AI coding agents jump straight into writing code—this tool wants them to think it through first

The most common way to use AI coding agents is to throw in a requirement and wait for them to change the code directly. In a tweet, 0xMarioNawfal said most people are using AI coding agents completely wrong. The approach he mentions, Superpowers, is to first have agents like Claude or Codex plan and test—think clearly about what to build—before starting to write.

The tweet gives three steps:

1. Install: add Superpowers to your coding agent.
2. Plan: tell it what you want to build.
3. Execute: have it write code, test, and debug step by step.

According to the tweet, the result is less bad code, less wasted time, and better outcomes. The repository is here: obra/superpowers.

This selling point is not complicated. What really gives people headaches with AI coding agents is often that they write too fast. They edit the wrong files, pile on tests, go in circles debugging, and end up reworking everything. The planning step looks like a way to constrain execution.

The comments are not unanimously positive. One user directly asked how detailed the planning step needs to be after installation. Another question is more practical: how does it get out of debugging loops? Someone else poured cold water on it, saying that adding a planning step usually just makes Claude write fifty lines of markdown explaining what it wants to do, and it still ends up touching the wrong files. This doubt is worth considering. If the planning document is disconnected from the state of the codebase, then no matter how long it is, it is just formality.

Some users also think the process matters more than the model. Another Chinese reply said that blindly starting to write does easily lead to rework, but whether these processes work ultimately depends on whether they suit the project. This judgment is fairly solid. For small projects and tasks with clear boundaries, the planning cost may outweigh the benefit. For complex projects, multi-person collaboration, and codebases with heavy historical baggage, planning first is more worthwhile.

Tests are not always better when there are more of them. One user mentioned that Astra kept adding checks to his fixes, and Claude eventually redid the work and deleted most of the checks. This example shows that tests need boundaries. If the agent does not know which checks are meaningful, it will use the number of tests to cover up a lack of judgment.

The pain point Superpowers addresses is clear: make AI coding agents leave a plan before acting, and test and debug during execution. Whether it can deliver depends on whether the planning is specific enough, whether testing converges, and whether execution is truly constrained by the plan. Tools can change processes, but engineering judgment still has to come from people.