Superpowers (Jesse Vincent)
github.com/obra/superpowers · 149K stars (Apr 2026)
What it is
Section titled “What it is”A skills framework + software development methodology by Jesse Vincent. Built on top of Claude Code with a hard 7-phase workflow: brainstorm → plan → write spec → write failing tests (TDD) → spin up subagents to implement → review → finalise.
The core rule: production code only exists to make a failing test pass.
Core skills (auto-triggered)
Section titled “Core skills (auto-triggered)”| Skill | Phase | What it does |
|---|---|---|
brainstorming | Design | Refines rough ideas through Socratic questions |
using-git-worktrees | Setup | Creates isolated worktree on new branch |
writing-plans | Planning | Breaks work into 2-5 minute tasks with exact file paths |
subagent-driven-development | Implementation | Dispatches fresh subagent per task with two-stage review |
test-driven-development | Implementation | RED-GREEN-REFACTOR enforced |
requesting-code-review | Review | Reviews against plan, blocks on critical issues |
finishing-a-development-branch | Finalise | Verifies tests pass, presents merge/PR options |
What we kept
Section titled “What we kept”- TDD iron law. Borrowed verbatim. Added to validator prompt as a HARD RULE: a behavioural assertion needs (a) a test that failed before the change, (b) executable evidence of the feature working, or (c) explicit “no behavioural test possible” acknowledgment.
- Subagent-driven development. Spirit of “fresh subagent per task” maps cleanly to our fresh-context-per-role.
- Git worktrees. They use this for isolation; we adopted the same with our
useWorktreesconfig.
What we dropped
Section titled “What we dropped”- The 7-phase rigid workflow. Our orchestrator picks the next role dynamically based on state, not a fixed sequence. A feature might skip planning if it’s a F-FIX-* issue conversion.
- Single mega-orchestrator. Their orchestrator is one Claude Code session that delegates to subagents. Ours is a
claudesubprocess emitting one decision verb per iteration. - 41× perf claim. Real but their context (chardet rewrite). Doesn’t generalise to our missions.
Differences in philosophy
Section titled “Differences in philosophy”| Superpowers | Papercup |
|---|---|
| 7 fixed phases, gates between | Dynamic decision verbs (orchestrator picks) |
| Mega-orchestrator + subagents | Fresh-context per role, no parent agent |
| TDD enforced at execution time | TDD enforced at validation time |
| Skills auto-trigger via Claude’s reasoning | Roles trigger via orchestrator decision verb |
Honest take
Section titled “Honest take”Superpowers is the most opinionated framework we surveyed — and that’s why so much transferred. The TDD iron law is the single biggest quality lever we have. We owe Jesse for making this case so loudly that it became impossible to ignore.
The 41× chardet rewrite story is real and instructive: when every change has a failing test as a safety net, the agent optimises aggressively because it can’t break anything silently.