Hermes Wiki
TechResearch/Superpowers_Agent_Framework/research-superpowers

Superpowers (obra) — Agent Skills Framework, and What It Teaches About Personal Branding

github.com/obra/superpowers — MIT-licensed, ~277K stars, created 2025-10-09 (~10.5 months old). Install via /plugin install superpowers@claude-plugins-official in Claude Code, or per-harness equivalents (Codex, Cursor, Antigravity, Devin, Droid, Gemini CLI, Copilot CLI, Grok, Kimi, OpenCode, Pi, Hermes — 12 harnesses supported at once).

Flagged by Mihir while researching how to market/sell himself (interview positioning, personal brand, content creation) — not primarily for the tool itself, though it's a genuinely notable one. This note covers both: what it is technically, and the personal-branding playbook it's a live example of.

What it is

"A complete software development methodology for your coding agents, built on top of a set of composable skills." Concretely: a library of Agent Skills (see research_agent_skills_setup for the SKILL.md format itself) that auto-trigger through a fixed pipeline rather than firing ad-hoc:

  1. brainstorming — Socratic refinement of a rough idea into a spec, shown back in digestible chunks for sign-off before any code gets written.
  2. using-git-worktrees — isolated branch/workspace, clean test baseline verified before work starts.
  3. writing-plans — breaks approved design into 2-5 minute tasks, each with exact file paths and complete code.
  4. subagent-driven-development / executing-plans — a fresh subagent per task, two-stage review (spec compliance, then code quality), or batch execution with human checkpoints.
  5. test-driven-development — enforced RED-GREEN-REFACTOR; code written before its test gets deleted, not just discouraged.
  6. requesting-code-review — runs between tasks; critical issues block progress rather than getting silently noted.
  7. finishing-a-development-branch — verifies tests, then presents merge/PR/keep/discard options and cleans up the worktree.

The distinguishing claim: skills trigger automatically and are framed as mandatory workflow, not suggestions — the agent checks for a relevant skill before starting any task, rather than the user having to remember to invoke one. Philosophy stated plainly in the README: TDD always, process over ad-hoc guessing, complexity reduction as the primary goal, evidence over claims.

Compared to ECC (also in this vault, similarly single-maintainer and similarly starred): both are "don't trust the model to remember the process, make the harness enforce it" tools. ECC frames itself around five primitives (skills/agents/rules/hooks/instincts) and an explicit per-harness support matrix; Superpowers is narrower and more opinionated — one fixed methodology (brainstorm → plan → TDD → subagent execution → review → merge), not a general toolkit you assemble yourself. Worth a direct hands-on comparison later if either gets adopted for real work.

The personal-branding playbook this is a live example of

This is the part Mihir actually flagged it for. Superpowers isn't just a tool — it's a working instance of a specific, replicable brand-building pattern for a senior individual engineer:

  1. Built by one person with existing credibility, not a company from day one. Jesse Vincent ("obra") — known well before this from RT (Request Tracker) and the Perl community — released Superpowers under his own name and blog first. The README's "Community" section literally reads "Superpowers is built by Jesse Vincent and the rest of the folks at Prime Radiant" — person first, company second. Reputation compounds onto a name, not just a GitHub org.

  2. Give away something genuinely useful, for free, with no strings. MIT license, immediately installable, real production-grade methodology (TDD enforcement, code review gates) — not a lead-gen teaser version. The value has to be real before anyone will amplify it.

  3. Reduce install friction to near zero, everywhere your audience already is. 12 supported coding agents at launch-adjacent maturity (Claude Code, Cursor, Codex, Gemini CLI, Copilot CLI, Devin, Droid, Grok, Kimi, OpenCode, Pi, Hermes) — each with its own one-line install command in the README. The tool meets people in their existing tool, rather than asking them to switch to yours.

  4. Publish the origin story as content, not just the tool. The original release announcement on his own blog is referenced directly from the README's Philosophy section — the why behind the tool is itself a piece of content that drives adoption, separate from the docs.

  5. Convert organic community into an owned channel. A Discord server for "community support, questions, and sharing what you're building with Superpowers," plus a separate release-announcement mailing list — both explicitly built so future reach doesn't depend entirely on GitHub's algorithm or a single post going viral again.

  6. Monetize the credibility, not the tool itself. The tool stays MIT-licensed and free. The README's "Commercial Services" section points enterprise users to [email protected] for support/tooling/managed spend — Prime Radiant is the company built around the reputation Superpowers generates, not a paywall bolted onto the tool. This is the actual funnel: free tool → technical credibility at scale (277K stars) → inbound enterprise sales conversations that would otherwise take years of cold outreach to generate.

  7. Instrument what you can't otherwise measure — transparently. The README's own "Visual companion telemetry" section explains, in plain language, that stars/downloads don't tell you real usage, so an optional, disclosed, opt-outable telemetry ping exists specifically to answer "how many people are actually running this." Worth noting for anything Mihir builds and open-sources: say what you're measuring and why, out loud, in the README itself.

Applying this to Mihir's own positioning

The transferable shape, not a literal template: pick something Mihir already has real depth in (per Profile's trajectory — AWS/cloud infra, network engineering moving into platform/senior developer work), build one genuinely useful, freely-shareable artifact around it (a tool, a well-documented methodology, a strong technical write-up), publish the reasoning behind it under his own name first, and only think about monetization/consulting after the credibility exists — not as the opening move. The order matters more than the individual tactics: value and reputation come first, the sales conversation is a consequence, not the pitch.

Maturity / caveats

  • ~10.5 months old — the 277K star count is real (verified via gh repo view), but that's an extreme outlier for a single-methodology tool this young; worth periodically re-checking whether usage (not just stars) holds up, especially since the README's own telemetry section admits stars/downloads aren't a reliable usage signal.
  • "We don't generally accept contributions of new skills" — the skill set itself is curated/closed, not community-extensible in the way a plugin ecosystem usually is. The marketplace/installer surface is open; the methodology's actual content is not.
  • No hands-on trial yet in this vault — this note is desk research from the README, not a lived comparison against how Superpowers actually performs on a real task (unlike agent_harness_hands_on's benchmarked approach). Worth a real trial run if it gets adopted.

Sources

Hermes Wiki