The Technical Context
I wouldn't look for magic here. When Claude Code users enable Superpowers on top of the usual /plan with a chain like brainstorming, writing-plans, and subagent-driven-development, I almost immediately expect overhead. This is a typical trap in AI automation: you think you're adding 'smart modes,' but you're actually inflating the context and the number of internal steps.
User reports paint a consistent picture: a small task, a few commits' worth of changes, a fresh session, yet the agent grinds away for nearly an hour. On Opus 4.7 xhigh, this is especially frustrating because the wait time doesn't match the scale of the work.
I've looked into the mechanics of Superpowers, and it's all logical: the plugin enforces discipline through plans, TDD cycles, sub-agents, reviews, and checkpoints. It looks great on paper. In practice, each layer adds instructions, intermediate artifacts, and new passes over the context.
This explains the feeling that Claude Code has been 'getting slower every day.' It's not necessarily throttling. I haven't seen any official confirmation of artificial delays for Enterprise users. However, the 'token fat' hypothesis seems very plausible: the more scaffolding, the more the model reads its own output instead of the code.
Another interesting signal from the discussions is that people are uninstalling Superpowers and going back to the default /plan. This is telling. If removing a single skill-pack noticeably speeds things up, the problem isn't with a specific model but with the prompt orchestration architecture around it.
Impact on Business and Automation
For businesses, my conclusion is simple: not every agentic add-on is useful in production. If your AI implementation gets bogged down in multi-step 'smartness,' you're paying not for the result but for the rituals the model performs within the session.
Teams that prioritize strict processes, test coverage, and formal discipline might benefit. But those who need rapid AI integration into real development, support, or internal automation pipelines without unnecessary token noise will lose out.
I would test this very pragmatically: run the same task with and without Superpowers, logging the time, tokens, and number of iterations. At Nahornyi AI Lab, this is exactly how we build AI solutions architecture for clients, because a beautiful agentic scheme without measurement quickly becomes an expensive illusion.
If your Claude Code has already started to lag and your team is losing hours waiting, don't speculate about subscription conspiracies. Instead, break down your workflow step by step: where is the context bloating, where is the plugin interfering, and how can you build AI automation without the extra overhead? If you'd like, at Nahornyi AI Lab, I can help you quickly analyze this and build a working scheme for your process.