06. When AI Gets More Capable, Keep the Judgment

I read a post on Zhihu: From Superpowers to Matt Pocock: Five Things a 200k-Star Skill Library Taught Me. Matt's approach fits how I work with agents today better than Superpowers does.

The post says:

Superpowers bets that a methodology can take over and save you the judgment. It is too heavy. Matt bets that judgment has to stay with the person. That is not wrong, it is just not what I want. Nobody at the top can decide this for you. You have to work it out with your own head.

To absorb it you have to admit three things: think clearly before you start, write the knowledge down into files, and trust a process over an incantation.

With Superpowers I got the third point that day: the process has to be reliable, and the steering wheel has to be in your hands.

For someone willing to be grilled by AI for ten minutes, this is an amplifier. For someone looking for a fully automatic wishing machine, it is one more bookmark that collects dust.

I still think the judgment has to stay with us. Before any development task starts, the key conditions of the problem need to be spelled out:

• Whose problem is this
• How good the output has to be
• Which costs are unacceptable

AI can help break a task down and surface problems. The trade-offs are ours to judge. If the upfront process is too complete, we may end up throwing away more of the work later.

On this point, Superpowers goes off course easily when the early context is thin. Some of my past projects burned a lot of tokens without reaching a result I was happy with, mostly because the direction changed during design or even during development. So I start with grill-me to pin down boundaries and features, then use Superpowers to write the design doc. When a fork in the road shows up, I go back to grill-me and clear up whatever is still fuzzy.

That is the value of the follow-up questions in a skill like grill-me. When an idea gets asked about its use case, its failure conditions, and who the users actually are, things I had left vague finally come out of the AI's mouth. If some questions still have no answer, I write them into my notes and keep checking. I would not hand a task straight to an AI to write code; fixing it later costs too much. The cost here is the time spent revising and re-verifying, not tokens. Nobody likes inheriting a mess of code. I doubt AI does either.

grill-me gives me clearer boundaries and conditions, which removes a lot of uncertainty. Still, you cannot deny how strongly Superpowers constrains the process.

That process has decades of software engineering behind it.

Requirements → Design → Plan → Implement → Review → Verify

Once the constraints we worked out are written into files, I do not have to re-explain them every time. When requirements change, the design can change with them. That is the core value of a skill like Superpowers for me.

Nothing limits how many skills you can install now. Context is still finite, so I pick skills as needed to save room. My routine: grill-me first until my own thinking is clear, then Superpowers where a process helps.

You may have noticed that I rarely attach skills to a task by hand now. The agent can tell which skills fit the job and load them on its own.