GSD vs Superpowers vs Claude Code: A New AI King?

CChase AI
Computing/SoftwareSmall Business/StartupsInternet Technology

Transcript

00:00:00Should you be using GSD or should you be using superpowers?
00:00:04Or are both a waste of time?
00:00:06To answer that question,
00:00:07I did a head-to-head-to-head test with superpowers,
00:00:10GSD and vanilla Claude code.
00:00:13I had them create the exact same web app
00:00:15and we grade them on their final output,
00:00:17the amount of tokens used
00:00:18and how long it takes them to build it.
00:00:20And the winner is not the one you would expect.
00:00:23Now, before we run the head-to-head tests
00:00:25between GSD, superpowers and baseline Claude code,
00:00:28let's first talk very quickly
00:00:30about what GSD and superpowers actually are,
00:00:33how they work and how they differ between one another.
00:00:35So GSD and superpowers are cut from the same cloth.
00:00:40These are both orchestration layers
00:00:41that sit on top of Claude code
00:00:43and change the way how Claude code approaches
00:00:46complex projects.
00:00:48It introduces a much more robust planning system,
00:00:50a much more robust testing system,
00:00:52and they both utilize sub-agent driven development
00:00:56to deal with context rot.
00:00:57And the similarities become even more obvious
00:00:59when we take a look at their actual step-by-step process.
00:01:02Superpowers and its first three steps, what is it doing?
00:01:05It's planning stuff.
00:01:06It brainstorms, it uses get work trees,
00:01:09and then it writes plans.
00:01:10What does GSD do?
00:01:12GSD starts a new project, discusses the plan,
00:01:15and then breaks the plan down into phases.
00:01:17They're taking your big idea
00:01:19and they're chunking it up into smaller,
00:01:21specific atomic tasks that they are gonna have sub-agents
00:01:25deliver on down the line.
00:01:27Once the plan has been figured out, what does superpowers do?
00:01:30It does sub-agent driven development.
00:01:32Again, I keep referencing this.
00:01:33Instead of having Claude code
00:01:35execute everything in the same session,
00:01:36like the normal plan mode into execution does,
00:01:40filling up your context window all the time,
00:01:42instead, these sub-agents each get assigned a specific task.
00:01:45That way, they have essentially a clean context window,
00:01:48which means they should give us better outputs.
00:01:51And so that's what happens in step four and five,
00:01:53although it also includes test-driven development
00:01:55or superpowers.
00:01:56The GSD side of things, it's just one step, execute phase.
00:02:01It's pretty much doing step four and five from superpowers
00:02:04in its step four.
00:02:05Then we finish up.
00:02:06Superpowers requests a code review
00:02:09and then merges everything.
00:02:10What does GSD do?
00:02:12It verifies the work and then it shifts it,
00:02:14commits, creates the PRs, done.
00:02:16So very, very similar.
00:02:18And when we talk about the differences, it's pretty subtle.
00:02:21When it comes to superpowers,
00:02:22it's very heavy on test-driven development
00:02:25and this idea of red-green refactors.
00:02:28If we take a look at the actual test-driven development skill
00:02:31inside of superpowers, what does it talk about?
00:02:33It talks about the iron law,
00:02:35no production code without a failing test first.
00:02:37So anytime it's trying to create code for a feature,
00:02:40it creates a test for that feature.
00:02:41It fails the test first,
00:02:44and then it tries to write the minimal amount of code needed
00:02:47to pass that test.
00:02:48From there, it goes into red-green refactor.
00:02:50And if you wanna see the specifics of it,
00:02:52you can take a look at the skills inside of the GitHub.
00:02:53I'll link all this down below.
00:02:55On the other hand, GSD really emphasizes state and context.
00:02:59So it's constantly creating markdown files
00:03:01that reference what you're planning to do,
00:03:03what you've already done,
00:03:05and what's gonna get completed in the future.
00:03:06And things like the requirements of markdown file,
00:03:08the roadmap markdown file, the different phases.
00:03:11It's very explicit and it writes down everything.
00:03:13And the idea with that is with so much sub-agent execution
00:03:16and so much content resetting,
00:03:19we always want some sort of North Star telling us
00:03:21where we are and where we're going.
00:03:23That's GSD's idea.
00:03:24But really those differences are subtle.
00:03:26So a lot of it also comes down to feel,
00:03:28which is what we're going to see today.
00:03:29The other things we're gonna be looking at
00:03:31is the amount of time it takes
00:03:32for each of these to execute the task I'm gonna give it,
00:03:35as well as the amount of tokens,
00:03:36'cause cost is something we always need to be cognizant of.
00:03:39But with that being said, installing these is very simple.
00:03:41Superpowers is in the official plugin library on Claude Code.
00:03:45So if you're inside of Claude Code and you do /plugin,
00:03:48you will be able to see Superpowers right there
00:03:50and install it that way.
00:03:51And with GSD, you just need to run this one command
00:03:53and it will install everything.
00:03:55So what is our test gonna look like
00:03:56for these three guys today?
00:03:57Well, we are gonna have all three of them build us a website
00:04:00for our AI agency, Chase AI.
00:04:02And this website needs three things.
00:04:04One, it needs a landing page.
00:04:06This is the simplest ask.
00:04:08I just want a standard landing page, hero section,
00:04:11and about me services, and then a lead capture form.
00:04:14So this is where we're testing them on a very simple ask.
00:04:17And also I wanna see how they do with web design
00:04:20and skill calling.
00:04:21Are they gonna use the front end design skills?
00:04:23'Cause I'm not gonna explicitly tell them.
00:04:25For two and three,
00:04:25it's all about creating our blog generator.
00:04:27So for step two, I want there to be a page
00:04:30that allows people to see my blog
00:04:32where they can see the different posts,
00:04:34click on them, read them, little basic stuff.
00:04:36And then number three is the actual blog generator itself.
00:04:39This is a hidden admin page.
00:04:40I don't want it on the nav bar.
00:04:42And I wanna be able to give my webpage
00:04:44either a YouTube video URL or an article URL.
00:04:48I then want it to scrape everything from that URL.
00:04:51I want it to use the anthropic SDK
00:04:53to then create a clean blog post based on that information
00:04:57from the YouTube video or the article in my voice.
00:05:01I also want it to grab the thumbnail or hero image
00:05:03from the source and then save it all as a new blog.
00:05:06For time's sake, I'm not doing any authentication here.
00:05:08I'm confident all three of these would be able
00:05:10to execute that just fine with the SupaBase CLI.
00:05:13I then give them both a basic tech stack
00:05:15as well as some sort of aesthetic guidance.
00:05:18But the point here is to give them enough of a direction
00:05:22so we can all grade them on the same thing
00:05:24yet leave enough wiggle room
00:05:26so they aren't just following directions.
00:05:28I wanna see how they think through this prompt.
00:05:30And that's because we left certain things
00:05:32open to interpretation,
00:05:33like how to actually fetch the transcripts.
00:05:36How do I actually get the thumbnails
00:05:38from YouTube URLs I give it.
00:05:39What the actual blog generation system prompt should look like.
00:05:42What should that voice be?
00:05:44And again, whether to invoke any specific cloud code skills.
00:05:47So these are all different things
00:05:50that we should see variance in between GSD superpowers
00:05:54and out of the box cloud code.
00:05:55So I just released my cloud code masterclass last month
00:06:03and it is the number one way to go from zero to AI dev,
00:06:06especially if you do not come from a technical background.
00:06:09I teach you everything you need to know about this tool
00:06:11and we focus on real use cases.
00:06:14And just as importantly,
00:06:15I update this thing literally every week.
00:06:18And since its inception,
00:06:19we've already added almost three hours of additional content.
00:06:23You can find a link to it in the pin comment
00:06:24inside of Chase AI+ and we'd love to have you there.
00:06:27So let's get this test started.
00:06:28I have GSD superpowers and cloud code here.
00:06:32I will be very specific about which tab I'm in
00:06:34so you don't get confused,
00:06:36but I also have the status line down here,
00:06:37which will explicitly state which directory I'm in
00:06:40because they're all in different directories.
00:06:42So for superpowers,
00:06:43we can see that it loaded the superpowers
00:06:46brainstorming skill.
00:06:47And with superpowers, it's pretty fluid.
00:06:50It has like 14, 15 plus skills loaded
00:06:54when you actually install the superpowers plugin.
00:06:57And the idea is cloud code knows
00:07:00just based on how you're talking to it
00:07:02and where you are in the process,
00:07:03which skill it needs to invoke.
00:07:05This is a little different than GSD
00:07:07where you are going to use explicit slash commands
00:07:10like GSD new project when you are using GSD
00:07:13in your directory.
00:07:14So GSD was actually the first one to come back
00:07:17with some sort of questioning after the first couple minutes.
00:07:19So GSD was the first to come back with some questioning.
00:07:22It said our brief was pretty complete
00:07:24because we did give it a fairly robust prompt.
00:07:27But what I do like is it says,
00:07:28hey, here's some taste calls I'm making.
00:07:30And right away, it's calling out a few of the things
00:07:32that we mentioned that could be differentiators,
00:07:34things that we didn't include in the prompt.
00:07:37So we never specified what services
00:07:39we wanted on the landing page, it gives us four options.
00:07:41And then it calls out what it's gonna do for YouTube
00:07:43when it comes to the transcript, as well as the hero image.
00:07:46So I'm gonna go ahead and let it create the project.md file.
00:07:49Now let's take a look at superpowers.
00:07:51So right off the bat, superpowers says
00:07:52it's going to skip the visual companion offer.
00:07:55I'm gonna say I want the visual companion offer
00:07:58because that actually is one of the big differentiators
00:08:00between superpowers and GSD.
00:08:02So I would like to see it in action.
00:08:03And right away, it brings up some design decisions,
00:08:06specifically things for fetching the URL,
00:08:09which again, just like with GSD was one of those things
00:08:11we kind of left it for interpretation.
00:08:13It gives us three options with pros and cons,
00:08:16as well as a recommendation.
00:08:17And then it breaks down the actual thumbnail strategy.
00:08:19So it's a bit more in depth when it came back
00:08:22with these sort of suggestions than GSD was.
00:08:24And that same story plays out here with services,
00:08:27design system, as well as error handling and edge cases.
00:08:30So overall, it's been a bit more in depth
00:08:33with everything it's come back with.
00:08:34So I wrote this looks good,
00:08:35but I would still like to go through the visual companion
00:08:38to make sure we are locked in on the front end aesthetic.
00:08:41And it came back with the visual companion,
00:08:44which is one of its coolest features.
00:08:46So it spun up a dev server,
00:08:48and now it's asking me what we should do for the aesthetic.
00:08:50And it has actual options,
00:08:51four of them right in front of you, which I really like,
00:08:53because it's one thing when it tells you
00:08:56what it's gonna do visually,
00:08:58and spins up one dev server for one options.
00:09:00It's much different when you can see everything all at once.
00:09:03This is one of my favorite part of superpowers.
00:09:04Now, with that being said, these are all very samey.
00:09:08None of these totally jump out at me.
00:09:11Of all these, I would say probably
00:09:13the warm editorial is the best.
00:09:15Electric lime is gross, monochrome is boring,
00:09:18and linear polish just looks like AI slop.
00:09:20Oh, we'll go with this one for now.
00:09:24At least it's something visual.
00:09:25I love the visual companion.
00:09:27So after I told superpowers I liked option C,
00:09:29now it's giving me some more options to look at.
00:09:32So it took that aesthetic, those sort of colors,
00:09:34and now we're going into the hero section.
00:09:35So it continues to drill down on the webpage.
00:09:40So this is the first hero.
00:09:44The second one's a bit more centered.
00:09:46We have the third one with some stuff over here.
00:09:51Split with the featured look.
00:09:53I think, you know, in reality,
00:10:00I would probably do something like this
00:10:03and cut out what's here 'cause this is kind of lame,
00:10:05but I like this as a template, right?
00:10:06This is what we can start with.
00:10:09So we'll go with C.
00:10:10Now superpowers visual companion takes you
00:10:12through every section of your landing page.
00:10:15So we'll skip the rest of them
00:10:17'cause I think you get the point.
00:10:18So superpowers has now written the spec for our website
00:10:21and it's asking us to review it.
00:10:23Once we take a look at it and we give it the thumbs up,
00:10:25then it's actually going to use the writing plan skill
00:10:28to produce the implementation plan.
00:10:30So this is kind of the rough draft blueprint
00:10:32of what it's going to do.
00:10:33And here's a look at that design spec.
00:10:35It is very comprehensive,
00:10:38but the part you should be taking a look at is at the bottom
00:10:41and that is the key judgment calls.
00:10:43What are decisions that superpowers has made for you
00:10:45up until this point?
00:10:46'Cause this is where you need to provide some pushback
00:10:47if you haven't.
00:10:48So it's going to do slash studio as the hidden URL
00:10:50where we sort of do all the actual blog content generation,
00:10:54uses writing as the nav label,
00:10:56talks about the generated voice.
00:10:59So, ah, used to be a marine pilot.
00:11:00Now he's an AI consultant, all right, easy enough.
00:11:03And it did that off of its actual user level cloud memory.
00:11:07And it kind of talks about security.
00:11:08Like I said, we aren't doing authentication here
00:11:11purely for the demo.
00:11:12And it's even like, oh, that's kind of weird.
00:11:13I guess we're just doing security by obscurity.
00:11:15So it calls it out.
00:11:17So I'm just going to tell superpowers that it looks good.
00:11:19And now it's actually going to write out the plan
00:11:21and you can see that skill being loaded.
00:11:23So while we were doing all of that with superpowers,
00:11:25GSC has been executing its own research
00:11:28before it built out its plan.
00:11:29So it spawned four researchers in parallel.
00:11:32One for stack research, one for features research,
00:11:35and then two more for architecture and pitfalls research.
00:11:38You can see that right here.
00:11:39Each of these uses a hefty amount of tokens, right?
00:11:4275K, 33K, 51 and 61.
00:11:46But the idea is if you're doing something rather novel
00:11:49or that isn't very common,
00:11:52these sorts of researchers are going to work wonders
00:11:55in the long run.
00:11:56So what we did today is,
00:11:58or what we are doing today is fairly straightforward.
00:12:00Web design, blog generator, it's seen these things before.
00:12:03I still had it execute these researcher agents
00:12:06just to keep this test equal, so to speak.
00:12:10So then synthesized all of that research.
00:12:13You can see here it uses Sonnet 4.6 for this.
00:12:15So even though for the most part,
00:12:17I told GSD to go wild with Opus 4.6,
00:12:20when it feels like it's just synthesizing information
00:12:23and it's not sending someone out to do something novel
00:12:25or unique for that project,
00:12:27it will use smaller, cheaper models to do the synthesis.
00:12:30And this four agent research phase
00:12:32is robust compared to Superpowers.
00:12:34Superpowers doesn't really do this.
00:12:36But like I said, 30 tool uses, 91K tokens, 15 minutes.
00:12:39It takes time.
00:12:40Once it does the research,
00:12:41it then defines the requirements.
00:12:43Similar to the MD file we just looked at with Superpowers,
00:12:48GSD does the same thing, but even more.
00:12:51It does multiple documents.
00:12:52So it creates a requirements document.
00:12:54It creates a roadmap document.
00:12:56Really it takes kind of what Superpowers did,
00:12:58but just divvies it up into multiple docs.
00:13:01Those docs being roadmap, state, requirements,
00:13:04and eventually things like phases.
00:13:06And at 35 minutes in, you can tell this takes some time.
00:13:10If we actually pause and take a look at, you know,
00:13:13the standard CLOD code, its plan's been up and ready.
00:13:16We haven't had it execute anything yet for some time.
00:13:19This took total for it, I think about five or six minutes.
00:13:22And that was on the slow side, I felt like,
00:13:24versus GSD, which is still going 36 minutes later
00:13:29and hopping back to Superpowers.
00:13:31Superpowers just finished up its website plan dot markdown.
00:13:35While we wait for GSD to finish up its roadmap
00:13:38and its series of stuff,
00:13:39let's take a look at Superpowers once more.
00:13:41So it just created the website plan dot MD,
00:13:44which includes 28 tasks and 2,500 lines.
00:13:47Jumping back into VS code,
00:13:49if we go down into the docs of this folder
00:13:52and take a look at the specs,
00:13:53sorry, take a look at the implementation plan
00:13:56and said, this is what it's talking about.
00:13:58Like I said, about 10 times longer than the specs.
00:14:03So there's a lot, there's a lot going on here.
00:14:09Now Superpowers offers us two execution options.
00:14:12One is sub-agent driven, which is very similar to GSD
00:14:16where each task gets its own sub-agent
00:14:18and therefore its own context window.
00:14:20But like it says here, that's a trade-off
00:14:22because that's a lot for 28, as it says,
00:14:25mostly straightforward tasks.
00:14:27Like, is this a nuclear bomb option?
00:14:29Do we really need it?
00:14:29The second option is inline execution.
00:14:32So we're essentially just gonna be doing this
00:14:34in the same session, pausing for review as needed.
00:14:37And it's gonna be much, much faster.
00:14:39This inline execution is much more akin
00:14:42to what we're doing here inside the standard cloud code
00:14:45where we're just like, yes, bypass permissions,
00:14:47go forth and conquer.
00:14:48Now, because Superpowers is recommending inline execution,
00:14:50we are gonna go with inline execution.
00:14:52And we can see Superpowers executing plan,
00:14:55skill loaded successfully.
00:14:57So now it's gonna start getting to work.
00:14:58And right on cue, GSD has now finished its grand plan
00:15:03for our project.
00:15:05So it's created a project MD, requirements MD,
00:15:07roadmap MD, state MD, cloud MD.
00:15:09And it also has created a folder
00:15:12for all the research it's found.
00:15:14GSD is proposing eight phases with 65 requirements.
00:15:18And like we've talked about before,
00:15:20when it comes to executing this,
00:15:21GSD is very kind of rigid, slash clear,
00:15:25next slash command, clear, next slash command, right?
00:15:27So it's very dun, dun, dun, dun, next thing,
00:15:30next thing, next thing.
00:15:31And it's like very phased versus, you know,
00:15:34I will say that Superpower is a bit more fluid, right?
00:15:37You kind of just talk through it.
00:15:38It knows how to load the command as needed
00:15:40or you expect it to load the skills as needed.
00:15:43GSD is a bit more clear cut.
00:15:45Now, before we go ahead and begin executing with GSD,
00:15:49remember this is all in the planning phase right now.
00:15:51This is the total token count for its sub agents,
00:15:54just for planning and research.
00:15:56We're at 459,862.
00:16:00What does that mean in terms of usage?
00:16:02Who knows?
00:16:03Totally depends when in the day you're using it,
00:16:05what sort of plan you're on, all that,
00:16:06but 460K give or take for planning.
00:16:09Plus we're at 16% right now, call it 150.
00:16:13Let's give it a nice round number.
00:16:15We're gonna call it 600,000 tokens for GSD
00:16:18in the planning phase.
00:16:19And for total time spent,
00:16:21we'll call it roughly 40 minutes, give or take.
00:16:25Now to compare that to the baseline,
00:16:27the standard Claude code plan phase,
00:16:29that took about 10 minutes and it was about 50,000 tokens.
00:16:33Now as for Superpower's token use in the planning phase,
00:16:36it was about 200,000 tokens.
00:16:38So Superpower's 200,000, GSD 600,000, Claude code 50,000.
00:16:47Claude code 10 minutes, Superpower's 40 minutes,
00:16:51GSD about 40 minutes.
00:16:53So that is one of the big differences,
00:16:54A, between the two orchestration layers
00:16:56and standard Claude code is the amount of time.
00:16:58But in terms of token usage between GSD and Superpower's,
00:17:02big difference there because GSD is very heavy
00:17:05on the research.
00:17:06Like you saw, four parallel sub-agents
00:17:09doing a bunch of planning.
00:17:10Now, was that necessary for this project?
00:17:14Perhaps not.
00:17:15But for a big project, understand it will be necessary
00:17:18and that token difference will be there.
00:17:20But that is just one checkpoint,
00:17:21the planning and research phase.
00:17:23Now it's time for execution.
00:17:25Claude code has already started,
00:17:27Superpower's has already started,
00:17:28and I'm gonna kick off GSD as well.
00:17:30Now when it comes to the execution phase,
00:17:32specifically with GSD, it's more hands-on than the others.
00:17:36It's not like, okay, we did planning and research,
00:17:38it wrote its thing, I can just tell it go
00:17:40and I can just like leave for 30 minutes
00:17:42and come back to finish project.
00:17:43Each phase is probably going to require
00:17:46some level of input from you.
00:17:47If nothing else, to kick it off.
00:17:49Because what it wants to do is it wants to ask you
00:17:51to first discuss each phase to make sure
00:17:55you're on the exact same page with Claude code
00:17:58as to what's in your mind, right?
00:18:00What do you want that feature to actually do?
00:18:01What do you want that thing to actually look like?
00:18:03It gets very, very detailed.
00:18:05On one hand, kind of annoying, let's be honest.
00:18:07On the other hand, if this is something very complex,
00:18:09it's probably important that you get that right.
00:18:11So these are things you have to weigh and measure.
00:18:15And what we're going to weigh and measure is,
00:18:17at the end of the day, that all this back and forth
00:18:19actually give us a better product.
00:18:21So for the sake of your time,
00:18:23I'm not going to show you every single phase in GSD.
00:18:27Again, check the video I linked earlier
00:18:29where I do that with GSD,
00:18:31if you really want to see it play out.
00:18:32Just understand that's one of the big differences
00:18:34within GSD and superpowers.
00:18:37And obviously, Claude code as well.
00:18:40So speaking of superpowers,
00:18:42at this point, implementation is complete.
00:18:44We are at a total spend of 250K tokens
00:18:47and 15 minutes have passed since the planning phase.
00:18:49So it's asking me what we want to do and it recommends,
00:18:52let's just keep the branch as is.
00:18:54So I'm just going to say, "Hey, we'll go with your rec."
00:18:56Superpowers then comes back with a summary of what it's built,
00:18:59what it's verified working,
00:19:01things it couldn't verify
00:19:02that needs some sort of manual verification or changes,
00:19:05and then judgment calls it made.
00:19:07And at this point, I'm also going to update my API key
00:19:10so it actually works.
00:19:11Okay, so they all finally finished their execution.
00:19:14So what we're looking at now is their one-shot product.
00:19:18Right here, we have GSD, superpowers,
00:19:22and the baseline Claude code.
00:19:24Now for reference, for how long this all took,
00:19:27upfront GSD by far took the longest to get to this point.
00:19:30All this was off screen of me going through each phase,
00:19:33having it plan, having it execute.
00:19:35That took, frankly, over an hour.
00:19:38And total token spend for the execution phase for GSD
00:19:41was 600,000.
00:19:42So we were looking at total
00:19:44from the beginning of planning phase
00:19:46to having our one-shot being 1.2 million tokens
00:19:50in an hour and 45 minutes to get to this point.
00:19:52For superpowers, it only took about
00:19:54an additional 50,000 tokens for execution in about 15 minutes.
00:19:59So total time, total tokens for superpowers
00:20:02to go from first prompt to actual product
00:20:05was one hour total time, 250K tokens.
00:20:10And for Claude code, we were looking at 200,000 total tokens
00:20:14in about 15 minutes.
00:20:16So kind of wild, the difference there,
00:20:19GSD being by far the longest and the heaviest,
00:20:23and as expected, standard out-of-the-box Claude code
00:20:27being the fastest.
00:20:28So let's see if all that time and token spend was worth it.
00:20:30We're looking at GSD right here,
00:20:32and just sort of a plain background, right?
00:20:37Everything is black, pretty much.
00:20:40Very basic.
00:20:44We have sort of the orange coloring.
00:20:45Like this doesn't look terrible, but like this isn't,
00:20:48you know, you're not blown away.
00:20:50It's like, okay, this was first pass by AI.
00:20:52Like everything looks pretty standard.
00:20:54When I click on the blog,
00:20:55the blog is here with some example stuff.
00:20:58And you know, this looks fine as well.
00:21:01Now let's look at the blog generation piece of it,
00:21:04that little behind the scenes studio page.
00:21:06But when I follow the link, it gives me, we get a 404.
00:21:11So the blog studio generator doesn't even work
00:21:14on the first pass.
00:21:15So I told GSD what the issue is.
00:21:16So it's working on it right now.
00:21:18While it does that,
00:21:19let's take a look at what Superpowers gave us.
00:21:21So here's what Superpowers came back with.
00:21:22And the front end design looks just like what we saw
00:21:25in the visual companion.
00:21:26And again, nothing special.
00:21:28The Claude code as a rule kind of sucks at front end design.
00:21:32If you don't give it really, really good instructions
00:21:35or just load it with a ton of skills.
00:21:37And so, because we kind of left taste and front end design
00:21:41and design work in general of the interpretation,
00:21:43we got something that looks like it was made by AI.
00:21:45So this is okay.
00:21:47It's fine as a base.
00:21:48Here's a look at the blog.
00:21:50It's got pictures and you know, whole blog setup is there.
00:21:56If I go to the studio section,
00:21:58this does work on the first pass.
00:22:01I can see the generator.
00:22:02And if I put in a link to one of my recent videos
00:22:05and it creates a draft for us,
00:22:06grabs a correct thumbnail.
00:22:08And then what it talks about is actually correct.
00:22:09Cause in that video, I talk about things like codecs
00:22:12inside of Claude code, Obsidian and auto research.
00:22:16So it did exactly what it said it was going to do,
00:22:20which is great.
00:22:20Now here's a look at just Claude code out of the box.
00:22:24So pretty standard stuff, nothing crazy.
00:22:29You know, like if we're honest,
00:22:30is there a huge difference in terms of the front end design?
00:22:33If we don't give it great instructions between this
00:22:36and this and this?
00:22:39No, there really isn't, to be totally honest.
00:22:43You could tell me any one of these three
00:22:46created any one of these three
00:22:47and I would not be able to tell the difference.
00:22:49So let's take a look at the blog.
00:22:52Has some fake articles for us and you know,
00:22:57looks fine, very bland,
00:23:00nothing really going on there, but it works.
00:23:02Now let's see if the studio,
00:23:05the blog generator piece works for this.
00:23:08And just like with GSD, this doesn't work.
00:23:11Gives me the link 404 page can't be found.
00:23:14So just like with GSD,
00:23:15I told base Claude code to go ahead and fix this.
00:23:18And while it's attempting to fix a blog generator,
00:23:21let's go back and see what GSD did on its second attempt.
00:23:24Looks like GSD was able to figure it out.
00:23:27Let's paste the URL in here
00:23:29and see if it generates a draft.
00:23:30All right, so it comes back with this draft and markdown.
00:23:33I do like that.
00:23:34The fact that I can kind of edit things in line very quickly.
00:23:39And as for the actual content, it matches what it should.
00:23:41So it did a good job there.
00:23:44And then I can see the actual preview, which is great.
00:23:48So to be honest, I like GSD's implementation of this
00:23:51with the upfront sort of inline editor,
00:23:54more so than I liked superpowers.
00:23:55And we can see it now inside of our blog.
00:23:58And now lastly, we are back with the out of the box
00:24:01baseline Claude code.
00:24:03It fixed its errors.
00:24:04And now we have a look at the blog generator
00:24:06and similar to superpowers.
00:24:08Once I gave it, it just created it automatically.
00:24:11It doesn't give me any chance to edit it
00:24:13or see it as a draft like GSD did,
00:24:15but here's a low res thumbnail
00:24:18and it grabbed all the correct information.
00:24:20And here it is inside of the actual blog page.
00:24:23So what can we take away from all this?
00:24:24Which one of these three actually won
00:24:27in this head to head competition?
00:24:28Well, let's do a quick recap.
00:24:30In terms of total time spent on the task,
00:24:33Claude code just out of the box was about 20 minutes.
00:24:37Superpowers took about an hour
00:24:39and GSD came in at 105 minutes, an hour 45.
00:24:43In terms of tokens, Claude code was about 200K.
00:24:48Superpowers, 250K.
00:24:51And then GSD was 1.2 million tokens.
00:24:56So those are the objective stats.
00:24:59In terms of the subjective,
00:25:00like how did we think they actually did
00:25:02on what they created?
00:25:04Do we have any super strong opinions one way or the other?
00:25:08The answer probably is no.
00:25:11The answer probably is if I actually had mixed all these up
00:25:14and then now had some grand reveal that,
00:25:16oh no, this was actually superpowers that did this one.
00:25:19And this was out of the box, Claude code.
00:25:22You wouldn't care.
00:25:23You wouldn't even know the difference.
00:25:24The only real difference out of all these
00:25:26was that superpower was actually able to do
00:25:27what it was supposed to do on the first attempt
00:25:30for whatever credit you wanna give it for doing a one shot.
00:25:35I mean, truth be told, if it one shots it, great.
00:25:36If you get it on the second attempt, that's fine with me too.
00:25:40The thing you're probably thinking right now too is,
00:25:42well, test is flawed.
00:25:44This wasn't a complicated enough test
00:25:46to where superpowers and certainly GSD
00:25:49could sort of pull away from the pack.
00:25:51The problem with that line of reasoning is,
00:25:53okay, then what is the line in the sand
00:25:55for now this theoretical task is complicated enough
00:25:58to use something like GSD
00:26:00or to use something like superpowers
00:26:02to justify really the time,
00:26:03even more so than the tokens for a lot of people.
00:26:06Is that clear?
00:26:07Is that obvious?
00:26:09I would argue, no, not really.
00:26:11It really isn't.
00:26:12However, we can't admit theoretically,
00:26:16yeah, there probably some ultra complicated task
00:26:18that this makes sense.
00:26:19The problem is defining that and knowing ahead of time.
00:26:21Because if you are wrong about the complexity
00:26:26of the task in front of you and you choose wrong
00:26:29and you go with say GSD or you just go with superpowers,
00:26:32you just cost yourself 40 minutes versus Claude code
00:26:36or 80 minutes if you went the GSD route.
00:26:39That's a big deal.
00:26:41Because truth be told, if I did this again
00:26:43and you asked me who was the winner out of these three today,
00:26:46it was Claude code and it isn't even close.
00:26:48Why?
00:26:48It's not even the tokens, it's the time.
00:26:50Sure, you could probably say this was the worst
00:26:53out of the bunch if we're splitting hairs, but guess what?
00:26:56Maybe it's the worst,
00:26:57but I just got 40 extra minutes to work on it.
00:26:59Or I got 80 extra minutes to work on it versus GSD.
00:27:02Which one do you think is gonna be better?
00:27:04This GSD one I just created?
00:27:06Or me and Claude code with 80 more minutes?
00:27:09Or me and Claude code with 40 more minutes?
00:27:11Should be kind of obvious, right?
00:27:15So, you know, where do I stand at the end of this?
00:27:20Well, my take is that you kind of need a good reason
00:27:25to use these orchestration layers.
00:27:26If I was going to use one today, it would be superpowers.
00:27:31If I was doing a task that I didn't know
00:27:33if it was gonna be too complicated, right?
00:27:35That imaginary line in the sand
00:27:36that nobody knows where it actually lies,
00:27:39and I think we might be getting close,
00:27:40I would use superpowers.
00:27:41'Cause I know it's not going to crush me in terms of tokens.
00:27:44And I'll just go do something else for 60 minutes.
00:27:47Versus if I go to GSD,
00:27:49I kinda gotta be there at the keyboard, right?
00:27:52If I wanna get the full use out of it,
00:27:53I gotta go through all the planning stuff.
00:27:55And it's gonna take a long time and cost a lot of tokens.
00:27:58So if I'm wrong, it sucks badly, right?
00:28:01That really hurts to spend that amount of time with GSD.
00:28:03It hurt doing this video, just sitting there,
00:28:05going through all these tasks for an end result
00:28:07that really wasn't worth it.
00:28:08So if I really do think it's gonna be something
00:28:11so complicated that I need superpowers,
00:28:13then okay, I think you can justify that.
00:28:17But if it's really not gonna be that complicated,
00:28:19or even if it is a complicated task,
00:28:21is it just something that we can break down
00:28:23into like different features and slowly add on to it?
00:28:26And by slowly, I mean actually way quicker
00:28:28than the other options,
00:28:29because I'm just using out of the box cloud code,
00:28:32which is way faster than these other options.
00:28:34The other thing is when GSD came out,
00:28:38and I did a video when GSD came out too.
00:28:41Like I really liked GSD at the time.
00:28:42And same thing with superpowers.
00:28:44When these two things came out originally,
00:28:46cloud code wasn't in the place it is today.
00:28:48And I already can hear people complaining,
00:28:51well, cloud codes nerve today.
00:28:52Like that's not what I'm talking about.
00:28:54I'm talking like the way cloud code approaches problems
00:28:56and some of the scaffolding
00:28:57and some of the way the harness itself works.
00:28:59There's a lot like,
00:29:00just for the fact that like when you have a large plan
00:29:03and you want to execute it and ask you,
00:29:04hey, do you want to clear context and do it like this?
00:29:07That wasn't even a thing.
00:29:08Like cloud code was way more susceptible
00:29:10to things like context rot than when GSD first came out.
00:29:12When GSD first came out, I was like,
00:29:14oh my, this is a godsend.
00:29:15It actually handles context the way it should.
00:29:17Well, cloud code has brought in a lot of those things,
00:29:19which is to say the gap between baseline cloud code
00:29:22in these things has shrunk significantly.
00:29:25While at the same time,
00:29:26there is now a huge gap in terms of speed to execution.
00:29:30And we can't not talk about the speed difference.
00:29:34This 20 minutes versus the 60 minutes,
00:29:36the 105 is the biggest difference out of everything.
00:29:38And it's what you should index on a lot,
00:29:41in my opinion, at least.
00:29:44So in conclusion, less is more.
00:29:47I think for 99% of use cases and 99% of users,
00:29:53just using the baseline cloud code makes the most sense.
00:29:56It's going to be quicker.
00:29:57Even if the output isn't better,
00:29:59you have way more time to bridge that gap
00:30:01and actually leap over these other guys.
00:30:03If you think you are doing a project
00:30:05that is that complicated and you want some extra power,
00:30:08use superpowers because it's relatively lightweight
00:30:12versus GSD, which kind of just feels like a behemoth.
00:30:16And it doesn't feel great to use.
00:30:18I'm going to be totally honest.
00:30:19Using superpowers just is much more fluid.
00:30:21I just talked to it, invokes the skills.
00:30:23I'm not having to like, all right,
00:30:25now we're going to forward slash clear.
00:30:27Okay, I'm in a new set, it's a bit much.
00:30:29And I get why GSD2 came out, right?
00:30:32GC 2.0 was meant to alleviate those problems.
00:30:35But guess what?
00:30:36That doesn't work either
00:30:37because you can't use the cloud code max planet,
00:30:38which means I'm paying absurd prices.
00:30:40So, you know, so much for that.
00:30:45So hopefully that sheds some light
00:30:46on this whole thing for you.
00:30:47I think if you stick with standard vanilla cloud code,
00:30:49you're going to be just fine.
00:30:50Have superpowers in your back pocket if you really need it.
00:30:53Just have the skills on a project level.
00:30:55And frankly, it's tough to say you need GSD
00:30:58unless you're just doing something crazy
00:31:00and you just like having your handheld
00:31:02through every single phase.
00:31:03So that's all I got.
00:31:06As always, let me know in the comments what you thought.
00:31:09We'd love to hear about how you've been using superpowers
00:31:11in GSD and when I inevitably messed up in their application.
00:31:15If you want to get your hands on the cloud code masterclass,
00:31:18make sure to check it out.
00:31:19Link is in my bio, my pinned comment.
00:31:23And besides that, I will see you guys around.

Key Takeaway

Baseline Claude Code is the most efficient choice for 99% of development tasks because the 40 to 80 minutes saved over orchestration layers like Superpowers or GSD provides more time to manually bridge any minor output gaps.

Highlights

Baseline Claude Code completed the web app development task in 20 minutes using 200,000 tokens.

Superpowers required one hour and 250,000 tokens, but was the only tool to succeed on the first attempt without errors.

GSD utilized 1.2 million tokens and took 105 minutes to reach the same development milestone.

Superpowers features a visual companion that allows users to choose between four distinct UI aesthetics like 'Warm Editorial' or 'Electric Lime' during the planning phase.

GSD spawns four parallel research agents for stack, features, architecture, and pitfalls, consuming 460,000 tokens before any code is written.

The 'Iron Law' in Superpowers mandates that no production code is written without first creating a failing test to ensure test-driven development.

Timeline

Orchestration Layer Fundamentals

  • GSD and Superpowers function as orchestration layers sitting on top of Claude Code to manage complex projects.
  • Both systems utilize sub-agent driven development to maintain clean context windows and prevent context rot.
  • The workflow across both tools follows a structured path of planning, atomic task chunking, sub-agent execution, and final verification.

These layers change how Claude Code approaches development by introducing robust planning and testing systems. Instead of executing everything in a single session, tasks are broken down into specific units for sub-agents. This modular approach aims to improve the quality of outputs by resetting the context for each task.

Distinct Methodologies: TDD vs. State Management

  • Superpowers prioritizes test-driven development through a red-green refactor cycle.
  • GSD focuses on maintaining state and context via persistent markdown files like roadmap.md and requirements.md.
  • Superpowers enforces an 'Iron Law' where minimal code is written only to pass a previously failed test.

Superpowers relies heavily on automated testing to guide the development process. GSD uses a different philosophy, creating a 'North Star' for sub-agents by documenting every completed and future phase in markdown. These differences are subtle in feel but represent different priorities in software reliability and project tracking.

The Benchmark Task: Chase AI Web App

  • The test requires building a landing page, a blog viewer, and an admin-only blog generator.
  • The blog generator must scrape YouTube transcripts or articles and rewrite them in a specific voice using the Anthropic SDK.
  • Success is measured by final output quality, total token consumption, and time to completion.

The prompt provided to all three tools includes a specific tech stack and aesthetic guidance while leaving technical implementation details open to interpretation. This includes how to fetch YouTube transcripts and how to design the system prompt for the blog generation. The goal is to observe how each tool handles wiggle room and architectural decisions.

Planning and Research Phase Performance

  • Superpowers offers a visual companion tool that provides live dev server previews for UI design decisions.
  • GSD executes heavy upfront research using four parallel agents, taking 40 minutes before starting execution.
  • Standard Claude Code completes the planning phase in roughly 5 minutes using only 50,000 tokens.

During planning, Superpowers engages the user with UI options and design specs, including key judgment calls based on user memory. GSD is significantly more resource-intensive, utilizing nearly 600,000 tokens just to define requirements and roadmaps. This research-heavy phase is designed for novel or highly complex tasks where architectural pitfalls are common.

Execution and Resource Consumption Analysis

  • GSD requires 1.2 million total tokens and 105 minutes to produce a one-shot product.
  • Superpowers completes the task in 60 minutes using 250,000 total tokens.
  • Vanilla Claude Code is the fastest, finishing in 20 minutes with 200,000 tokens.

Execution in GSD is rigid and hands-on, requiring users to discuss and kick off each phase individually. Superpowers offers a more fluid experience, allowing for inline execution in a single session which is much faster than its sub-agent option. The baseline Claude Code is the least demanding in terms of both user time and API costs.

Final Product Comparison and Verdict

  • Superpowers was the only tool to produce a working blog generator on the first attempt.
  • GSD and vanilla Claude Code both required a second attempt to fix 404 errors on the admin page.
  • The front-end design quality across all three tools was nearly identical without highly specific instructions.

While Superpowers achieved a one-shot success, GSD's implementation included a superior inline markdown editor for blog drafts. However, the time gap is the deciding factor; the 80 minutes saved by using vanilla Claude Code over GSD provides a significant window for manual refinement. For most users, the speed of the baseline tool outweighs the architectural overhead of the orchestration layers.

Community Posts

View all posts