Transcription
[Music] [Applause]
Hi, welcome to another video. So, I've been seeing a lot of talk about spec-driven AI coding. Basically, you make a ton of documents before starting a project. So, let's say that you are working on a project, maybe it's a new one, and you start from like zero files. You can use things like Speckit or OpenSpec and whatnot to build a specification of what you want to build with a ton of tasks and a bunch of markdown files.
Now, let's say that you are building a movie tracker app and you want it to be a mobile app. That means you'll go to something like Specit or OpenSpec and it will plan out the project for you. Like it'll create a markdown file where it tells you that you should use Expo or Flutter or whatever. You'll need to use the TMDB API for fetching movies, O0 for authentication, and a bunch of other stuff accordingly. It makes a ton of documentation and basically builds a huge prompt with every minor detail for the AI to follow through.
And well, it sounds great in theory and it works well for some, but it's never really worked for me. I know coding and I don't really need something doing literally all the work for me. I also think none of these tools are the best at it either because if you don't know the difference between Expo and Flutter, then you wouldn't really know what to use anyway. You basically just go with whatever the AI gives you, and I never get good results that way, especially if you're making something you actually want to deploy to production.
Secondly, current models never really oneshot anything with this approach. I really like models that can oneshot things without errors, even if that means the functionality is limited. I think with current agents, we're already in a pretty good place with what they can handle. It just doesn't make sense for me. Or at least I wouldn't recommend anyone use things like Speckit or whatever for something you want to be production-ready or real. It's better to learn coding yourself then mash your head against these things that aren't really worth it.
Now, I do understand that some people still like this stuff and maybe want to experiment with it. So, I just wanted to cover something I've talked about in the past, something that I can actually recommend and that's worked well for me when scaffolding an app from scratch. And well, that's the age-old Taskmaster. It started out basically as a to-do app and stuff, but it recently became really good. I do use it to build out apps when I'm doing something new and from scratch. Remember, I use it after initializing the project.
This one is more focused toward developers like me who just want to increase the performance of their coder rather than trying to replace something like vibe coding which most of the others seem to aim for to make you feel like a real coder. That's kind of misleading. And honestly, the vibe coders of the world won't use CLI tools and clawed code anyway. That person would be better off with something like Lovable. Anyway, Taskmaster used to be a CLI and your AI coder could use that, but now it's an MCP and it's pretty great.
All you have to do is copy this string and either place in the Anthropic API key and Perplexity API key or just place it in the OpenRouter API and it will use that. This basically does two things. It first creates a PRD or project requirement document that details everything we might need to do. Then it parses that PRD to create smaller tasks that are given to your AI coder as a to-do list and then you're good to go. It can also do research on topics your AI might not know about like newer libraries and stuff.
This generally takes only two or three MCP calls and that's it. It creates a PRD, then goes ahead and parses that PRD into different task files. You can also get it to break plans into even smaller tasks. It can plan out the complexity of each task, their dependencies on one another, and so on and so forth. If you use Cloud Code, then you won't need to give it anything. It'll just work. It's pretty good and easy to use and works majorly without any big issues. It's also pretty fast as an MCP. It's pretty quick and the MCPs are pretty good in not overflowing the context with crap like what a lot of other stuff can do for you. It also adheres more to the prompt or task that you give and doesn't wander off. So that is pretty great.
It is the only great spec-driven development thing that I can recommend people to use. Plus, this was made even before the to-do lists was a thing in Claude Code. It's pretty simple to use. You just copy the MCP config and then paste it into what you want to use. Like if I want to use Kilo, I just go to the MCP settings and paste it in there. Now I can go ahead and start using it.
This is a benchmark that I'm working on and it works, but I want to have a UI for it. So, I can ask it to build me a plan for how it would do that with the Taskmaster MCP, and it will go ahead and invoke the Taskmaster MCP. Then, the MCP will initialize the project and get onto the stuff that we want it to do. It will create the tasks, parse them, and create subtasks, and so on. I won't get into too much detail about how it all works. You can actually check out my Taskmaster 2 video for an in-depth walkthrough as it's very similar to what I covered there.
It's really good when you use it with things like GLM, which can be finicky in planning as it's not great at oneshot planning. But here, with the help of this MCP and multi-stage planning, you can get pretty good results. It's a solid option if you're looking for something like spec-driven development that's more mature and not just focused on the vibe coder crowd. Plus, it plugs in seamlessly with whatever coder you want to use. And if you think your context might get bloated with the MCP, there are other options like using a terminal CLI, which is great if you'd prefer not to use MCPs. So that's also doable.
I think this is the best spec-driven development setup that I can recommend to other people. I just don't get things like Speckit and OpenSpec because they're not very good for the smaller tasks that I do. That's majorly what I wanted to talk about. Overall, it's pretty cool. Anyway, share your thoughts below and subscribe to the channel. You can also donate via Superthanks option or join the channel as well and get some perks. I'll see you in the next video. Bye. [Music]