← Back to search

Ponytail 5 Makes Claude Code 41% Faster

AI News Today | Julian Goldie Podcast · 2026-10-21 · 12 min
relevance 65 2508 words Episode page ↗ Audio ↗
Show full episode description
Ponytail 5 Makes Claude Code 41% Faster
✨ Episode Outline — click any point to jump to it in the episode
Problem solved
AI coding agents overbuild, writing far more code than a task needs, which inflates tokens, cost, time, and future breakage; the Ponytail 5 skill makes them write only what the job requires.
Benefits
  • 53% less code, 41% faster, 45% fewer tokens, 26% lower cost on average
  • 98% of risky logic ships with a test vs 68% without the skill
  • Free MIT-licensed single-file skill that works across ~20 agents (Claude Code, Codex, Hermes Agent, OpenClaw, Cursor, Gemini, etc.)
  • Review and audit commands explain findings in plain-language steps, no code reading required
  • Live cost band in Claude Code shows tokens, agent time, reuse rate, and test status while working
Use cases
  • Date picker request: without Ponytail agent wrote 325 lines; with Ponytail 5 it reused the existing input plus browser calendar in 10 lines
  • ponytail review caught a renamed field that crashed a newsfeed in an untouched file; found 100% of planted problems vs 87% without
  • ponytail audit on a warehouse stock tracker found 1,200-item orders failing, silently skipped import rows, ignored product-code typos, and untested risky paths, ranked Must/Should/Nice to fix
  • Non-coders can audit an agent-built lead tracker with one command and get a ranked risk list
KPIs / results
  • Benchmark: Claude Code with Opus 5.5, 39 tasks, each run 5 times, with vs without the skill
  • Code 53% less, time 41% faster, tokens 45% fewer, cost 26% lower
  • Planted bugs caught by agent's own tests: 66% with Ponytail 5 vs 46% without
  • Blind comparison vs old Ponytail won 110 to 67; passed all 30 runs on the six hardest safety tasks
Tools / build
  • Ponytail 5 skill (single-prompt file, MIT, ~20 agents)
  • ponytail review command
  • ponytail audit command
  • Ponytail band (v5.3) cost strip in Claude Code
  • Ponytail board local dashboard with session import
0:00 / 0:00
Ponytail 5 The AI skill that cuts code by 53%. So Ponytail makes your AI agent write half the code, it finishes faster, it costs less, and the work it hands back is better tested than before. Those are measured results and you'll hear exactly how they were measured in a second. Add Ponytail 5 to Claude Code and the agent on average writes 53% less code, finishes 41% faster and it uses 45% fewer tokens. Tokens are the little chunks of text an AI reads and writes. And you pay for everyone and the whole job costs 26% less to run. Normally, when something gets cheaper and faster, it gets worse. Ponytail 5, focus on going the other way. Now, these are numbers from them. So test it yourself out. As always, when I've tested it in the past, it is actually pretty good. And it does match the averages when I've tested it personally. You just want to test this out yourself and I just want to be 100% clear on that. With it, 98% of the risky parts of the work come with a test and without it, 68% do as well. So today you'll learn what Ponytail 5 is, what's new in this version, the real numbers behind it, and why it matters even if you've never written a single line of code. Now, a lot of people build their own tools with AI agents, a booking page for your website, a simple tracker for your leads, a dashboard that shows how your content is doing. You don't write the code, the agent writes it for you. But here's the catch that nobody tells you. Every line the agent writes costs you something. It costs tokens or it costs time and every extra line is one more thing that can break later. AI agents love to overbuild. You ask for something small, you get something huge. The Ponytail team showed this with one simple request. They asked an agent to add a date picker to a website. A date picker is that little box when you click and a calendar pops up so you can choose a date. Without Ponytail, the agent built a whole calendar by hand, 325 lines of code. With Ponytail 5, the agent looked around first, it saw the website already had a basic input box and it knew every web browser already has a calendar built in. So it joined those two things together with 10 lines. Same result for the person clicking this button, but 325 fewer lines to pay for. Wait for and fix one day. Picture two builders. You ask both of them to hang one picture on you all. The first builder brings a ladder, a drill and build to your brand new shelf. The second one sees there's already a nail in the wall and just uses that. Ponytail basically turns your agent into the second builder. And the name comes from a person you've probably met at work. The readme actually describes Ponytail perfectly. Long Ponytail, oval glasses, been at the company longer than anyone can remember, and you show him 50 lines of code and he looks at them, says nothing and swaps them for one. Ponytail puts that person inside your AI agent. It's simple under the hood. The whole thing is one prompt, one file instructions called a skill. Everything else in the project just helps load that one file into different AI agents. It's also free, is under the MIT license, and it works with 20 different agents. Pretty much any agent, to be fair, that can use a skill can work with this. So Claude, Code, Codex, Hermes Agent, OpenClaw, Cursor, Gemini, OpenCode, Kimi, Grok, and more. So that's what it is. Now the bigger question, do the numbers hold back? Because a claim like half the code is easy to make and hard to prove. Let me run you through how they tested it. So they use Claude Code with Opus 5.5. That's one of the strongest models you can use right now. Then they ran the same agent twice, one with Ponytail and one without with the same tasks, same model. The only change was the skill and they used 39 tasks. One of them was a real working project with a website front and a system running behind it. And they ran every task five times. So one lucky run couldn't like skew the results. Here's what actually came out. So Code, written on average, was 53% less. Time, 41% faster. Tokens, 45% fewer. Cost, 26% lower. Now if you compare that to the last version, Ponytail 4.13, the old version already cut the code roughly in half down to 52% of the normal amount. So Ponytail 5 brought it down to 47%. Time went from 62% to 59%. And cost dropped a lot more, from 84% of the normal amount down to 74%. So version 5 isn't a small polish. The team actually says it was rebuilt from the ground up. But less code means nothing if the code is worse. So they actually checked quality three different ways. First, with tests. So a test is a small check that makes sure something works the way it should. With Ponytail 5.98% of the risky logic shipped with a test. Without it, 68%. Second, they planted bugs on purpose. Then they checked it. And they checked if the agent's own tests would catch them. With Ponytail 5, the test caught 66% of the planted bugs. Without it, 46%. Third, they ran a blind test. People compared answers from Ponytail 5 and the old Ponytail without knowing which was which. Ponytail 5 won 110 to 67%. On the six hardest safety tasks, the ones where a shortcut could put someone else's data at risk. Well, Ponytail 5 passed all 30 runs. Less code and no less safe. So the agent writes less, finishes faster, costs less, and tests more. Combination is correct. Now, if you're sitting there thinking, this sounds great, but I'm not technical. And I wouldn't even know where to start. That's exactly why the AI Profit Boardroom community exists. We've helped over 3,300 business owners use AI, and some of them had never touched AI before they joined. Inside, you get our agentic operating system. And it's one place where you can plug in all your AI agents like Claude, Hermes, OpenClaude, and a skill like Ponytail can sit across all of them. We've also added tools for videos, SEO agents, AI adverts, and you can build in your own workflows. And you get the Ziphar Affiliate Roam app set up and a video tutorial, plus our daily updates as we improve it. And on our four weekly coaching course, you can bring your own Claude code or Hermes setup and ask live how to make your AI agents build leaner. Cheaper tools that help you get more leads and more customers. Link in the comments description or go to the AIprofitable.com. So what's actually different inside Ponytail 5? Well, there are three big changes. And the third one is the one that most people use every single day. So the first change is about when the agent gets lazy. Ponytail works like a ladder. Before the agent writes anything, it climbs down a list of questions and it stops on the first one that works. For example, does this need to exist at all? Is it already in the project? Can a built-in tool do it? Can the browser or phone already do it? Is there a tool already installed that does it? And can it be one single line? And if only every answer is no, it writes the smallest thing that works, plus one small test if there's any real logic in it. So here's what's new. In Ponytail 5, that ladder only starts after the agent understands the problem. So it reads the code your change touches. It follows how the information actually moves through the project. Then it picks a rung. The readme puts it in a way that's easy to remember. Lazy about the solution, never about reading. The old risk with any write-less tool is that it cuts the wrong things. Ponytail helps and has a clear list of things it will never cut. Checking information coming in from outside. Handling the moments where data could get lost. Just security and accessibility, which means making sure people, for example, who want to use it actually can. And every reply ends the same way. The agent tells you what it skipped, what it didn't check, and any risks you should know about. Now, you might say, okay, fewer lines? That just means the agent is being sloppy. But actually, the rule was never write as many lines as possible or as few lines as possible. The rule is to write only what the job actually needs. The code ends up small because the rest wasn't actually needed. Lower cost and faster speed or a side aspect. Perfect. The second change is the review. So ponytail is a command called ponytail review, and you run it after your agent makes a change. In the old version, the review only looked for code to cut. Now it reviews like the senior person who gets a phone call at 2 a.m. When something breaks, it checks for bugs. It checks safety. It checks what happens under real load when lots of people use the thing at once. It looks for risky code with no tests and slow spots and what can be cut. And here's a big shift. It reads the code around your change, not only the change itself. In the benchmark, someone actually renamed one field, one small label, and a totally different file that nobody touched. The one that makes a website's newsfeed start crashing on every single visit. So review that only looks at the change would never see that. Ponytail 5 actually caught it. The numbers, ponytail's 5 review found 100% of the problems they planted. Without the skill, 87%. So for problems hiding outside the change itself, ponytail found 100%, without it 78%. Each finding is written in plain steps. Now, what does the code do? What goes wrong? How does it fix it? And what happens if you skip the fix? That's what you get inside the review. So you don't need to read code to understand the review. You just need to read four short answers. And the third change is the audit. And this is the one that turns ponytail from a tool for coders into a tool for anyone who has a project, an agent built for them. The review checks one change. The audit checks your whole project. It starts by mapping everything. Where things start. How information moves around. How much use the project is expected to handle. Then it ranks what it finds. Must fix. Should fix. Nice to have. And it tells you what to fix first. In the benchmark, they ran it on a stock tracking project for a warehouse. The audit found that from big office orders with 1,200 items failed completely. It found that the import quietly skipped bad roads without telling on one. And it found that a typo in a product code got silently ignored. And it found that the riskiest parts had no test at all. Then it gave one clear answer. Fix number one first. The old audit only listed things to delete. The new one tells you where your project will hurt you in order. So picture this. You had an agent build a lead tracker six months ago. And you've never really checked it, right? One command. And you've got a ranked list of what could go wrong written so you can understand it. Now, the newest piece is two. This came in version 5.3 and it's called the band. If you use code code, Ponytail now shows a small strip right above where you type. And it shows what the current task is costing you. How many tokens you used. How much time the agent spent. How often it reused the code that already existed instead of writing new code. And whether the test passed after the last change. For most people using AI agents, cost is a mystery until the end. You ask for something. The agent works for a while. And you find out later how much it used. The band shows it whilst it's happening. There are buttons on the band too. So you can change how strict Ponytail is. You can run the review or the audio with one click. And you can copy the task as a note. And you can shrink the band down to one line if it gets in the way. If you don't want it at all, one command actually hides it. So the same command brings it back. And there's the board too. It's a page that opens on your own computer. It shows every task across all your projects. Cost over time. Agent time over time. Every session with whether its tests pass. Even the days you worked in your streak. The example board in the readme shows about 61 sessions. And just over 20 hours of agent time in 30 days. That's the kind of thing that's easy to lose track of without a page like this. By default, everything on the board stays on your machine. Nothing gets sent anywhere. And there's an import command that pulls in your past quad sessions. So your history shows up too. So overall, Ponytail 5 is a good reminder of where AI agents are handling and heading. So the agents can already build almost anything you ask. The real gains now come from how carefully they build. Less code that does the same job. More tests on the parts that matter. And clear answers about what got skipped. And if you want help putting Ponytail 5 to work inside your own business, along with our token minimization playbooks, that we've designed to reduce the amount of tokens you use and build faster, come join us inside the AI Profit Boardroom. You get our agent OS, where Claude Code, Hermes, OpenClaw, and other AI agents plug into one place. So a skill like Ponytail can run across every agent you use, building your lead trackers, your content tools, your SEO agents, and your AI avatar workflows without the bloat. You get the zip file 30-day roadmap for setting up, a full video tutorial, and daily updates as new versions come out. And on our four weekly coaching calls, you can bring your own setup and ask live how to make your AI agents build faster and cheaper so you spend less time fixing tools and more time getting leads and customers. You also get daily tutorials showing you step-by-step how to use updates like this the moment they land. And there are over 3,000 biz zones inside, plenty of them running Claude Code and Hermes Agent every day with a member map so you can find people need. Someone's always online to help 24-7 as well because it's a big community. Link in the comments description or go to theairprofitform.com to get access. Thanks for watching.