I Started Treating AI Coding Like an Engineering Process, Not a Magic Trick

#ai

The most useful AI coding improvements I have seen lately are not about making the model sound more confident. They are about making it less reckless.

That is why Superpowers caught my attention.

Superpowers is an agentic skills framework and software development methodology from Jesse Vincent. The idea is simple: give coding agents a stricter process so they stop jumping straight into implementation, stop inventing requirements, and stop claiming work is complete before verification.

In practice, it pushes the agent toward a workflow that looks much closer to real engineering:

clarify requirements before touching code
write a design before implementation
split work into small reviewable tasks
use git worktrees to isolate changes
follow TDD instead of writing tests after the fact
verify before claiming completion
use review-oriented agent workflows when useful

That matters because the failure mode of AI coding is often not “it cannot write code.” The failure mode is that it writes too much code too confidently.

If you have ever asked an agent to fix a bug and watched it declare victory without reproducing the issue, you know the problem. If you have ever asked for tests and got something that could never fail, you know the problem too.

Superpowers does not magically solve software engineering. It adds friction in the right places. For tiny edits, that friction can feel like overhead. For larger features, bug fixes, and refactors, it can prevent a lot of wasted time.

I wrote a full review of Superpowers on Open Agent Skills, including how it works, where it helps, and where it still feels heavy:

https://openagentskills.dev/skills/superpowers

If you use Claude Code, Codex, Cursor, Gemini CLI, Copilot CLI, or similar agentic coding tools, Superpowers is worth a look. The bigger lesson is even more important: the future of AI coding may not just be smarter models. It may be stricter processes around models.

Links: