Claude Code Never Has To Build From Scratch Again
AAI LABS
Computing/SoftwareSmall Business/StartupsInternet Technology
Transcript
00:00:00AI agents have gotten really good now.
00:00:02Building is the easy part of making something now.
00:00:04The problem is that whenever you ask your agent to do something,
00:00:08it has to start from scratch.
00:00:09In order to fix this problem, we've found two skills,
00:00:12which are basically just files with instructions written in them
00:00:15so that your agent can follow a workflow.
00:00:17So in other words, we've got two workflows
00:00:19that basically allow you to build faster with your AI agent,
00:00:22whether you are using cloud code or codex.
00:00:25And the first skill also reveals a really interesting thing about GitHub,
00:00:28which changed the way that we download stuff from it now.
00:00:31So if you're new to the channel, then welcome.
00:00:33We're a software company and this is AI Labs,
00:00:35where we show you how to optimize your workflows with AI
00:00:38using proven methods from our own team.
00:00:41And in this video, we're going to go over both these skills
00:00:43and look at how they enable you to build faster.
00:00:46Before we actually introduce the skills,
00:00:48we need to take a closer look at the problem.
00:00:50So whenever you get a new idea
00:00:52or you want to build something to solve your problem,
00:00:54you go to your agent, turn on the plan mode
00:00:56and you explain your idea to it.
00:00:58It plans for a while
00:00:59and then it tells you how it's going to build what you want.
00:01:01But there is a huge angle
00:01:03that you and your agent have completely missed.
00:01:05What if what you want to build already exists,
00:01:08whether it's a paid tool or an open source one?
00:01:10Your agent never checks
00:01:11if the thing that you want already exists or not.
00:01:14And this doesn't only happen
00:01:15if you're starting a new project.
00:01:17Let's say that you already have a product
00:01:19that you are working on
00:01:20and you decide that you want to add a new feature in that product.
00:01:23You're going to open up your coding agent
00:01:25and tell it about this new feature,
00:01:27but it's just going to start exploring ways
00:01:28that the problem can be solved.
00:01:30It never looks if that problem is already solved
00:01:32and put out on the internet.
00:01:33It never looks how other competitors have solved that problem.
00:01:36This is a serious problem
00:01:38because you're spending all that time
00:01:39and tokens on a problem that has already been solved.
00:01:42So someone built a skill to solve this exact problem.
00:01:45The skill is called advise project approach.
00:01:48And the first thing it does
00:01:49is that it works out where you actually are.
00:01:51Because telling it about a brand new idea
00:01:53versus telling it about a feature you want to add
00:01:56to something you've already built
00:01:57are two completely different problems
00:01:59and it treats them in different ways.
00:02:01If it's a new idea,
00:02:02it works out how the whole thing should be built.
00:02:04But if you've built something already,
00:02:06it looks at what you've already got
00:02:08and tells you what to keep and what to change.
00:02:10But it doesn't go straight to searching either.
00:02:12Before it looks anything up,
00:02:14it checks whether it even knows enough about what you're doing.
00:02:16If your idea is still vague,
00:02:18it comes back and asks you some questions first.
00:02:20For example, who this is actually for,
00:02:22what does it need to do for that person
00:02:24and what your budget constraints are,
00:02:26can you actually get something paid
00:02:27or do you want to stick to a free option?
00:02:29And if you've already given it all of that,
00:02:31then it just skips the questions entirely.
00:02:33And whatever you don't answer,
00:02:35it doesn't quietly guess either.
00:02:36It writes down what it's assuming
00:02:38so you can see it and tell it where it's wrong.
00:02:40So by the time it actually starts researching,
00:02:42it's looking for something specific.
00:02:44It goes and finds the paid tools that already do this
00:02:47and the open source projects doing the same thing.
00:02:49But if official docs and GitHub
00:02:50aren't going to be enough on their own,
00:02:52it comes back and asks whether it should go into X,
00:02:55Reddit and YouTube as well,
00:02:56where it can get additional information.
00:02:58With the time that we've spent with it,
00:03:00testing it in different places,
00:03:01it doesn't usually need to use those sources
00:03:03as it finds the answer pretty quickly.
00:03:05But there's a trap in looking at what everyone else has built.
00:03:09Your agent goes straight for whatever's most popular.
00:03:11And popular on GitHub basically means
00:03:13whichever project has the most stars,
00:03:15which is a bookmark count that only ever goes up.
00:03:18So it tells you a lot of people liked it at some point.
00:03:20It doesn't tell you it fits what you're doing.
00:03:22And the bigger version of that is the setup itself.
00:03:25The biggest project doing your thing is usually big
00:03:28because there's a whole team behind it
00:03:30and a lot of people using it.
00:03:31So the way they built it is right for them.
00:03:33If your agent just copies that,
00:03:35you get all of their complicated setup
00:03:37and none of the reasons for it.
00:03:38So the skill makes itself write down two things
00:03:41for every project it finds.
00:03:43What actually applies to you
00:03:44and what's only there because of how big they got.
00:03:47And that's the difference between a plan built for your situation
00:03:50and a plan built for somebody else's company.
00:03:52And then there's the money.
00:03:53Let's say that you want to add DMs to your app.
00:03:55You don't have to build that.
00:03:57There are services like Stream and Sendbird
00:03:59that have already built the whole thing
00:04:00and you just plug it into your app.
00:04:02But then look at what it costs.
00:04:04Sendbird's free plan is about 100 users.
00:04:06Stream starts at around $500 a month for 10,000.
00:04:09And past that, on both of them,
00:04:11you're paying for every extra person who opens a chat.
00:04:13It's this cost angle that your agent usually misses.
00:04:16This is why the skill has to check
00:04:18what it costs you at three separate points.
00:04:20First, while you're building it and nobody's using it,
00:04:22then the day real people actually turn up
00:04:25and then once it grows.
00:04:26So you find that out right now instead of six months in.
00:04:28But before we move deeper into this,
00:04:30it would be great if you subscribe to the channel
00:04:32and hit the hype button.
00:04:34This small gesture of support goes a long way for us.
00:04:36So before we jump in and see how the skill actually works,
00:04:39you need to understand the problem
00:04:41that we needed to solve here.
00:04:42Skill updates only go one way.
00:04:44We change something and the person who downloaded it
00:04:46has no idea.
00:04:47Their copy just stays the way it was.
00:04:49So we needed something of our own
00:04:51where we could publish once
00:04:52and then everyone who's part of our community
00:04:54gets access to the updated version.
00:04:57The update would be sent out to our own team first
00:04:59and then to the whole community as well.
00:05:01This problem was actually highlighted
00:05:03by someone in the community
00:05:04and their concern was actually valid here.
00:05:06So to install the skill,
00:05:08you just need the link to the GitHub repo,
00:05:10which we're going to paste in the description below.
00:05:12And you can go inside whatever agent you're using,
00:05:15paste that link in and tell it to install the skill.
00:05:17We're using Claude code here.
00:05:18So if I look inside the .claude folder
00:05:20and go inside skills,
00:05:22you're going to see that the advise project approach skill
00:05:24is sitting right there.
00:05:25After installing it,
00:05:26we referenced the skill and told it what we wanted.
00:05:29We wanted to build a web app
00:05:30in which we could drop our skills
00:05:32and anyone who's downloaded them onto their laptop
00:05:34automatically gets those updates
00:05:36without having to worry about
00:05:38whether the one they've got is still the current one.
00:05:40And so it started researching.
00:05:42It started doing some minor research
00:05:43and after that,
00:05:44it realized that it needed to ask us some questions.
00:05:47And then first of all,
00:05:48it asked us who it was for.
00:05:49It gave us numerous options
00:05:51and it gave us good advice
00:05:52that before we actually implemented it
00:05:54for the full community,
00:05:55we could build it for our own team
00:05:57because our own team does use internal skills as well.
00:06:00So this is why we told it
00:06:02that we needed to build it for the team right now.
00:06:04And then another thing that made it more complex
00:06:06was which agents this project was supposed to be for.
00:06:09So if it was only supposed to be for Claude Code,
00:06:11it had found out that the solution was already there
00:06:14because Claude Code has something called plugins
00:06:16and those would get updated automatically.
00:06:18But since a lot of our audience
00:06:20and we ourselves use Codex as well,
00:06:22this is why we included other agents as well,
00:06:24which complicated things.
00:06:25And then it asked whether the thing that we were building
00:06:28was actually going to be used to edit the skills.
00:06:30And we clarified that no,
00:06:32it was only supposed to store the skills
00:06:33and send updates to the people
00:06:35who had those skills in the first place.
00:06:37And it basically did more research
00:06:39and then it confirmed whether we wanted the updates
00:06:41to be automatically pulled to the device.
00:06:43And we clarified that yes,
00:06:45pushing to the app should be manual
00:06:46whenever we felt that the skills needed an update,
00:06:49but anyone who had those skills
00:06:50should automatically get them.
00:06:52And it made this docs folder
00:06:53and inside that docs folder,
00:06:55it was actually writing that project approach
00:06:58But the approach that it took had some flaws
00:07:00that it couldn't figure out how to fix them on its own,
00:07:03which is why we needed to use another skill.
00:07:05But before we look at that second skill,
00:07:07let's have a word by our sponsor.
00:07:08Lovable.
00:07:09Check this self-checkout e-commerce store
00:07:11for watches we made with just a single prompt
00:07:13and after approving a plan on Lovable.
00:07:15We started with explaining our app
00:07:17with a prompt in plan mode.
00:07:19First, it asked further questions
00:07:20about the app we're building.
00:07:21Then it proposed a plan
00:07:22where the whole app's details
00:07:24and the decisions it made are verified.
00:07:25And after some time, we got our store's first view.
00:07:28Lovable also provides its MCP server,
00:07:31which can be connected to any AI agent
00:07:33and you can build using cloud code, cursor, or codex.
00:07:36Payments are usually where AI builders stall,
00:07:39but Lovable sets it up and handles the global tax,
00:07:41letting the developer test safely
00:07:43before the app goes live.
00:07:45Therefore, you get from an idea to a live app
00:07:47that takes real orders from customers.
00:07:49So start building with Lovable
00:07:51using the link in the description below.
00:07:53Before we move on to that skill,
00:07:55let me explain why we even need a second skill
00:07:57when the first one just handed us the whole plan.
00:07:59What the first skill came back with
00:08:01was the skills CLI.
00:08:03That's a tool you run in your terminal
00:08:04to install skills into your agent
00:08:06and it had figured out a lot of the parts
00:08:08that we wanted here.
00:08:09It also told us that we didn't need
00:08:10to make a web app for it
00:08:11and there was a better approach to this.
00:08:13It said that for now,
00:08:14while we perfected the app
00:08:16and made it just for the team,
00:08:17we didn't need to build something for the storage.
00:08:19The skills could be stored on GitHub for now.
00:08:21When we would launch it for the community,
00:08:23we would look at the storage problem again.
00:08:25As we already said,
00:08:26most of what you want to build
00:08:27has already been built by somebody else
00:08:29and it's sitting on GitHub
00:08:30and that's exactly what the first skill
00:08:32is good at going and finding.
00:08:34But when the thing is experimental
00:08:35and nobody has shipped it yet,
00:08:37you get stuck.
00:08:38But that doesn't mean nobody has worked on it though.
00:08:40The concept might already be there.
00:08:42And a small part of our app had that problem.
00:08:44What we're actually building
00:08:45is something that automatically pushes instructions
00:08:48onto other people's laptops
00:08:49and the first skill flagged that
00:08:51as the part to be careful with.
00:08:52Then it stopped
00:08:53because there was nothing out there
00:08:54for it to point us at.
00:08:56So that's where the second skill comes in.
00:08:57It's called Neuro Archive.
00:08:59It helps by searching a platform called Archive.
00:09:01It's where actual researchers publish their work.
00:09:04People at universities and labs
00:09:05whose entire job is to spend months,
00:09:08sometimes years on one specific problem
00:09:10and then write down exactly what they found
00:09:12and what didn't work.
00:09:13Other researchers then read that and pull it apart.
00:09:16And all of it is free and public.
00:09:17Almost everything you're building with right now
00:09:19started there.
00:09:20The paper behind the technology
00:09:22that made ChatGPT possible
00:09:23went up on Archive years before there was a product.
00:09:26So when you're building something experimental,
00:09:28that's where the concept is already sitting.
00:09:30The way the skill runs
00:09:31is that it takes your problem
00:09:32and picks the research areas that actually match it.
00:09:35Then it pulls the papers
00:09:36and the way it reads them is the part that matters.
00:09:39It doesn't read them all in one go.
00:09:40Every paper gets its own sub-agent
00:09:42which is basically another agent
00:09:44that your main agent can start
00:09:46with a fresh context window
00:09:47and the main agent can give it a prompt by itself
00:09:50so you don't need to do anything.
00:09:51That's one reader per paper
00:09:53and none of them can see
00:09:54what the others are looking at
00:09:55and after reading all of them,
00:09:57it commits to one answer.
00:09:58So the problem that this skill solved
00:10:00was that we were pushing files
00:10:01onto people's laptops automatically
00:10:03and we had no idea
00:10:04what security issues this was going to cause.
00:10:06It pulled 12 papers
00:10:08and then it went back and forth
00:10:09over which one actually applied to us
00:10:11and it landed on one.
00:10:12A paper about package systems
00:10:14and it implemented the technique
00:10:16that the paper talked about,
00:10:17tested it and made sure that it was working.
00:10:19So just like the previous skill,
00:10:21we gave it a link
00:10:22and it installed it in our skills folder
00:10:24as you can see right here.
00:10:25Then it loaded the skill
00:10:26and as you can see,
00:10:27it spawned sub-agents
00:10:28and they all read each paper one by one.
00:10:31And after implementing the fix,
00:10:32we ended up with a working version of our tool.
00:10:35It's a CLI tool
00:10:36which just means it runs inside the terminal.
00:10:38And since these coding agents
00:10:40already run in the terminal themselves,
00:10:42they can use it directly
00:10:43and do the whole thing on their own.
00:10:45So here's how the thing actually works.
00:10:46On our side, it's two commands.
00:10:48You point it at a skill you've already got
00:10:50and then you publish it.
00:10:51You can actually tell Claude Code to publish it
00:10:53and it would be able to help you with that.
00:10:55And everyone else's agent
00:10:56pulls it down in the background
00:10:58whenever they reopen their agent.
00:11:00Everyone just needs to install it on their laptop
00:11:02and it automatically attaches to Claude Code and Codex
00:11:05and just does the work.
00:11:06So right now we're using both of these separately
00:11:09and we'll link both of them down below.
00:11:11But running them separately has a problem.
00:11:13The second one is expensive to run
00:11:14and most of what you're building doesn't need it.
00:11:16So you don't run it.
00:11:17And then the one time you actually needed it
00:11:19is the time you didn't bother.
00:11:20So we've built a separate skill named Head Start
00:11:22which basically combines the two.
00:11:24It runs the first one on everything
00:11:26and it only pulls in the second one
00:11:27when it hits something that's actually new.
00:11:29And in our testing that's worked really well.
00:11:31That one's in our community AI Labs Pro.
00:11:34That's also where you'll find other skills and resources
00:11:37that we show you in our videos
00:11:38along with a place to interact with our team.
00:11:41So if you found value in what we do
00:11:43and want to support the channel
00:11:44this is the best way to do it.
00:11:46The link's in the description.
00:11:47That brings us to the end of this video.
00:11:49If you'd like to support the channel
00:11:50and help us keep making videos like this
00:11:53you can do so by using the super thanks button below.
00:11:55As always, thank you for watching
00:11:57and I'll see you in the next one.
00:11:59Thank you.