SuperPowers
Give Claude Code a spine.
14 skills. One plugin install. Your coding agent goes from eager junior to disciplined engineer.
The problem
Vanilla Claude Code is brilliant and undisciplined.
The model is great. What's missing is the scaffolding a senior engineer takes for granted.
It jumps straight to code
You ask for a feature. It starts typing files before asking what you're actually building.
It re-explains itself every session
Skills, conventions, style: all wiped at /clear. You become a human README read aloud.
It claims done without checking
"It should work now." Did you run the tests? Did you read the output? Probably not.
It writes tests after the fact
Or never. Vanilla Claude treats TDD as a suggestion. Superpowers treats deleting your pre-test code as a feature.
It debugs by vibes
Hypotheses stacked three deep, none verified. Symptom fixes. Shotgun refactors.
It reviews its own work
Single model, single pass, "looks good to me." No skeptical second reader.
How it works
From prompt to PR
Watch the full Superpowers workflow, or step through it manually.
15 steps from prompt to merged PR
Philosophy
Four beliefs, enforced.
Not aspirations. Not best practices. Iron laws with delete buttons and gates that stop the agent mid-turn.
Test-Driven Development
Write tests first. Always. Code written before a failing test gets deleted — not kept 'for reference.'
Systematic over ad-hoc
Process over guessing. Four debugging phases, one hypothesis at a time, no vibes-based fixes.
Complexity reduction
Simplicity as the primary goal. YAGNI, DRY, and tasks small enough a junior can follow them.
Evidence over claims
Run the command. Read the full output. Then claim the result. "Seems," "probably," "should" are red flags.
These aren't aspirations. They're gates. The agent is stopped, mid-response, until the rule is satisfied.
Install
One command. Done.
Pick your host, copy the command, reload. Skills are active on the next session.
The Anthropic marketplace. Single command.
What happens next
Architecture
Five layers. One hook to rule them all.
No runtime orchestrator. The LLM steers itself, guided by the session-start injection and the 1% rule.
What's not here:
Skills catalog
All 14 skills, grouped by what they actually do.
Tap a card for the trigger, iron law, collaborators, and the files it ships. Every claim is pulled from the actual SKILL.md in obra/superpowers.
Vanilla vs Superpowers
Same prompt. Different species of output.
The prompt below goes to vanilla Claude Code and to Claude Code with Superpowers installed. Watch the difference unfold.
Build me a rate limiter for our API.
Waiting for prompt...- No clarifying questions
- No failing test first
- No persistence / scale discussion
- "Let me know if you want tests!"
Honest caveat: Superpowers isn't for everything.
- One-off throwaway scripts? Overkill. Skip the skill.
- Your team doesn't want TDD? The iron law will fight you.
- No subagent support (e.g. Gemini CLI)? Two-stage review collapses to one pass.
How it compares
The discipline slot,
not the capabilities slot.
Different tools, different angles. Superpowers gives your agent judgment — pair it with anything.
Superpowers
this pageEngineering methodology
Teams wanting disciplined agent-coding habits
Get Shit Done (GSD)
Spec-driven phase machine
Solo devs shipping long features reliably
anthropics/skills
The Skill format itself
Learning what a Skill is
wshobson/agents
Domain specialists (breadth)
On-demand domain experts
Impeccable
Frontend design quality
Designers fighting AI-ugly UI
claude-code-templates
Meta-installer / dashboard
Discovery / multi-source setup
Your workshop host
This page is a Datasense artifact.
Datasense runs hands-on workshops for teams adopting agentic coding. We distill frameworks like Superpowers down to what works in production, teach the mental models, and walk teams through real features together.
Resources & FAQ
Go deeper. Verify everything.
Frequently asked