Why I’m Thinking of Uninstalling the Superpowers Skill in My AI Coding Workflow

I have been using the Superpowers skill in Codex almost every day, and it was once praised in the article “The Most Underrated Skills” as the most complete AI programming workflow.

Back then the model would start writing without much guidance, so Superpowers forced a workflow of requirement discussion, planning, TDD, and review, which prevented a lot of rework.

Since the model upgraded, Codex now includes its own Plan mode, SubAgent, Review, and verification processes, yet Superpowers remains excellent but feels overly eager.

Good Discipline Turning Into a Default Tax

According to the latest using-superpowers rule, if any skill has a 1% chance of being relevant it must be invoked, even before entering Plan mode, and a brainstorming check is required. This explains why Superpowers and brainstorming appear together in the call‑ranking chart.

The problem is not just token consumption; a small change I wanted triggers an entire new workflow—planning, questioning, sub‑agents, and layered reviews—so the work rhythm is taken over.

Such forced processes are like engineering insurance: worthwhile for high‑risk projects, but paying the automatic renewal for every conversation is not cost‑effective.

I Prefer a Surgical Knife

Recently Matt Pocock’s grill-me and grill-with-docs skills have become popular. Matt previously created Total TypeScript and ts-reset, and I have also contributed to his open‑source Skills.

# Install Matt’s skill set with one command
npx skills@latest add mattpocock/skills

# Or install only the two core skills
npx skills add mattpocock/skills --skill=grill-me
npx skills add mattpocock/skills --skill=grill-with-docs

Unlike Superpowers, which can run hundreds of lines of prompts, the entry files for grill-me and grill-with-docs are only seven lines each, sharing a common grilling module of just twelve lines.

These skills do not try to teach the model a full software‑engineering course; they simply enforce a few key actions: ask one question at a time, attach a suggested answer, look up factual information from the codebase, defer decisive choices to a human, and require mutual confirmation before execution.

The author believes this is the most valuable aspect of skills in the strong‑model era: the model can already write code, but the scarce resource is asking the right question before coding and knowing where to pause for human approval. grill-with-docs also maintains a CONTEXT.md and ADR files during the questioning process. The former turns project jargon into a unified language, while the latter records hard‑to‑retrace decisions with real trade‑offs.

Many agent documents are long, and each new conversation would require re‑explaining them. These two files turn a one‑off dialogue into reusable decision assets, making the pipeline to-spec → to-tickets → implement increasingly efficient.

A good workflow should capture decision assets while keeping control in human hands.

My New Choice

For small, clear‑cut changes I act directly; for well‑defined requirements I use Plan mode; for ideas with many branches I invoke grill-with-docs.

When facing high‑risk refactoring, I still call Superpowers for TDD, system debugging, and review, but I no longer keep the whole framework resident in every conversation.

Conclusion

I plan to disable Superpowers for a week, let Codex’s native capabilities handle ordinary tasks, and use grill-with-docs for fuzzy decisions. I will add strict engineering discipline only when needed, and if quality noticeably drops I can reinstall Superpowers in a minute.