This Is The Only Claude Code Plugin You'll EVER Need (Superpowers)

BBetter Stack
컴퓨터/소프트웨어창업/스타트업AI/미래기술

Transcript

00:00:00This is Superpowers, the agentic skills framework with more than 50,000 stars on GitHub
00:00:05that stops your coding agent from rushing and making mistakes by forcing it through a structured
00:00:10workflow that contains brainstorming, implementation, red-green test-driven development,
00:00:15and even using sub-agents to execute tasks and review code in parallel.
00:00:19But what makes this different from just using plan mode or something like spec-driven development?
00:00:24Hit subscribe and let's get into it.
00:00:27So this is a project called XDL, which is a CLI tool used to download videos from Twitter.
00:00:32I want to create a web UI for this tool that not only downloads videos to the user's browser from
00:00:38a URL, but also uses it to create an article. So I used Opus 4.6 with plan mode in clause code
00:00:46to try to create that UI. And this is what it did.
00:00:50Now, unfortunately, when I ran the code the first time, there are some issues.
00:00:53So I went through with the code to try and fix those issues, which hopefully it's resolved.
00:00:58So now it looks like everything is working.
00:01:00I'm going to go to Twitter and copy the link to this tweet.
00:01:03Sorry, Kevin, paste it here and see if it'll download the video.
00:01:06So it's extracting and it's been able to download it.
00:01:09So if I click here, it opens it in a new tab, which is not what I expected, but hey, it does the thing.
00:01:15And if I try to create an article from that tweet, so I'll paste it in here and it generates,
00:01:20we can see the different steps it's taking. So it's first extracting the video and then
00:01:24it's doing the audio. And now that it's done, we get the article in perfectly formatted Markdown,
00:01:29which is really impressive. And if we take a look at the exact same task done with Opus 4.6,
00:01:35but using superpowers, we can already see that the design is much better with an option to
00:01:39download and generate an article. We'll paste in the same Twitter URL, hit download, and again,
00:01:45it extracts the video, but this time it downloads to my browser and I can click on it to view the
00:01:51video. Now, if you were to use the same tweet to generate the article here, we can see the steps
00:01:56that it's taken. And now that it's finished all the steps, it's streaming the article to my browser.
00:02:01And I have the option to copy this and paste it wherever I want. Now already you can see the
00:02:06superpowers version is better than the version without using superpowers. We'll go through the
00:02:10code a bit later on in the video, but first let's go through how I did this. So after you
00:02:15install the superpowers plugin, we should now have some new slash commands like brainstorm,
00:02:20execute plan, and write the plan, which will write the plan from the brainstorm. Now we only actually
00:02:25need this command since superpowers will direct us to the next stage at the end of a previous one.
00:02:31So let's run this and we'll paste in a prompt. We're going to hit enter. And now it loads the
00:02:35brainstorming skill and begins to understand the current project. And here it's going to ask me some
00:02:39questions to further improve the plan. So I'm going to say it reacts plus feet, and then I'll say step
00:02:44by step for the article generation. And now it's asked me to confirm the structure of the architecture,
00:02:49which I think looks good. It also asks about the layouts and then asks about the endpoint and server
00:02:54logic. So after I've answered all the questions, it then goes ahead to create a plan, which it's
00:02:58now put inside this directory. And if we take a look at the plan, we can see it's very detailed,
00:03:03giving an overview, the stack, the structure, as well as design, layout, API endpoints, and much
00:03:09more. In fact, this is very similar to what Claude code would give us if it was writing its own plan.
00:03:15But the next step is where things get interesting. So after I confirmed the plan and so I'm ready
00:03:19for implementation, it then goes ahead and writes another implementation plan. It does this by first
00:03:25looking at the original design plan that we created. And it breaks this down into smaller, more
00:03:31manageable chunks of tasks that can be completed by parallel sub agents. So now it's written another
00:03:36plan and it's saved it to the plans directory again. And if we take a look at that one, we can
00:03:41see instead of giving a detailed overview of the project and architecture, it splits things up into
00:03:46tasks. So we have one task here to scaffold the project and it has some steps inside those tasks.
00:03:52We have another task here to create Vite and React client skeleton. And again, it has some steps and
00:03:57this keeps going on with every task that's needed to complete the project. From here, superpowers
00:04:03asks me to pick an approach. Do I want the sub agent driven approach or the parallel session
00:04:08approach? This approach, we use a separate sub agent for each task without any human confirmation.
00:04:14This approach will create another session from the plan session, complete the tasks in batches,
00:04:18and then check with me. So the human, if I'm happy with the batch is completed before it continues.
00:04:24Now for the sake of speed, I'm going to go with the sub agent driven approach. And now it's gone
00:04:28ahead and created the tasks that need to be completed for this project. Now, unfortunately,
00:04:32these tasks look different from the previous tasks because I forgot to record the screen the first
00:04:36time. So I've gone ahead and started again, but I use the exact same prompt. So what's happening now
00:04:41is that each sub agent is first creating a test. So it's writing a test that's going to fail. And then
00:04:47it writes the minimal code to make that test pass. And then when it's finished working on the feature,
00:04:52it verifies the feature passes the requirements that were specified in the plan and then moves on to
00:04:57check the quality of the code, which checks if it's clean, well-built and maintainable. So now it's
00:05:02finished most of the tasks. It's now going to test everything it's done. And here we have it, it's
00:05:06completed all the tasks and it even found one bug in the process that it was able to fix. And if we
00:05:11take a look at the code, it's split by server and source. I'm assuming source is the front end or
00:05:16client. We have some components in here. So article tab, download tab, et cetera. If we take a look at
00:05:21the pipeline progress, we have stages in an object, a TypeScript interface. Now it looks like it's using
00:05:27some kind of CSS in JS for the styling. And in hindsight, I should have specified to use Tailwind,
00:05:32but this is fine. And check this out. This is something Opus plan mode wouldn't do by default.
00:05:37You can see through every step of the way, Superpower has made a git commit from scaffolding
00:05:43the project to adding the CLI wrapper and adding Hano and everything in between. So if I check the
00:05:48git status, there's nothing for me to commit because the branch is clean, which is much less work on my
00:05:53part. Now you may be looking at this and wondering where are the test files? I'll talk about that
00:05:57later. So if I'm being honest, I think Superpowers is an impressive project. I love the fact that it
00:06:02focuses on skills. I think it has 14, which do so many different things from brainstorming to
00:06:08planning to implementing and even reviewing the code afterwards. I love the focus on TDD,
00:06:12specifically red green TDD, where it writes the tests first. So they fail, making them red and
00:06:18then writes minimal code to pass them. But sometimes this doesn't always work because the agent chooses
00:06:24not to do that. Take a look at this. So here we can see the writing plan skill explicitly said TDD. And
00:06:31for some reason, Claude chose not to do it. I asked it to confirm that and it said, yes, it's on me.
00:06:36The skill says TDD and I still didn't do it. And I asked it, why did you not do it? And then it says
00:06:42it focused on shipping quickly over following the process. I'm not sure exactly why this happened for
00:06:47a state of the art model, but it just goes to show that you shouldn't blindly accept what the model
00:06:53does. It's important to read through the plan and make sure it does what you expect it to do.
00:06:57This is one of the reasons why I don't think I'll be using Superpowers for every single one of my
00:07:03projects or feature requirements purely because they're quite small. And if I have a small feature
00:07:08I want to add, I'm happy to go back and forth with Claude to plan this out, write a plan document,
00:07:13and then wipe the context to execute that plan. But if I do find myself in the situation where
00:07:19I'm implementing multiple features, which does happen from time to time, then Superpowers is a
00:07:24great tool to use just because it's very good at breaking down complex requirements into chunks
00:07:31that can be completed by sub-agents. And yes, the Claude code tasks feature can kind of do this,
00:07:37but I love the fact that Superpowers creates a plan for this instead of going ahead to implement
00:07:43the code. Now, how does Superpowers compare to something like Beads or SpecKit or the whole
00:07:48spec-driven development architecture? For me, it seems like a more simple version of planning out
00:07:53before executing. Of course, not as simple as something like Ralph, but maybe in between
00:07:58something like Ralph and Beads. But it seems like to me, there are more and more tools like this
00:08:02helping people to write better code with their AI agent. Now, this is a good thing,
00:08:06but everyone is different. And I think it's good to take a bit from here, a bit from there,
00:08:10and end up making your own workflow that is perfect for you. So maybe that's just something I might do
00:08:16in the future. And if you're lucky, I'll make a video about it to show you exactly how I did it.

Key Takeaway

Superpowers is a robust agentic framework that enhances Claude Code by enforcing structured workflows and parallel sub-agent execution to build complex, well-tested software more efficiently than standard planning modes.

Highlights

Superpowers is an agentic skills framework for Claude Code with over 50,000 GitHub stars.

The tool enforces a structured workflow including brainstorming, implementation, and red-green TDD.

It supports parallel execution using sub-agents to complete complex tasks and review code.

A comparison with Opus 4.6 plan mode showed Superpowers produced a more functional and better-designed UI.

The framework automates Git commits for every step of the development process, ensuring a clean branch.

The speaker notes that while powerful, agents can sometimes skip TDD to ship faster, requiring human oversight.

Timeline

Introduction to Superpowers Framework

The speaker introduces Superpowers as an agentic skills framework designed to prevent AI coding agents from rushing and making errors. It utilizes a structured workflow that incorporates brainstorming, implementation, and rigorous test-driven development. A key feature mentioned is the use of sub-agents to execute tasks and review code in a parallel fashion. This section sets the stage by questioning how this framework differs from standard plan modes or spec-driven development. The framework boasts significant community support with over 50,000 stars on GitHub.

Case Study: Twitter Video Downloader and Article Generator

The speaker demonstrates building a web UI for 'XDL', a CLI tool used for downloading Twitter videos and generating articles. He first attempts the task using Opus 4.6 with standard plan mode, which results in several bugs and a less-than-ideal user experience. In contrast, the version built with Superpowers features a superior design, better file handling, and a more intuitive streaming article generation process. The Superpowers version successfully downloads videos directly to the browser rather than just opening them in new tabs. This comparison serves to highlight the practical improvements in output quality when using the specialized framework.

Workflow and Command Overview

This section dives into the technical setup, starting with the installation of the Superpowers plugin and its unique slash commands. Commands like 'brainstorm' and 'execute plan' allow the user to interact with the agent as it develops the project architecture. The speaker shows how the agent asks clarifying questions about the tech stack, such as React and Vite, before generating a highly detailed plan. The resulting plan document is stored in a dedicated directory, covering everything from API endpoints to layout design. This phase emphasizes the importance of a solid foundation before any code is actually written.

Implementation Strategies and Sub-Agents

Once the initial plan is confirmed, Superpowers creates a secondary implementation plan that breaks the project into manageable chunks. The user is given a choice between a 'sub-agent driven approach' for speed or a 'parallel session approach' for more human control. The speaker opts for the sub-agent approach, which delegates specific tasks to different AI instances to work on simultaneously. Each task is broken down into specific steps, ensuring that the project remains organized even as it scales in complexity. This modularity is presented as a major advantage over monolithic coding attempts.

TDD, Code Quality, and Automated Git Integration

The speaker explains the red-green test-driven development cycle where the agent writes a failing test before writing the minimal code required to pass it. After implementation, the agent verifies the feature against requirements and performs a code quality check to ensure it is maintainable. A standout feature highlighted is the automated Git integration, where Superpowers makes meaningful commits at every stage of the process. This leaves the developer with a clean branch and significantly less administrative work once the coding is complete. The speaker notes the presence of specialized components like an article tab and download tab within the generated source code.

Limitations, Comparison, and Conclusion

In the final section, the speaker addresses potential pitfalls, such as the agent occasionally ignoring TDD instructions to speed up delivery. He stresses that users should never blindly accept AI output and must remain involved in reviewing the plans and code. While he wouldn't use Superpowers for tiny tasks, he finds it invaluable for complex, multi-feature requirements that benefit from sub-agent distribution. The tool is positioned as being more complex than simple scripts like Ralph but more streamlined than heavy spec-driven tools like Beads. The video concludes with the suggestion that developers should mix and match tools to create their own custom AI workflow.

Community Posts

View all posts