A stab at superpowers

I've been using AI workflows to build applications for a little over one year now. I decided, for the first time, to use a pre-baked industry methodology: superpowers.

Superpowers caught my attention because it resembles how I currently work with AI. I plan my work and chunk it down into smaller deliverables, which I execute through tdd.

I have a few dimensions I wanna assess out of this experiment:

• the end result feature correctness and UX;
• the generated code;
• cost, both time and tokens;
• ergonomics: how I feel about this workflow;

Running the experiment

I used my agentic hobby project, conveyor, as the guinea pig for this. Conveyor is a TUI that lets me see which stage of the SDLC my iterations are: Pull Request, merge queue, merged or deployed. A feature it lacks is tabs. I wanna be able to cycle through PRs I wrote and PRs I need to review, for instance.

To keep it simple, this iteration focused on adding tabs just for the PRs column.

I gave two distinct Claude Code sessions running Sonnet 5.5 on medium the same initial prompt:

Let's plan a feature: tabs. Right now each column shows a fixed view based on its config. I would like to add the ability of having multiple configs for each column, and you'd cycle through them by pressing a key that defaults to . I'm ok breaking the current tab behaviour. I envision these tabs appearing at the top of the column. The currently selected one gets a visible highlight.

Both got the PR through the automatic review cycle without any further interaction with me, and in about 40 minutes I had two working versions of the same feature. I was excited to try them out!

Artifacts & Results

Artifact	Current workflow	Superpowers
Issue/plan	https://github.com/piacsek/conveyor/issues/79	https://github.com/piacsek/conveyor/issues/80
Pull request	https://github.com/piacsek/conveyor/pull/82	https://github.com/piacsek/conveyor/pull/81
Screenshot

Cost	Current workflow	Superpowers
Build wall time	41min	42min
Token cost - main session	104.9k	196.7k
Token cost - subagents	483k	1.28M
Token cost - total	587.9k	1.477M

My thoughts

Both implementations worked first time! And both of them kept retrocompatibility with the existing configuration schema.

However, I must say that I'm intrigued. My current workflow beat superpowers in all the dimensions I was interested in analysing: it basically took the same time for almost a third of the token cost to produce an outcome I enjoyed more.

• UI: I don't like the tab number showing up, it's confusing since it's so close to the PR count above;
• Code: I prefer the Tab abstraction over the View one - it clicks faster in my mind;
• Ergonomics: To my surprise, my current workflow asked me more questions than superpower's brainstorming skill: 5x1 - I felt more in the loop;

I wasn't expecting such one-sided result. If you've used superpowers, please comment how you operate it! I'm interested in experimenting more with it.

Thank you for reading, have a great day!