obra/superpowers

Worth a trial if you already let a coding agent handle multi-step work and want it to stop improvising. Superpowers is not a new model or IDE; it is a set of skills plus startup instructions that change how an existing agent behaves inside your normal harness.

The problem

You ask an agent to build a feature, and it starts editing files before it has understood the request. You then spend time correcting scope, re-explaining conventions, and reviewing a diff that includes untested or unwanted work. Superpowers changes the first step: the agent asks what you are actually trying to do, turns that into a spec, shows it in readable chunks, and only then writes an implementation plan. From there, the work is split into tasks handed to subagents, with inspection and review between steps. The result is a slower but more legible start and a code change you can trace back to an agreed design.

What's different

The core mechanism is composable skills plus initial instructions that make the agent load them. According to the project description, skills trigger automatically, so you do not invoke a command before each phase. The methodology explicitly emphasizes red/green TDD, YAGNI, and DRY, and the author claims agents can work autonomously for hours at a time without deviating from the plan project description. That is a workflow claim from the README, not an independently measured outcome.

The trade-off is ceremony. A small typo fix can still be routed through spec and planning steps if the skills misjudge the task, and long autonomous runs consume tokens and time. The project acknowledges this failure mode through a dedicated diagnosing skill: ask the agent what went wrong in a session and it reads the transcript and reports what happened diagnosis guidance. That is useful for debugging the framework itself, but it also tells you the system is not silently self-correcting.

Who should try it

It fits developers using Claude Code, Antigravity, or Codex App/CLI who want a repeatable process around agent-assisted feature work, especially when several tasks must stay coherent across a longer session. The documented installs go through plugin marketplaces or a single plugin-install command per harness, so trying it does not require replacing your editor installation.

It is a poor fit for someone who wants a fast one-shot patch, or who works without one of these coding-agent harnesses. It is also a poor fit if you dislike reviewing intermediate plans; the framework is designed around sign-off points, so skipping them removes much of its value. Teams that already have a strict internal process may find the methodology redundant or conflicting.

Requirements and limitations

You need a supported agent harness, and installation is per harness: using both Claude Code and Codex means installing Superpowers separately for each installation. The README lists Claude Code, Antigravity, Codex App, and Codex CLI, with marketplace-based paths for Claude and Codex Claude Code setup Codex App setup.

The ongoing cost is attention. You still need to read specs, approve plans, and review subagent work; autonomy is bounded by the plan you sign off on. Misbehaving sessions are expected enough that the project documents a diagnostic skill rather than presenting the workflow as hands-off. The README's commercial-services section indicates enterprise support is available, but the open-source path is the documented install and license commercial services.

HotRepo's take

HotRepo tracked this repository from 2026-09-11 to 2026-10-27, with 14 recorded board days and a rise from about 285k to about 306k stars during that sample; the last three recorded days averaged +851/day, though gaps in the series mean that is not continuous daily growth project page. The Hacker News threads are thin evidence: the top-scored one has 5 points and no comments HN thread, and the sampled comment is a copy of the README's "How it works" text, so it does not add independent experience HN comment. Before adopting, check the installation section for your specific harness and confirm the skills load at conversation start; without that, the methodology will not trigger automatically.