You know that sinking feeling when your AI assistant tries too hard? It’s the moment you ask for a simple one-line fix, and instead it launches a full planning session, spawns a sub-agent, runs a review, and burns 50,000 tokens before you’ve even typed a single character.
That’s exactly where I found myself last week with Superpowers — the most beloved Skill in the Codex ecosystem. I’ve been using it daily since March. I even wrote a glowing review calling it “the most complete AI coding workflow.” Back then, it was a lifesaver. Models would dive into code without asking, and Superpowers forced them to stop, plan, and review. It saved me from countless rewrites.
But models have evolved. Codex now has native Plan mode, SubAgent, Review, and validation. And Superpowers? It’s become a default tax — a fee you pay in tokens and friction, even when the job doesn’t need it.
I looked at the latest rule file. If there’s a 1% chance a skill might apply, it must be invoked. Even before entering Plan mode, it checks whether you’ve done brainstorming. No wonder Superpowers and brainstorming appear together in every invocation graph. The tool is so eager to protect you from mistakes that it no longer trusts you to make a simple edit.
That’s not just a token problem — it’s a rhythm problem. Every conversation becomes a ceremony. Every small change triggers a workflow. The tool isn’t serving you; you’re serving the tool.
So I started hunting for something quieter. Something that stays silent until you need it, then cuts like a surgical knife.
That’s when I rediscovered Matt Pocock’s grill-me and grill-with-docs. Two Skills, each with exactly 7 lines of prompt. The shared helper is 12 lines. No lecture on software engineering. Just a few sharp rules: ask one question at a time, suggest an answer, check the codebase for facts, stop when a decision needs a human, confirm alignment before executing.
This is the real value of Skills in the strong-model era. Models already know how to code. What’s scarce is the ability to ask the right questions before the code is written — and to know exactly where to stop and wait for a human.
But the real gold is in grill-with-docs. As it grills you, it maintains two files: CONTEXT.md and an Architecture Decision Record (ADR). The first turns project jargon into shared language. The second captures only the decisions that are hard to reverse — the ones that carry real trade-offs. Most agent-generated docs are long but useless in the next session. These two files turn a one-shot conversation into a reusable decision asset. And they make the downstream pipeline — to-spec → to-tickets → implement — progressively cheaper.
I’m not saying Superpowers is bad. I’m saying it’s a sledgehammer when you need a scalpel. Great discipline should be optional, not compulsory.
Here’s my new rule: match the overhead to the risk. Small change? Let Codex do it natively. Clear requirement? Use Plan mode. Fuzzy idea with many branches? Open grill-with-docs. High-risk refactor? Sure, I’ll still call Superpowers for TDD, system debugging, and review — but I’ll never let it run every conversation again.
I’m uninstalling Superpowers for a week. If quality drops, I can reinstall in one minute. But I have a feeling I won’t need to. Because the best tool isn’t the one that does everything — it’s the one that knows when to do nothing.
FAQ
Q: Isn't Superpowers still better for complex, high-risk tasks?
A: Yes, absolutely. And the article doesn't say to abandon it forever. The point is to stop using it as a default for every single conversation. Reserve it for the 10% of tasks that genuinely need the safety net. For the other 90%, lighter tools save you tokens, time, and frustration.
Q: What's the practical takeaway for someone who uses AI coding tools daily?
A: Audit your workflow. If you feel like you're fighting your AI assistant's process more than you're collaborating with it, look for tools that enforce alignment without forcing ceremony. Prioritize tools that accumulate reusable decision assets (like ADRs) instead of just generating long docs. And most importantly: match the overhead to the risk of the task.
Q: Isn't this just a preference for less structure? What about the risk of AI making mistakes without guardrails?
A: The guardrails we need today are different from the ones we needed six months ago. As base models get smarter, the biggest risk isn't the AI making a coding error — it's the AI wasting your time with unnecessary process. The real guardrail is human judgment at the right moments. Lighter tools don't remove guardrails; they move them to the points where they actually matter.