Superpowers

Executive Summary

Superpowers is a development methodology packaged as reusable skills for coding agents. Jesse Vincent and Prime Radiant develop the project, whose GitHub repository includes harness integrations and instructions for planning, testing, debugging, and review.^[1]

Verified September 15, 2026: the latest release is v6.3.0, published August 12. Its planning workflow explicitly scales the amount of process to the work, and repeated small edits can share one subagent dispatch. Those changes matter more than the popularity counters that dominated the previous profile.^[2]

Project attribution and distribution come from the repository; licensing comes from the license file.^[1][3]

---

Product Overview

Planning Scales with the Task

The current brainstorming skill separates three paths:^[4]

All three paths require explicit human approval before coding. The lighter paths reduce documentation ceremony; they do not remove the approval checkpoint.^[4]

Testing and Completion

The TDD skill calls for observing a meaningful failing test, writing the minimum implementation, and then refactoring. It distinguishes behavioral tests from assertions that merely exercise mocks, and asks the agent to identify which production change would make a test fail.^[5]

The completion skill separately requires fresh verification output. It distinguishes a passing linter from a successful build, a changed file from a fixed bug, and a subagent's success message from a reviewed diff. This is a useful review rubric even when a team adopts only part of the methodology.^[6]

---

Technical Architecture

Skill Loading and Host Integration

The bootstrap instructions tell the agent to load relevant skills before acting, while explicitly placing user instructions above skill instructions. This is behavioral guidance interpreted by the host model; describing it as mechanical enforcement overstates what the source establishes.^[7]

Superpowers is listed in Anthropic's official Claude plugin marketplace.^[8] Its installation command is:

/plugin install superpowers@claude-plugins-official

Install separately for each harness; the guide covers Claude Code, Codex, Cursor, Gemini CLI, Copilot CLI, Pi, Hermes, and others.^[1]

Integration does not imply identical capabilities. The Pi mapping requires an optional companion tool for subagents and falls back to sequential execution when it is absent. Task-list tooling is also optional.^[9]

Subagent Development

The current skill describes a fresh implementer for an independent task, one task reviewer covering specification compliance and code quality, and a broader review of the completed branch. It also allows one dispatch for a batch of small, similar edits.^[10]

Progress and review artifacts live in a plan-specific workspace under .superpowers/sdd/. A ledger supports resuming after context compaction. Fix rounds initially reuse the implementer; later rounds escalate to a fresh, more capable agent, with a five-round limit before adjudicating remaining findings. These are documented coordination rules, not measured guarantees of lower cost or fewer defects.^[10]

The README still uses the older “two-stage review” shorthand. The detailed skill above is the more specific source for the current combined task-review design.^[1][10]

---

Strengths

• Review criteria are inspectable. The completion skill demands concrete evidence for each success claim instead of accepting an agent's confidence.^[6]
• Tests are tied to behavior. The TDD guidance asks whether a test would detect a real production failure, which is more useful than maximizing test count.^[5]
• Planning has a smaller form. Bounded work can use an in-chat design, preserving a decision checkpoint without requiring architectural documentation for every edit.^[4]

Cautions

• Prompt compliance remains a dependency. Bootstrap instructions establish intended behavior, not a permission boundary or a proof that an agent followed every step.^[7]
• Review decisions still require judgment. After bounded fix loops, the controller can adjudicate findings and record unresolved items; completion should not be read as zero remaining concerns.^[10]
• Optional visual tooling makes a network request. The project says the brainstorming companion loads a Prime Radiant logo with a version identifier. It documents SUPERPOWERS_DISABLE_TELEMETRY and other opt-outs. This is the project's disclosure, not an independent traffic audit.^[1]

---

What Developers Say

In a June 11, 2026 field report, GitHub user sstoehrm described building three small example applications with v5.1.0. Reviews caught a vacuous test, a deserialization validation gap, and a version mismatch. The same run required roughly fifteen subagent dispatches, and writing plans took longer than executing any individual plan. Some prescribed code also proved incompatible with the actual toolchains.^[11]

That report is useful because it records both benefits and costs in a specific workload. It is not a benchmark of v6.3.0. The newer release's lighter planning and batched dispatch address the same class of overhead, but their effect on that workload has not been measured in this research.^[2]

---

Pricing & Licensing

Software: MIT-licensed, including commercial use and modification subject to notice requirements, with no warranty.^[3]

Services: Prime Radiant advertises commercial support, additional tooling, and managed spending; the repository does not list rates. Coding-agent subscriptions and model usage remain separate costs.^[1]

Competitive Positioning and Fit

Compare Superpowers with BMAD and Spec Kit in the category comparison. The editorial reason to evaluate Superpowers is its explicit treatment of development habits: how an agent investigates, tests, reviews, and demonstrates completion.

Best fit: Teams willing to inspect and tune an agent's process, then measure review effort and defect discovery on representative work.

Poor fit: Teams expecting skill installation alone to guarantee correctness or satisfy access-control requirements. Small tasks also need calibration so orchestration costs do not dominate the change.

Viability Assessment

The August release provides evidence of ongoing maintenance, including revised planning, worktree cleanup behavior, and additional harness support. Its changes also mean old tutorials can describe a different workflow. Evaluate a pinned version and read the relevant skill implementation when upgrading.^[2]

Bottom Line

Superpowers makes a coding agent's development process explicit and reviewable. The current version is less uniform about planning overhead than the old profile suggested. Its value still depends on the host agent following the process and the team checking whether that process improves delivered work.

---

Research by Ry Walker • methodology