Skip to content

superpowers.mdx

Latest commit

 

History

History
90 lines (77 loc) · 4.98 KB

superpowers.mdx

File metadata and controls

90 lines (77 loc) · 4.98 KB
Raw
title Superpowers - agentic skills framework
tagline Composable skills + a methodology that makes your coding agent step back, spec, plan, TDD, and run autonomously for hours without deviating.
attribution Jesse Vincent
tier framework
canonical_url https://github.com/obra/superpowers
requires Claude Code primary. Also runs on Codex CLI, Codex App, Factory Droid, Gemini CLI, OpenCode, Cursor, GitHub Copilot CLI.
upstream
repo ref paths
obra/superpowers
main
README.md
CLAUDE.md
skills/subagent-driven-development/SKILL.md
setup_command # Claude Code, official marketplace (recommended): /plugin install superpowers@claude-plugins-official # Or via Superpowers' own marketplace (for related plugins too): /plugin marketplace add obra/superpowers-marketplace /plugin install superpowers@superpowers-marketplace # See README for per-harness install (Codex, Cursor, Factory, Gemini, # OpenCode, GitHub Copilot CLI). Each harness installs separately. # Then just talk to your agent, Superpowers skills auto-trigger. # Try: "I want to build X", agent will spec, plan, TDD, then run.
when_to_use Real software where you want methodology, not vibes, when the agent should step back and spec, write an implementation plan, TDD the build, and dispatch sub-agents in parallel. Particularly strong for greenfield + multi-hour autonomous runs.
when_not_to_use One-off scripts or quick fixes where the methodology overhead exceeds the work. Skills auto-trigger, but the install + discipline only pays back at non-trivial project size.
tags
skills
sdlc
tdd
subagent-driven-development
multi-harness
mit
use_cases
building-features
spec-authoring
test-generation
sources
title author url year
obra/superpowers
Jesse Vincent
2025
title author url year
Superpowers marketplace (related plugins)
Jesse Vincent
2025
related
patterns antiPatterns practices glossary tools workflows
orchestrator-workers
plan-execute-verify-commit
spec-driven-development-kiro-style
red-green-tdd-with-agent
one-shotting
cookie-clicker-mode-passive-supervision
custom-slash-commands
spec-driven-workflow
verify-everything
human-in-loop-checkpoints
slash-command-custom-command
agent-fleet-agent-cluster
claude-code
codex-cli
cursor
gstack
grill-prototype-rewind
tdd-red-green-refactor-skills
spec-kit-specify
compound-engineering-plugin
loop
trigger steps gate exit
Task
Spec
Plan
TDD build
Spec satisfied?
Done

A complete software development methodology by Jesse Vincent built on top of composable agentic skills. The headline thesis: the moment your coding agent sees you're building something, it doesn't jump into code, it steps back, asks what you're really trying to do, teases out a spec in chunks short enough to digest, then writes an implementation plan clear enough for a junior engineer with no taste to follow. Only then does it start writing code, via subagent-driven-development with continuous review.

187k+ stars (Oct 2025 release), MIT, and shipped as plugins for the major harnesses including Anthropic's official Claude Code marketplace.

The 14 skills

The framework ships 14 composable skills under skills/ that auto-trigger based on what the agent is doing:

  • brainstorming, explore solution space before committing
  • writing-plans, implementation plans clear enough for a junior to follow
  • executing-plans, work the plan, don't deviate
  • subagent-driven-development, fan out to subagents per task; review their work
  • test-driven-development, red/green/refactor, YAGNI + DRY
  • systematic-debugging, disciplined diagnosis instead of guess-and-check
  • verification-before-completion, never declare done without proof
  • requesting-code-review / receiving-code-review, pair-review skills
  • using-git-worktrees, isolated parallel work
  • dispatching-parallel-agents, orchestrate concurrent sub-runs
  • finishing-a-development-branch, clean shipping discipline
  • writing-skills, meta-skill for creating new skills
  • using-superpowers, entry-point skill that points the agent at the rest

Why it works

Three structural choices the README hammers home:

  • Skills, not prompts. Each skill encodes a workflow + quality gate, not a generic prompt. The agent activates skills automatically based on what you're doing.
  • Subagent-driven development. Tasks fan out to subagents with fresh context windows; the parent reviews their work. Long autonomous runs become tractable.
  • Methodology, not vibes. "Step back, spec, plan, then build" replaces "vibe code until something works."

Use for real software builds where you want the agent to spec → plan → TDD → ship, autonomously for hours. Strong default for greenfield and feature work alike.

Avoid for quick scripts where the install + methodology overhead outsizes the task.