the other

13:44 version. So, I’m going to create an account here, okay? To see if it works, right? Look, I managed to create the account, register a client, they have a 200 limit. Alright. Then they bought 10 dollars of bread. I noted the purchase, dashboard, closing. So this one is the “no pontail” version, right? Which is the one where we have more legible text , white backgrounds so it doesn’t cut

14:16 off the font. with the lines, so it’s fully functional, okay? Team mode, new client, closing. So it added everything we asked for, okay? Now the componil, which is 3002. Let’s run the test here. So it still kept that aesthetic where the line cuts the letters, right? Look, it worked. There’s a Next issue there. The other one didn’t have it. This one is easy to solve too, right? Just tell it to fix it. I’ll add a new client.

14:47 Look, even the buttons are hard to see like this, right? Born there. So, 200 limit, I registered, look. Registered here, she bought 10 dollars of bread. The history is here, there’s the closing screen, 10 dollars there in the value, dashboard, right, team, and everything else. So, aside from the Next error, everything worked the same. We could say the difference is a matter of layout, right, besides the token savings and everything else, and the simpler code that Pinte likely adopted,

15:21 right? So, man, for me the result was the same here regarding the system, because I could change everything here with layout instructions, okay? Now we’re going to look at the result, okay ? I’ll open a terminal here in the P folder. I’ll run the token counter. This guy here shows what the difference was between the sessions, okay? So we have the “no pontail” and the “componil “ here. And here is a general

15:52 side-by-side summary, it’s easier to read. So if it were via API, right, I would spend 10,100 and almost 6 with. Look at the difference in output tokens . So this is where Pilio shines, right? Because it works on code simplification and that generates savings, much less development time, that’s also cool, delivering faster, fewer lines of code, right? Practically 1/3. Also interesting. This can result in performance, right? The browser reading the system. Ah, code file, fewer files

16:23 to load too, a faster build, there’s a test file that has none, and one less dependency. This also helps the bundle, right, to be smaller. So we had a 42% lower cost here with Ponytail, right? And the time was 38%lower as well with Ponytail. Here are the libraries that each one installed, right? They were actually different. Interesting, right? You can see that Ponytail didn’t install BerkeleyDB, right? Which is an ORM. It probably did all the queries manually, right? So here it became clear that there is indeed savings.

16:57 Okay, guys. So, of course, it takes a lot of testing, right? I don’t like to just start recommending things out of nowhere, but it’s clear here that we have a comparison and Ponytail came out better, right? We also saw the visual result, nothing significantly better with just Cloud without Ponytail, which would be fair, right? Not saving as many tokens and getting the layout right, it probably wouldn’t even reach the same value. We can also be using some new commands that have arrived, which is ponytail dep. Okay? This

17:28 command here will show if something was simplified on purpose, we will have a comment on what was done, right? In other words, it is basically looking for comments in the code to gather these debts and show them to us. So, what is the central idea of this topic here, of this command here? Some things are kept very simple for the project scope, okay? When the project grows, this approach might be bad. So it goes there, notes what it did and what needs

18:02 to be done later, okay? So let’s see if the closing panel summaries load all customers and entries for the market into memory and do the math in JavaScript instead of aggregating in SQL. This works well up to a few thousand lines per market. So it is talking about technical debt here. When the volume grows a lot, it is better to move this to SQL. So it is already commented, right? So, theoretically, we could make this improvement later, okay ? Everything stays recorded. And there is also ponytail audit. This one

18:33 here might not find anything, right? But it serves to perform an audit on projects that are already underway. And then we install Ponytail. What will it do? It will identify points for project improvement, right? The ideal would be to come here, install it on this one and do it, right? But I will leave it with you so you understand that it can be done. Basically, it would convert the natural project, let’s say, the entire project into a Ponteio project, right? That’s the function of the alt command. Let’s see if it finds anything , because even if it was made entirely

19:07 by Ponteio, the AI might have different results, right? So it might have made an error, in quotes, and then it would see in that audio that it could improve , right? It would also be a process of continuous improvement to run this here every once in a while. So, look at that , it found something. So, that’s interesting, right? It went there to understand what the project does, right , and so on. It’s as if it came out of nowhere, right? Then, look, the date comes from the server clock. What is the today function? It gives the date of each payment purchase and calculates due dates and delays. The problem is, it uses the server timezone; almost all hosting runs on UTC. Okay, correction,

19:40 just one line. new date to locale string and it would set the America São Paulo timezone. If you don’t fix it, dates and due dates come out wrong at night, right during the busiest time for the market, see? So it’s anticipating the problem for when we host this project, right? And it provides the solution. So the project I created entirely in component has adjustments we could make to implement these improvements and have a more “ Ponteio-ized” code, right? Let’s put it that way. And we just need to ask via prompt, right? Fix that, let’s go, right? Something like that. It would fix all these problems. So, man, these improvements in this new version are

20:13 really cool, okay? I liked it a lot. I’m going to test more so I can bring results here for you. And if you want, leave it in the comments and I’ll try to bring more videos about Ponteio. Guys, final notices, okay? I’ll leave the repository here for you, probably in the video links in the description. It’s actually very easy to find, it’s a big repository, just type Ponteio there and you’ll find it, okay? And read the documentation or the changelog, just to understand what changed, which were the commands we used, so you understand where it comes from, okay? And if you are thinking about integrating payments

20:45 into your system, consider Appmax, because they have this zero fee on Pix. This helps a lot for those who are starting out, guys! It’s very easy to integrate. AI can already do all of this practically automatically, right? You just need your key that you create when you set up an account here, right? And get it from their system. Got it? Download Fábrica. I’d say it’s our newest resource here. It guides you to create a project by thinking it through first, planning it, and getting it running with SDD from scratch until it’s live. And also how you can use interesting strategies to attract

21:16 customers and users to your app or system, okay? The whole step-by-step, it’s an illustrated ebook, people are praising it a lot, almost 200 pages, all for free for you who are here on the channel, okay? Link in the description. Well, that’s it, guys. In the other Pilio video from months ago, people asked to run it side-by-side to see if it really saved tokens. Here I showed you that it does, right? And with this new version, it’s even more aggressive in token savings. Just a reminder, right? You saw it there, it generates issues with technical debt. This needs to be mapped out, because if

21:48 this project grows, you’re in trouble, right? The interesting thing about this new way of developing is that we deal with varied projects now with few clients, because it’s launching so many things, right? And because it gave people who don’t understand code the power to do this. That is interesting. And the bridge helps, right? Because it works with simple, performant, and fast-to-launch projects. But when it evolves, you have to watch the technical debt so things don’t get left behind, so you don’t get blindsided,

22:19 have your server crash over something silly, or have the project slow down naturally, because the strategies are so aggressive that in the long run some aren’t sustainable, right? So you have to keep an eye on that. Comment below what you thought so I can know your opinions. Also, leave that like, subscribe, it helps me a lot here on the channel and supports my work too. Deal, guys? So, thanks, see you later, I’m out.