Superlite: A Faster Alternatie to Superpowers
contextfirst.blog
contextfirst.blog
Everyone is pushing Agent Fleets and Swarms, and I'm sure that there are many success stories, but from our internal experiments many devs don't see any speed advantage for complex work. Here's what one of them told us recently: "The speed advantage is offset by the manual refactoring work. The code is 'correct', but it's just not what any of us would have written".
Now one could argue that you could just ignore the code if it's working (that is in fact what Ex Co instructed us to do) and just keep running these fleets if the end result "works", but I'm not fully convinced this will end well in the long term.
Then again, I'm stuck with decades of legacy code and systems that noone understands that were introduced by humans...
I broke apart a 1k length script into functional components because the local AI just couldn't figure out where edits landed. But once that was complete, it had no problem tweaking features in those scopes.
I think the real TAM is local coding models starting around 32GB sometime in 2027. Obviously, Capitalism gets to decide what we're blessed with but knowing that you can get Qwen3.8-Flash-Next running at around 40GB VRAM with SSD offloading means data centers are either going to become way more efficient, or they're going to die on the vine.
But that's despite many efforts made to keep that code understandable, maybe even documented. If now that's not even tried the results will be a thousand times worse - and the dependency on the LLM absolute.
The LLMs are so much faster than I can check the output anyway. If I have a long-running task I just work on another one in parallel while I wait on the first one.
Many are using a sub, which means (AFAIK) you can't override the system prompt:
https://platform.claude.com/docs/en/release-notes/system-pro...
I imagine it's harder to follow relevant instructions once they are buried in a mass of otherwise irrelevant, but undifferentiated instructions.
Swarms don't necessarily need to be about speed. You can get a lot of value from small swarms doing things like analysis or verification work.
But I think many people are still missing the point about how AI systems should learn. Arguably, with people we tend to hire and "grow" in-house experts as opposed to outsourcing everything, so why shouldn't we build a system that learns from experience, from its failures?
At first I thought the key was in skills, but skills are static, while you need a continuously learning system
Even if you have a basic skill-based workflow, a rudimentary retrospective > process improvements artifact that can be ingested on the next run can have a positive impact.
I'd fine with not what we'd have written ... if it worked.
Trying ChatGPT for a simple one page of HTML JS having draggables, I said clamp touch point to visible viewpoint. Fine. Then I said add 3x drag gain. It put that after the clamp, whereas anyone/thing having an ounce of intelligence would have urecognised the only possible purpose of the clamp and so put the gain before.
This is fundamentally why we cannot trust stochastic parrots to code. They lack the human's understanding of purpose which cannot be abstracted from the code yet is essential for success.
Try rereading the report please.
I have found most of the other skills to emulate an 'enterprisy' way of developing, having everything checked multiple time by 'councils' and whatnot, making things slow, vacuum tokens like mad and the whole process reeks of cargo culting.
Making active decisions and weighing different alternatives improves my understanding, and I think the result is better, though that might just be because I understand it. Which is worth it on its own.
I imagine by this point the models already have the plugin code in their training corpus, plus plenty of RL data from users interacting with the plugins.
The last straw was two weeks ago when I discovered on a run Fable accounted for this problem in by first writing all the code in a scratch git repo and running unit tests, before copying that code into the plan. Since then I’ve been back to vibing. I lose overnight runs but gain sanity.
Theoretically the value of systems like Superpowers should be regulating bias like this.
You gain sanity only if (insanely) you never test the output.
The latest frontier models exhibit good process that was trained into them during the RL phase, and whilst that might make them massively more capable in terms of one-shot, or general coding ability, models without skills simply don't have a durable process.
If you have a need for procedure in your use of agents, you absolutely need skills like this. I'm not commenting on the relative merits of this or superpowers, but rather the seemingly religious debate, "skills are good" vs "skills are bad" discussion that keeps coming up.
Most of the comments here seem to focus on specific things people were trying to get out of skills, and whether those are still needed.
That is going to vary a lot.
If your goal with something like superpowers was to have a process that made your model work "properly" on tasks, sure that's probably not needed anymore.
If your goal with something like superpowers was to get a consistent set of artifacts alongside what was developed (quality varying, so i mean consistent here as "one design plan, one implementation plan", not "good"), nothing has changed, models do not do this consistently, in a durable fashion, without something telling them to do this.
Yes, for example, newer Opus will often decide to write a spec. But not always, and not with any consistent format or ...
If you want that to happen, you'd still need a skill making it happen.
If you don't care about that, then sure, you don't need the skill anymore.
People seem to have used various skills to achieve different things, and so have wildly different views on what is still necessary.
I do use skills for other things, I don't believe that skills are bad. I'm just not sure what they add in this case.
I am not so convinced yet that subagents with haiku and sonnet really deliver better results versus opus, nor if the speed really matters that much if you are truly autonomous. I mean he is talking about saving like 20 minutes here. There are 24 hours a day, you are much more likely to be gated by product specs and knowing what you want and good ideas than engineering if you are hammering things now.
Why not just tell Claude to figure it out and build what it needs? Why do I need a plugin or a factory or whatever.
I just ask “can you go faster?” And claude suggests some stuff and mostly just works autonomously.
Basic issue is this:
1. The models are probabilistic in their output. This is objectively verifiable fact. Give same prompt, repeat and you'll get different output.
2. With these markdown heavy approaches, the system has to keep a track of the system state not just in code but in tons of markdown files as well.
3. As the system evolves due to features, bug fixes, the corresponding markdown files are to be updated as well to keep in sync.
4. This update consumes tokens for sure so now token spend goes through roof.
5. But because the models are probabilistic, the real problem is that it is guaranteed that some of the places would not be updated or updated with misinformation in those markdown files.
6. Next agent run when loading the context from these markdown files would have misinformation in its context that it would discover mid task, would have to rectify the situation. token spend would be quadrupled.
7. Only those edge cases of markdown files would hit that has been visited so there would be many inconsistencies still and they would pile up and compound with the passage of time.
8. Anyone who has significant experience with such approaches would testify to that.
9. There's no solution to that other than just using agents as power tools and holding the context and desired outcome in your own head.
The word you're looking for is unreliable.
> But because the models are probabilistic, the real problem is that it is guaranteed that some of the places would not be updated
untrustworthy.
Claude Sonnet 5.5 late enough for you? It wrote some HTML JS and I gave it a screenshot of a fault.
It IDed the fault. Impressive.
It explained it with "It is caused by".
Yes.
"It is caused by".
End of output.
Not impressive.
No, I have not met a programmer so unreliable.
I get this sense that because models are great, the work gets done, and the results are perceived as good enough by the user, then it creates this illusion that the tooling is responsible for it.
I feel like it's kinda like saying "if it wasn't for VS Code then my API wouldn't be this good" which is a very strange correlation to put out there.
I see this tendency on measuring workflows and productivity chasing instead of software, this can create an addictive trap and the illusion that the workflow beats looking at the problem, defining it, and optimizing for how it needs to change.
One thing I learned working across several industries and problem domains throughout my life is that there is no "one size fits all" here and most often productivity hacks are not true bottlenecks.
We stopped doing big design up-front for a reason, yet it’s roaring back.
We know the devil is in the implementation details and you can’t anticipate everything before you try to implement it. The implement/discover/adjust approach cycle is the best cycle for a reason.
The approach of handing plans down to dumber agents is also a bad idea, for the same reason. A dumber agent will plow right through unanticipated friction in the implementation, and it won’t know to stop and raise an issue.
Even smarter agents will try to complete a plan as written without altering course.
Nothing beats implementing quickly, reviewing, addressing, then reviewing again.
If you’re on a sub, you can afford to waste tokens on this planning theater, I suppose.
When I saw the title Superlite, I thought it's a Superpowers fork optimised for less token usage...
Yeah, for five minutes you needed to use a specific style, very specific language patterns and what have you, in order to get results that were a cut above. But then the models and the harnesses got better and that moment passed.
There are other ways to achieve this same effect without loading the full superpowers suite though.
We use to say “web development is changing fast” when a thing changed after two years, and now it looks like it was changing at a snail’s pace compared to AI.
Plan mode tends not to ask as many questions and go as deep, so often edge cases and other details get missed.
Even when starting Claude in dangerous mode, you still need to babysit it! Annoying.
works for me
Fast is easy.
Fast and good, not so much.