13:39Ben AI · Tutorial
Anthropic's 44-Skill Small Business Plugin, Ranked
Ben walks through Anthropic's updated Small Business plugin, now 44 skills and 35 connectors deep, and ranks the ten actually worth installing.
September 17thThe argument in one line.
Opus 5.5 is now good enough at reverse engineering behavior that showing Claude your best and worst finished examples builds a better skill than writing out the process step by step.
Read if. Skip if.
READ IF YOU ARE…
- You already give Claude detailed written instructions for a repeat task and want a faster, often better-performing alternative.
- You have past output you can rank by performance, like emails, posts, or proposals, and want Claude to find what actually worked.
- You're building or maintaining a library of Claude skills and want a sharper process than writing out every step by hand.
SKIP IF…
- You're looking for a walkthrough of the REA GitHub repo itself as a coding tool; this video is about applying its underlying idea to prompts and skills, not a repo tutorial.
- Your task has no clean finished output to point at, like an open-ended triage or decision process; the video says this method doesn't fit that case.
The full version, fast.
The viral Reverse Engineer Anything repo gives Claude agents the tools security researchers use: point it at an app and it studies the app's behavior, not its code, then rebuilds and tests a matching version. The real story is that Opus 5.5 does this kind of reverse engineering well even without the repo, which changes how prompting and skill-building should work. Instead of describing a task in detail, show Claude a finished example of what good looks like and let it infer the process. The author proves this by rebuilding his newsletter skill from his six best and six worst performing emails: comparing hook, structure, CTA placement, and tone produced a sharper skill than the one he originally wrote by hand.
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 – 00:40
01 · Intro
The REA repo passed 200,000 GitHub stars in a week; Ben previews the video and the free skill he's giving away that automates the whole reverse-engineering process.
00:40 – 02:06
02 · What Is the REA Repo?
REA is an MCP plus a Claude Code/Codex skill that gives an agent the inspection tools security researchers use: point it at an app, it studies behavior (not code), then an agent rebuilds and tests until it matches.
02:06 – 03:20
03 · Why It Matters
People have used the approach to clone Adobe apps, Whisperflow, and Screen Studio, and to rebuild entire tech stacks; the real point is that Opus 5.5 reverse engineers well on its own, which is why showing an example now beats describing a task.
03:20 – 05:11
04 · Applying It to Skills
A skill is a handbook Claude follows every run; feeding it six best- and six worst-performing newsletters, ranked by open and click rates, produced a sharper skill than the one Ben originally wrote by describing the process.
05:11 – 06:03
05 · Step #1: Gather Examples
Gather real performance data, best and worst, when you have it; with no data of your own, borrow a few strong examples from a creator you admire instead.
06:03 – 06:57
06 · Step #2: Find the Winning Patterns
Ask Claude to compare structure, hook, CTA, tone, length, and subject line, tag every finding as seen, inferred, or unknown, and let it isolate what actually drove performance.
06:57 – 08:19
07 · Step #3: Build the Skill, Then Test It
Only after approving the findings does Claude build the skill; the result shortened the newsletter, moved the CTA earlier, and was tested by repurposing a past video into a reel in the shown style.
08:19 – 09:23
08 · When It Doesn't Work
The method generalizes to anything with one clear finished output (proposals, reports, decks, posts) but breaks down on processes with no clean output, like email triage, where recording the task once is the fallback.
Lines worth screenshotting.
- Reverse engineering now works on your own work, not just other people's software: show Claude a finished example and it can infer the process that produced it.
- Anthropic's own documentation states that examples convey the desired style and detail more clearly to Claude than descriptions alone.
- A newsletter skill built from 6 best and 6 worst performing emails outperformed the same creator's original hand-described skill.
- In one test, open rates were nearly identical between best and worst performing emails, but click-through rates diverged sharply, pointing to body copy and CTA placement as the real lever.
- Tagging every pattern finding as seen, inferred, or unknown stops the model from inventing reasons for why something worked.
- The reverse-engineered newsletter skill shortened the email, moved the CTA earlier, and gated extra detail behind the link.
- Without real performance data, borrowing three examples you personally admire works as a substitute training set.
- The method only works on tasks with a clear finished output; a process without one, like email triage, still has to be built by doing it once with Claude or recording the screen.
- The underlying shift isn't a tool, it's a mindset: start every AI task by picturing the finished output, then let the model reverse engineer the steps to get there.
- People have used this reverse-engineering approach to clone Photoshop, Illustrator, and After Effects features, and to rebuild tools like Whisperflow and Screen Studio.
Show Claude the output, skip the instructions
SKILL-BUILDING METHOD
Opus 5.5 is now good enough that showing Claude your best and worst finished work builds a sharper skill than writing out the process by hand.
02What Is the REA Repo?
- REA is an MCP plus a Claude Code/Codex skill that gives an AI agent the same inspection tools security researchers use to study a target app.
- Point REA at an app and it reports back how a feature actually behaves, evidence included, before any code gets written.
- An agent then writes its own implementation and tests it until the behavior matches, without ever copying the original code.
03Why It Matters
- People have used this approach to rebuild pieces of the Adobe suite, clone tools like Whisperflow and Screen Studio, and in one case rebuild an entire software stack.
- Rebuilding a whole stack trades a subscription bill for ongoing maintenance and bug-fixing time, which is often the worse deal.
- The real shift isn't the repo, it's that Opus 5.5 reverse engineers behavior well on its own, repo or not.
- Anthropic's own documentation states examples convey desired style and detail more clearly to Claude than descriptions alone, which is why showing beats describing.
04Applying It to Skills
- A skill is a handbook Claude follows every time a task runs, so the fastest way to build one is with great finished examples instead of a written-out process.
- Feeding Claude six best-performing and six worst-performing newsletters, ranked by open and click rates, let it infer what the top ones were doing right.
- The reverse-engineered newsletter skill beat the original hand-described one, built faster and closer to the author's actual voice and winning pattern.
05Step #1: Gather Examples
- Gather actual performance data when you have it, your top-performing outputs AND your worst-performing ones, so Claude can compare rather than guess.
- With no data of your own, borrow from someone whose work in your niche you already admire; three strong examples is enough to start.
06Step #2: Find the Winning Patterns
- Ask Claude explicitly to compare structure, hook, CTA, tone, length, and subject line, and to explain why each example likely worked or didn't.
- Require every finding to be tagged seen, inferred, or unknown; that one instruction stops the model from inventing plausible-sounding reasons.
- In the newsletter test, open rates were nearly identical between best and worst performers, but click-through rates differed sharply, isolating body copy and CTA placement as the actual lever.
07Step #3: Build the Skill, Then Test It
- Only after reviewing and approving the findings should Claude build the skill; it wrote the entire newsletter skill without further process instructions.
- The resulting skill shortened the newsletter, moved the CTA earlier, and gated extra detail behind the link, all patterns it found in the winning examples.
- Test a new skill by feeding it a fresh input, a thought, an idea, or an old video, and checking the output against the patterns it was supposed to learn.
08When It Doesn't Work
- The method generalizes to anything with one clear finished output, a proposal, report, deck, or post: show Claude your three best and let it reverse engineer the rest.
- It breaks down on processes without a clean output, like an email triage workflow; for those, still build the skill by doing the task once with Claude, or by recording the screen and voice while doing it.
- The underlying lesson outlasts any specific repo or tool: think about the output first, since the skill that matters most going forward is knowing what good actually looks like.
Terms worth knowing.
- REA (Reverse Engineer Anything)
- A viral open-source MCP and Claude Code/Codex skill that lets an AI agent inspect a target app's behavior and rebuild matching functionality without copying its code.
- MCP (Model Context Protocol)
- The standard that lets an AI agent connect to external tools, in this case the toolkit REA uses to inspect and trace a target app.
- Skill
- A saved, repeatable process definition that a Claude agent follows every time it runs a given task.
- Seen / inferred / unknown tagging
- A labeling convention that forces a model to mark whether a pattern claim is directly observed, a guess, or unsupported, which prevents it from inventing reasoning.
Things they pointed at.
Lines you could clip.
00:00
“The reverse engineer anything GitHub repository got more than 200,000 stars in one week.”
01:30
“The key is it doesn't copy the actual code. It just studies how the app behaves and then rebuilds it.”
03:18
“The real takeaway here is that Opus 5.5 is just becoming really good at reverse engineering things, and we don't really need this repo for it at all.”
05:30
“I gave it six of my best performing newsletters and six of my worst performing newsletters.”
06:40
“Open rates are very similar for my best performing and worst performing emails, but click-through rates are very different.”
07:20
“It created the entire newsletter skill without me giving any instructions on what process to follow to get to a good outcome.”
09:18
“The better these models get, the more the skill that matters is knowing really what good looks like.”
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.
metaphor
The reverse engineer anything GitHub repository got more than 200 ,000 stars in one week. And the entire internet has been using it with Opus 5 .5 to reverse engineer apps. But the same principle can also be applied to automate your work far faster and better.
So in this video, I'll explain what the reverse engineering repo is, exactly how to apply it to your cloud prompts and even skills to automate work far faster, better, and easier, and give away a skill that does this entire process for you. Now, if you don't know... me yet i'm ben i'm a three -time founder and now run an ai agency and one of the largest ai communities for professionals and business owners where i teach how to use ai to actually automate work or build an ai business so what is this reverse engineer anything repo that's been going absolutely viral it's essentially just an mcp and a skill for ai agents in cloud code or codex that give them the same tools that security researchers use to look inside softwares to reverse engineer a feature or an app now you point it basically at an app and it figures out how the app or a feature works and then reports back what it found.
And then an AI agent can write its own version and test it until it behaves the same way. And the key is it doesn't copy the actual code. It just studies how the app behaves and then rebuilds it.
Now, reverse engineering has existed, of course, for decades. And people have been doing this with AI for a while already. But the bigger point here is that Opus 5 .5 is just becoming very good at doing this, even without this repo.
So people have been using this repo or even just Clods, Opus 5 .5. to rebuild the entire Adobe suite with Photoshop, Illustrator, After Effects, etc. People have also been cloning tools like Whisperflow or Screen Studio, rebuilding existing apps with only the features they need or even rebuilding their entire tech stack like Peter Levels.
Now, I don't necessarily recommend entirely rebuilding your software stack because there is still such a thing as maintenance and bug fixing, which is often just not worth the time invested for the price you pay for these software tools. Again, the real takeaway here is that Opus 5 .5 is just becoming really good at reverse engineering things.
And we don't really need this repo for it at all. Now, you might be thinking, I'm not building apps. So why does this matter to me?
It's because the same principle doesn't just apply to building apps. It can also be used to automate work far faster and better. Now, I strongly believe that with these new frontier models, the mindset shift that we sort of have to make is that instead of trying to describe a task or a good output, you just want to show Cloud a good example or output that you...
already have then let it go off and reverse engineer what to do to get to an outcome like that again for example instead of describing how you should do your accounting show it your last month finished accounting sheet and let it figure out by itself exactly how to do this for the new month now many people are still used to describing in detail what these models should do because we're used to worst performing models but entropic themselves say examples convey the desired style and level of detail to cloud more clearly than descriptions alone and with these models becoming better and better showing what good looks like is often all we need because these models can figure out and reverse engineer how to get to that outcome by themselves so it's a good principle to keep in mind when using these newer frontier models for any task really is start with the output in mind and this principle doesn't just apply to your one -off tasks it also applies to your repetitive tasks and to skill building now if skills are still new to you that clots
follows every time you run it so essentially if you can instead of describing the steps and the process a skill you're building should follow just start with some great examples of the outcome and let a model like opus 5 .5 reverse engineer the skill and the entire process from there itself because you'll see it's often faster and the skills come out better for example i applied this principle to redo my newsletter skill i gave it six of my best performing newsletters and six of my worst performing newsletters i let it identify patterns in the structure the copy and call to action of the best performing ones and then i let opus 5 .5 reverse engineer a skill and a process for high performing newsletter generation and it actually led to a far better newsletter skill than the original one that i built by just describing the process and giving some examples so a good principle to keep in mind with these new models for building skills is that great examples can really lead to great skills so in a second i'll show you an easy three -step process to build better skills through this
method really fast but before i show you very quickly if you enjoy this type of content you'll really enjoy my ai accelerator where you'll get access to my 30 -day cloud course that takes you from beginner to pro in 30 days and in less than one hour a day it teaches everything you need to know and nothing you don't in order to use cloud to actually automate work on top of that you'll also get access to a full course on growing an ai business from getting your first clients to scaling based on everything me and my team have learned over the last three years and you also get access to unlimited one -on -one live ai consulting with AI experts.
And besides that, we have a big library of skills, templates, and resources across all business departments to give you a head start. So definitely check it out. If you want to dive deeper into AI, the link is in the description below.
And if you're a business looking to partner with my AI agency, you can also book a free call with us in the description below. Now back to the video. So the three -step process to reverse engineer good skills I recommend is the following.
First, you want to gather good output examples or even let Cloud do that for you. So if you have actual data, on your output examples, you want to, of course, gather your top performing ones, but also your worst performing ones.
So Cloud can actually identify performance patterns between those. For example, in my case, I gave Cloud my six best performing and six worst performing newsletters ranked by open rates and click rates. Now, if you don't have existing examples, of course, you can also take inspiration from other people.
For example, you can gather the top performing posts from another creator that's in your niche that you like, or if you have a use case where you don't even have data. Of course, you can also just reverse engineer anything you actually like. For example, here for a skill that I wanted to build to create YouTube shorts and Instagram reels, I just gave it three example reels with a style that I liked.
Then the second step is to let Clod identify the winning patterns. Essentially, let Clod figure out why these outputs seem to perform better. And of course, Clod can do this even better if it also has access to those bad performing output examples.
So in my newsletter example, I just told Clod, first, compare them, find the strong... structure, hook, CTAs, tone, length, and subject line patterns and say why each one likely worked or didn't. Then tag every finding as seen, inferred, or unknown.
Now, this last line is key because it prevents Cloud from inventing reasons and lets you sort of check what it's actually guessing. Clot will then compare these, identify patterns and make assumptions about what factors could lead to better performing outputs.
For example, it found that open rates are very similar for my best performing and worst performing emails, but click -through rates are very different. So the body and the call to action are really what matters for my high -performing newsletters. And after reviewing then, you'll get to the last step, which is actually telling Clot to build the skill by reverse engineering these winning patterns.
And then, of course, you want to test. So after I reviewed the patterns, I just told it, now build the skill. It created the entire newsletter skill without me giving any instructions on what process to follow to get to a good outcome.
Then with the newsletter skill, I can just throw in either a thought or an idea that I have or a past YouTube video. And the skill produces a newsletter that is honestly really in my tone of voice, follows the winning patterns and is far better than my original newsletter skill, which actually took me a while to build because I had to describe the entire process to go through to get to a new.
For example, my new skill made the newsletter far shorter, put the CTA earlier, it gatecapped some information for inside of the link, which are all patterns that seemed to lead to better performing newsletters. Now for my Instagram and YouTube short reels skill, I tested it by repurposing one of my YouTube videos into a reel in the style that I gave it.
And as you can see, it's really impressive. It follows the reverse engineered style really well. And a nice thing with this, of course, is that you only have to find these good outputs once and then that's it.
can replicate and automate it forever. If you don't write newsletters, the same works for anything really with a clear finished output. For example, let's say you send a proposal to every new client, you can show Clod your three best ones and let it reverse engineer the skill from there.
Now, of course, we can't apply this to everything because some things don't have really clear outputs, like for example, an email triads process. And this is where I still recommend building your skills by doing the task once with Clod and then turning it into a skill or by you. using the record, a skill featuring cloud that records your screen, your voice, and all the clicks you make, and then turns that into a skill, which follows a very similar process in the end of .
But really the main takeaway here is that with these models getting better, you want to think output or outcome first and reverse engineer from there. And I think that's really where this is all going. More and more, the better these models get, the more the skill that matters is knowing really what good looks like.
Now I've turned this entire three -step process into a free skill that you can use to build your own reverse engineered skills. You just give it your best and worst examples. It finds the winning patterns and waits for your approval and then builds the skill and tests it against your own examples.
You can download it for free in the first link in the description below. As always, thanks for watching. If you enjoyed this video, I highly appreciate a like and a subscribe.
And if you want to learn more about building cloth skills, you can check out the video here.
The Hook
The bait, then the rug-pull.
A repo that let strangers clone Photoshop's layer engine in a weekend didn't stay a coding trick for long. Ben Van Sprundel walks through why the same instinct, point Claude at a finished example and let it infer the process, now beats writing instructions by hand, and proves it by rebuilding his own newsletter skill from twelve ranked emails.
Named ideas worth stealing.
05:21list
The 3-step skill reverse-engineering process
- Gather your best and worst output examples
- Let Claude find the winning patterns
- Build the skill, then test it
The repeatable process Ben uses to turn ranked past output into a working Claude skill instead of writing one by describing the steps.
Steal forany repeat Claude skill built from Joe's own ranked output, like email P.S. copy, VOD promos, or breakdown takeaways.
06:38concept
Seen / inferred / unknown pattern tagging
A rule that forces Claude to label every pattern claim by confidence level, instead of presenting guesses as fact.
Steal forany pattern-analysis prompt where Joe needs to trust which findings are real versus the model's best guess.
How they asked for the click.
VERBAL ASK
08:57link
“Download the free skill in the first link in the description below”
Single clear link dropped at the very end after the method is fully delivered; the paid AI Accelerator pitch is a separate mid-roll ask around 4:23 with its own return line back to the video.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
- 🚀 [Join my AI Accelerator] Get all my courses, resources & 1-on-1 help ↗
- 🧠 My FREE Second Brain app ↗
- 📘 Step-by-step playbooks to land your first AI client ↗
- 🤝 Book a free call with my [AI agency] ↗
- 💼 We're hiring. Join the team ↗
- 👨💼 LinkedIn ↗
- REA (Reverse Engineer Anything) on GitHub ↗
- Anthropic's skill best practices ↗
- 📺 My video on building Claude skills with Record a skill ↗
Visual structure at a glance.
