Most agent failures start in the same place. You ask it to build something. It dives straight in. Fifty minutes later you have 400 lines of code solving the wrong problem.

The problem is that we use AI agents the way we used Google: type what you want, hit enter, hope for the best. But agents are not search engines. They are colleagues. And the most important difference between a good and a bad colleague is not how fast they can write code. It’s whether they stop and ask.

🎯 What it looks like

This week I installed a plugin for Claude Code called Superpowers. It’s made by Jesse Vincent, sits at 149k stars, and it’s one command to get going:

/plugin install superpowers@claude-plugins-official

What the plugin does is force the agent to work in a specific order. Instead of coding first and asking later, it starts by brainstorming: what are you actually trying to achieve, who is the user, which edge cases matter, what is the success criterion. Only once you’ve approved the spec does it start planning. And only once the plan is in place does it start writing code, in small pieces, tests first.

⚡ Why this matters to a product manager

This is not a developer tip. It’s a way to change how your team uses agents.

Most places I see agentic coding in practice, it looks like bikeshedding on steroids. The developer asks the agent to build a button. The agent builds a button. Two hours later they’ve built a button that works, but nobody asked whether it was the right button in the right place for the right user.

Superpowers removes that trap by forcing the conversation up into the right layer. What are we trying to achieve. Who is the user. What is the success criterion. Those are PM questions, but now the agent is asking them, and the developer has to answer before any code gets written.

That means the conversation between PM and developer has to happen before the feature hits the editor, not after. That’s a pretty big change in how teams work, if you take it seriously.

👉 Your tip this week

Ask your developers how their agents plan before they build. Are they using something like Superpowers, or does the agent jump straight into the code? The answer tells you more about the quality of what’s being built than any burn-down chart.

If you work in Claude Code yourself, install the plugin and try it on a small feature. Not the big refactor, start with something that normally takes an hour. You’ll feel the difference immediately.

Best, Martin