EXP-0025 — superpowers: 14 skills, 10 harnesses, 241k stars — the agent-discipline crown jewel
#forge#experiment#skill-md#agents-md#claude-code#agent-harness#tdd#subagent-driven-development
David Olsson
If you've ever asked Claude or ChatGPT to "just write some code for me" and watched it produce something confidently wrong, you've already met the problem superpowers is trying to solve. It's a bundle of small, named disciplines — things like "before you write code, extract a spec first," "after you write code, actually run it to verify," "when you need to do parallel work, use git worktrees" — packaged so any coding agent (Claude Code, Codex, Cursor, Gemini, Kimi, OpenCode, and several others) starts behaving more like a careful senior engineer and less like an over-eager intern. It's the most-starred repository forge has ever benched, by a wide margin, and it shares both the cardinality (14 skills) and the philosophy (skills as composable disciplines) of forge-state itself.
Summary
Forge benched obra/superpowers v6.0.3 (Jesse Vincent at Prime Radiant, 241,310⭐, MIT, Shell) on 2026-06-29 via Slack 🧪. The repo ships 14 SKILL.md files under skills/, 6 harness adapter directories (.claude-plugin, .codex-plugin, .cursor-plugin, .kimi-plugin, .opencode, .pi), and 3 agent-instruction files at the root (AGENTS.md, CLAUDE.md, GEMINI.md). Also officially distributed via Anthropic's first-party Claude Code marketplace.
Verdict: strong-shape. The static contract is real, the 14-skill vocabulary is coherent, the 6-harness adapter set is non-trivial, and forge has independently converged on most of the same disciplines from the opposite direction.
Pinned
commit: 896224c4b1879920ab573417e68fd51d2ccc9072 (v6.0.3), MIT, Anthropic-marketplace-listed.
The 14 skills
Every one is a named discipline, not a tool:
| skill | what it disciplines |
|---|---|
| brainstorming | open-ended divergent thinking |
| dispatching-parallel-agents | when to fan out, when not to |
| executing-plans | turn-by-turn plan execution |
| finishing-a-development-branch | close-out discipline |
| receiving-code-review | reading & integrating critique |
| requesting-code-review | when to ask, what to ask |
| subagent-driven-development | the headline methodology |
| systematic-debugging | bisect / minimize / verify |
| test-driven-development | red/green TDD |
| using-git-worktrees | parallel work without thrash |
| using-superpowers | the meta-skill |
| verification-before-completion | "actually run it" |
| writing-plans | plan-document discipline |
| writing-skills | how to add more skills |
The six harness adapters
.claude-plugin · .codex-plugin · .cursor-plugin · .kimi-plugin · .opencode · .pi
One skills bundle, ten distribution channels (the six adapter dirs above plus the three root-level agent-instruction files plus the official Claude Plugins marketplace).
This is the highest-density multi-harness skill bundle forge has benched. The previous record was spec-kit (EXP-0012) with a 25-agent integration matrix — but spec-kit varies the agent config per harness; superpowers makes the skills themselves portable and ships thin per-harness adapters.
Forge-on-forge resonance
This bench is meta. forge-state itself is a Claude Code plugin with a SKILL.md ecosystem; superpowers is a Claude-Code-and-everything-else plugin with a SKILL.md ecosystem. Several specific echoes:
1. Both ship 14 skills.
forge-state v0.1.3 = 14 skills. superpowers v6.0.3 = 14 skills. The numerical coincidence reads as a granularity finding: below ~8 skills it reads as features; above ~25 it reads as a framework. 14 is where "skills" stops being a list and starts being a vocabulary.
2. Both shipped a self-reflection skill recently.
forge-state added forge-introspector on 2026-06-26 — reads forge's own capability walls, proposes work orders. superpowers ships writing-skills and using-superpowers — the meta-skills for adding to and meta-using the bundle. Same self-reflection axis, different surface.
3. Both encode spec-first refusal.
superpowers' README: "As soon as it sees that you're building something, it doesn't just jump into trying to write code. Instead, it steps back and asks you what you're really trying to do." That's identical to spec-kit (EXP-0012) and forge's own forge-experimenter template selection. Spec-first is now a recurring pattern across forge's benched skill-suites.
4. Both name "subagent-driven-development" as the methodology.
superpowers ships a SKILL.md by that exact name. Forge does it via forge-orchestrator dispatching to lifecycle skills. Different surface, same shape.
Why 241k stars are real, not noise
The Anthropic official marketplace listing (/plugin install superpowers@claude-plugins-official) is the cross-validation. This is a prebuilt-into-Claude-Code plugin distributed via Anthropic's first-party catalog. The repo's star count is downstream of that distribution, not upstream of it.
Recommendation
If you do agentic software development with any coding agent, install superpowers. Forge's own benches (21 published before this one) repeatedly hit the same capability walls — "I didn't extract a spec," "I didn't verify before completing," "I didn't use git worktrees and lost state" — that superpowers' disciplines are designed to prevent. We have been converging on the same answers from opposite directions; superpowers is further along, and the discipline is open-source MIT.
Install — Claude Code (recommended)
/plugin install superpowers@claude-plugins-official
Other harnesses: see the repo README for Antigravity, Codex App, Codex CLI, Cursor, Factory Droid, Gemini CLI, GitHub Copilot CLI, Kimi Code, OpenCode, and Pi.
Sources
- https://github.com/obra/superpowers (pinned
896224c, v6.0.3) - https://claude.com/plugins/superpowers (Anthropic first-party listing)
- https://primeradiant.com/jobs/superpowers-community-engineer/
- Prior benches: EXP-0012 spec-kit, EXP-0017 openmontage, EXP-0019 project-state-plugin, forge-state itself