⚠️ Version / star counts may change over time
Superpowers is an Agentic Skills Framework and a software-development methodology built for AI coding agents. It packages a whole set of engineering disciplines — TDD, systematic debugging, structured planning, code review — into plain-text C1 files, turning AI from “an assistant that writes whatever you ask” into “a disciplined senior engineer.”
🧭 Part 1: Background & Overview #
| Item | Detail |
|---|---|
| Author | Jesse Vincent (GitHub: obra), Prime Radiant team |
| Release | October 2025 (first blog post) |
| Repository | C1 |
| GitHub Stars | ~290k (2026-09, one of the fastest-growing developer tools ever) |
| Forks / Contributors | 25.9k forks / 51 contributors |
| License | MIT (open source) |
| Latest version | v6.4.1 (2026-09-19) |
| Core composition | 14 C1 skill files + Session Hook + multi-platform manifests |
Author background #
Jesse Vincent is no newcomer — he has long built infrastructure tools others depend on:
- Request Tracker (RT): a 1990s ticketing system
- Perl 6: project manager from 2005–2008
- Keyboardio: co-founder (mechanical keyboards)
- K-9 Mail: Android email app, later acquired by Mozilla and renamed Thunderbird for Android
Core insight #
Superpowers starts from one counterintuitive claim:
What AI coding agents lack is not “capability” but “discipline”.
Models already know TDD and debugging, but when a shortcut is available they “convince themselves” to skip steps. Superpowers’ approach is to solidify discipline into plain text — no fine-tuning, no proprietary SDK.
🧩 Part 2: The 14 Core Skills #
Skills fall into four categories, each an independent C1 file:
🧪 Testing #
| Skill | Purpose |
|---|---|
| C1 | Strict RED-GREEN-REFACTOR cycle + anti-pattern catalog |
🐞 Debugging #
| Skill | Purpose |
|---|---|
| C1 | Four-phase root cause analysis (reproduce → collect data → analyze cause → verify fix) |
| C1 | Evidence-based completion verification — no claiming “done” without proof |
🤝 Collaboration #
| Skill | Purpose |
|---|---|
| C1 | Structured brainstorming; clarify requirements before coding |
| C1 | Break features into 2–5 minute tasks |
| C1 | Execute a plan that is already written |
| C1 | Dispatch a fresh subagent for each task |
| C1 | Orchestrate multiple agents in parallel |
| C1 | Proactively request code review |
| C1 | Handle review feedback |
| C1 | Isolate work with Git Worktree |
| C1 | Wrap up a branch: merge / open PR / clean up |
🔧 Meta-Skills #
| Skill | Purpose |
|---|---|
| C1 | Teach AI to write new skills (a skill for skills) |
| C1 | System bootstrap |
Two-tier skill architecture #
- Core Skills: shipped with the plugin, provide general methodology
- Personal Skills: stored in C1 , customizable; override core skills on match
🔄 Part 3: The Seven-Stage Development Flow #
From idea to delivery, the agent follows a structured pipeline instead of jumping straight into code:
| Stage | Skill | Key point |
|---|---|---|
| 1. Brainstorming | C1 | Socratic questioning to clarify the problem; produce a design doc; wait for approval section by section |
| 2. Worktree isolation | C1 | Set up an isolated workspace in a new branch; confirm existing tests pass first |
| 3. Write plan | C1 | Break into small tasks, each with exact file path + verification steps |
| 4. Subagent implementation | C1 | Dispatch a fresh subagent per task; two-stage review (spec + quality) |
| 5. TDD implementation | C1 | Write failing test → minimal code to pass → refactor → commit |
| 6. Code review | C1 | Categorize by severity (Critical / Major / Minor); Critical blocks progress |
| 7. Branch wrap-up | C1 | Verify all tests pass, then merge / open PR / continue |
💡 Key takeaway: never skip the human-review step of the Plan. Plan quality directly determines execution quality — time spent planning saves hours of rework.
⚖️ Part 4: Three Iron Laws & Red Flags #
Each strict skill opens with an uppercase, non-negotiable Iron Law, plus a “rationalization (Red Flags)” table to stop the agent from convincing itself to cut corners.
| Skill | Iron Law | Common excuses it blocks |
|---|---|---|
| TDD | NO PRODUCTION CODE WITHOUT A FAILING TEST FIRST | “the test passed on first run”, “keep deleted code for reference”, “just this once” |
| Verification | NO COMPLETION CLAIMS WITHOUT FRESH VERIFICATION EVIDENCE | vague words like “should”, “probably”, “seems to” |
| Debugging | NO FIXES WITHOUT ROOT CAUSE INVESTIGATION FIRST | treating symptom disappearance as fixed, skipping investigation and trying random fixes |
These aren’t motivational posters — they hard-encode a senior engineer’s code-review feedback into commands. The goal isn’t to “teach” the agent (it already knows), but to prevent it from self-persuading to violate discipline.
Real-world results #
- chardet 7.0.0: rewritten with Superpowers, 41× faster, 96.8% accuracy, fixing dozens of backlog issues; a full test suite covering 2,161 files and 99 encodings — a direct product of TDD.
🖥️ Part 5: Multi-Platform Support & Installation #
The same C1 directory runs across multiple agent hosts via thin manifests pointing to the same Markdown:
| Platform | Installation |
|---|---|
| Claude Code (deepest integration) | C1 |
| Claude Code (community marketplace) | /plugin install superpowers@superpowers-marketplace then C2 |
| Cursor | C1 or search the marketplace |
| Codex CLI | tell Codex to read C1 |
| Gemini CLI | C1 |
| OpenCode | C1 |
Also supported: GitHub Copilot CLI, Qwen Code, Goose CLI, Auggie.
How skills trigger #
- No need to memorize commands — the framework auto-detects the task type and activates the matching skill
- Or manually:
/superpowers:write-plan,/superpowers:execute-plan, C3 - Bootstrap is under 2,000 tokens; specific skills load on demand — very token-efficient
🎯 Part 6: Use Cases & Limitations #
✅ Best suited #
| Scenario | Why |
|---|---|
| Multi-file refactoring | needs planning + test protection |
| Production features needing test coverage | TDD enforced |
| Projects needing team consistency | unified methodology |
| Long-running migration tasks | subagent works autonomously for hours without drifting |
| Complex features needing architecture design | brainstorm first |
⚠️ Less suited #
| Scenario | Why |
|---|---|
| One- or two-line quick bug fixes | the discipline process is overkill |
| Prototypes / demos | no engineering discipline needed |
| Single-file small edits | same as above |
| Simple config updates | same as above |
🚧 Known limitations #
| Limitation | Note |
|---|---|
| Subagent context injection | subagents may not receive the C1 injection, causing inconsistent skill activation (known issue) |
| Model over-inference | Opus 4.5 sometimes “guesses” skill content from the description instead of reading it; mitigated by tweaking descriptions |
| Token consumption | a full brainstorm→plan→execute flow uses more tokens |
🔗 References #
| Source | Link |
|---|---|
| Official repository | https://github.com/obra/superpowers |
| Author’s first blog post (2025-10) | https://blog.fsck.com/2025/10/09/superpowers/ |
| Deep dive (Heyuan) | https://www.heyuan110.com/posts/ai/2026-02-01-superpowers-deep-dive/ |
| Marc Nuri’s review | https://blog.marcnuri.com/superpowers-claude-code-skills-framework |
| Termdock guide | https://www.termdock.com/en/blog/superpowers-framework-agent-skills |
| Chinese enhanced version | https://github.com/jnMetaCode/superpowers-zh |
Sources: official GitHub repository, the author’s blog and public reviews (September 2026)