Hacker News Reader

Comment by artisin | Hacker News Reader

I gave Superpowers 5.x a whirl for a week, and aside from consuming a stupid amount of tokens, it did materially worse across all my personal benchmarks and general day-to-day development compared to plain Codex/Claude. I'm convinced it's either some 4D ploy by the AI cartels to set tokens ablaze, or it only provides Superpowers to those without any power to begin with. Rating: 1/5 Pinocchios. Would not recommend.

View on HN

I'm a certified Superpowers hater. It's just not necessary with the modern models and fills up the context windows with garbage and adds an insane number of turns for no benefit.

I had similar prompts back when the models were terrible at instruction-following, so it was actually useful to fill up their context with a mass of instructions so they'd be less likely to forget rules.

Now I've got a few small slash commands or pasted prompts that work perfectly every time as the models follow them exactly.

Yeah I like the idea, but I'd rather just not use most of these plugins, superpowers included. Code seems to include the best tricks in itself anyway at a fast pace.

Likewise. I use Reasonix with planning-with-files and that's been better than even openspec. Cache hits with the help of Reasonix on Deepseek make requests practically free. And that's unsubsidized, with American providers like Digital Ocean or cloudflare.

6.x feels much more efficient with respect to token usage to be fair.

I picked up superpowers back when it first started gaining traction; the first iteration felt like an “oh shit” moment for me, then the sheen quickly wore off. Higher spend, slower throughput and mediocre results made me eventually drop it and go back to plan mode, which had improved significantly during that time.

Coming back, 6.x does feel different and I’m back on the superpowers train. I’m finding it great at taking discrete tasks from beginning to end with very little hand holding.

I run every session with a /goal as well: “Spec + Plan is written and you have implemented the plan without my involvement. You have validated that the implementation is complete and ready to merge”

It’s also great in situations where you may need to complete a plan over multiple sessions, because you get a whole ton of state with superpowers that new sessions can pickup on.

This. I found superpowers a huge token guzzler. And more generic a skill is the worse it seemed to perform. I have found that skills are something you need to build yourself and for your needs and most importantly be willing to throw away. One team I know blindly checked this into every repo they had. They also had the highest cost per pr across all teams in our org (of about 60 eng teams). AI has already given people superpowers. How sad is it that they now need to be told how to just chat and prompt and use AI effectively as a pair programmer:(

This. I found the original superpowers was a great start, but rewrote all those skills to fit my workflow, and iterated on that.

Writing skills is the new writing code.

I haven't used superpowers yet, but it seems a major focus of this release was to reduce clocktime as well as token spend.

From TFA (well, blog):

> The long and the short of it it is that across about 36 hours of work and what would have been $650 of unsubsidized token spend, our Anthropic eval benchmarks were looking like we'd reduced wall-clock runtime for Superpowers builds by 50% and token spend by 60%.

What I like about superpowers is that for my workflow, I spend most of my time brainstorming with the claude session. Using its brainstorm mode helps to keep it from shifting to just writing code. That's basically never what I want until I actually want it. Once we've locked in a design, whatever, I don't care.

But when I don't trigger brainstorm mode, even using the built in plan mode, it's just never as in depth of a brainstorm partner for me.

It sounded like it might not hurt and seemed endorsed by Claude and codex because they both had plugins for it by default BUT I ripped it out after I kept seeing Claude/codex TDD things like when I asked them to make a pydantic model immutable. I’d end up with unit tests testing that my immutably configured pydantic model was immutable or tests that setting foo=bar was actually foo=bar in app config.

I found it slowed me down significantly at first, and produced more verbose code. After a few weeks of using it, I think I've gotten used to it (sometimes I explicitly bypass it, but it's good enough to know which skill to use).

Yeah on the token consumption, I'll be doing something small at work, and it'll consume a lot of tokens.

And the fact that this article’s story is basically “I prompted Fable with a goal and went to sleep and the model got it done” is telling me that the latest models have gotten past the need for Superpowers… even the creators of superpowers is just using a simple /goal!

  The model 
    |
  The harness
    |
  The harness of the harness?!!

This structure doesn't make any sense to me. It's like adding a half leather half shag wrap to your steering wheel. Not to mention the harness itself is updated almost multiple times daily. I'm sure these framework authors are keeping tabs on the harness of the harness performance for every release of the harness.

It's harnesses all the way down.

I only use superpowers when I want to stick with composer2.5(fast) for everything. If I use it with other models it's terrible. With composer2.5 it is slightly better, though not much.

Same here. I tried something similar to Superpowers and it went completely overboard for a small bug fix - writing a TDD, generating artifacts, etc.

Reply on news.ycombinator.com