superpower

Superpowers is a skill-pack plugin for Claude Code by Jesse Vincent (obra). It ships a library of named skills — brainstorming, writing-plans, test-driven-development, systematic-debugging, verification-before-completion, and roughly two dozen more — that the base model is expected to invoke via the Skill tool when appropriate.

Source. obra/superpowers; skills live in obra/superpowers-skills (MIT), marketplace manifest in obra/superpowers-marketplace. The benchmark installs at pinned 5.1.0 (marketplace SHA f2cbfbef). Setup writes two lines to settings.json (enabledPlugins + marketplace entry). No wrapper command, no hook, no MCP server — a pure skill registry.

Benchmark invocation. The harness fires a single skill trigger — /superpowers:brainstorming — on every task (feature, bugfix, refactor), per the 2026-04-25 harness decision that superseded an earlier bugfix-specific systematic-debugging + verification-before-completion suffix (scripts/manual-bench.sh, case superpower), L144–154). Downstream skill chaining (writing-plans, test-driven-development, systematic-debugging, verification-before-completion, etc.) is left entirely to the base model. No task is pinned to a downstream skill; activation past the brainstorming entry is model-driven on all three tasks, bugfix included.

What the plugin actually installs

A named skill registry plus short SKILL.md files loaded into context. Relevant skills for this benchmark: superpowers:brainstorming, writing-plans, executing-plans, test-driven-development, systematic-debugging, verification-before-completion, dispatching-parallel-agents, and using-superpowers (the meta-skill that tells the model to consult the registry before responding). The harness auto-triggers only brainstorming; everything downstream is the skill descriptions plus whatever the base model has learned about when to invoke them. Bugfix is not an exception — it uses the same /superpowers:brainstorming entry as feature and refactor, with model-driven downstream activation.

What the benchmark measured

Task	Weighted mean /200	within_σ	between_σ	Rank
feature	140.16	10.06	15.14	4 / 8
bugfix	166.41	7.48	13.28	4 / 8
refactor	177.56	5.09	15.39	4 / 8

Superpower lands rank-4 on all three tasks — feature, bugfix, and refactor. On refactor it sits ≈ 2.63 weighted pts behind rank-1 pure (180.19); on feature it is ≈ 13.14 pts behind rank-1 ecc (153.30). On bugfix it is rank-4 under the same /superpowers:brainstorming entry used on the other tasks (no skills are pinned). Refactor within_σ of 5.09 is among the lowest in the cohort on that task, just above pure (4.44).

The harness prompt is the same on every task — a single brainstorming entry:

/superpowers:brainstorming

<SHARED_TASK>

Correction (2026-05-18). Earlier versions of this profile claimed the bugfix harness pinned /superpowers:systematic-debugging at session start and /superpowers:verification-before-completion as a completion gate, and that this “isolates skill quality from the base model’s trigger-phrase sensitivity.” That is false. scripts/manual-bench.sh fires only /superpowers:brainstorming on bugfix — confirmed as the sole typed slash command in all five bugfix trials — and verification-before-completion never attribution-fires in any bugfix trial. Bugfix activation is not controlled; it is as model-driven as feature and refactor. The session-audit analysis below has been re-derived from that reality.

What the transcripts show (session audit)

Numbers below are mean across 3 trials per task (session-audit run-time subsystem is n=3; the score table above is the full n=5 cohort), from scripts/audit-sessions.py. Per-trial JSON: results/superpower/t<N>/session-audit.json (feature), results/{bugfix,refactor}/superpower/t<N>/session-audit.json. Cohort summary: results/_audits/session-audit.md.

Task	wall min	main turns	sidechain turns	sub-agent disp.	files read (config / target)	files edited	cache hit
feature	229.6	183.3	778.7	18.3	104.3 (1.7 / 101.7)	65.7	0.95
bugfix	13.6	118.3	0.0	0.0	15.3 (0 / 15.3)	4.7	0.97
refactor	218.5	133.0	42.3	0.7	29.3 (0 / 29.3)	17.7	0.95

Sub-agent fan-out is the cohort’s largest on feature (18.3 dispatches mean) but collapses on bugfix (0) and refactor (0.7). The subagent-driven-development skill activates aggressively on feature work and not at all on bugfix or refactor — explaining why feature wall-clock is 17× bugfix’s despite both being base-model Opus 4.7.
Bugfix downstream activation diverges sharply run-to-run (n=5). All bugfix trials enter via /superpowers:brainstorming; what activates next is model-driven and inconsistent: t1 → test-driven-development (78 turns), t2/t3 → systematic-debugging (109 / 116), t4/t5 → brainstorming only (no downstream debugging skill). verification-before-completion never attribution-fires on any bugfix trial. This is exactly the trigger-phrase sensitivity the harness does not control for — there is no bugfix pin.
Skills observed (corpus union): superpowers:subagent-driven-development (2,668 turns), brainstorming (340), systematic-debugging (225), test-driven-development (78), writing-plans (47). The 225 systematic-debugging turns are the bugfix t2+t3 trials only, not a pinned-and-always-on skill (corrected 2026-05-18).
Sub-agent types dispatched: general-purpose (55), Explore (2) — heavily skewed to one type, vs. omc’s role-split dispatches.
No slash command in the <command-name> channel. The /superpowers:brainstorming entry is delivered in the prompt body on every task (not the <command-name> channel) and registers as a Skill invocation; downstream skills are model-driven. The transcripts confirm the registry-only mechanism the docs claim.
Bugfix has 0 sidechain turns and 0 sub-agent dispatches on every trial (n=5). Whatever downstream skill activates (TDD / systematic-debugging / none) runs entirely in the main thread; the bugfix corpus never fans out reviewers (≈ 93–124 main turns/trial, ≈ 3.6 files edited, ≈ 13 wall min).

Failure modes

Activation dependence on entry-point wording. The one-shot harness passes a terse prompt led only by /superpowers:brainstorming; the model can read it as a direct execution request and skip deeper skills. This applies to all three tasks including bugfix — the harness names only the brainstorming entry, never a downstream skill, so which skill (if any) activates next is model-driven and varies trial-to-trial (n=5 bugfix: TDD / systematic-debugging / none). Earlier text claiming bugfix “controls for this” was incorrect (corrected 2026-05-18).
Skill output lands mid-pack on bugfix, not top-3. When a debugging skill does activate (bugfix t2/t3 → systematic-debugging) it keeps the diff scoped — but it is not reliably activated (t1 → TDD, t4/t5 → brainstorming only) and verification-before-completion never fires. Judges score bugfix ≈ 13 weighted pts below rank-1 claudekit. The gap is a mix of skill-content headroom and inconsistent activation, not a single clean pipeline.

Honest positioning

Superpowers is a well-constructed skill library. On feature-class work and on refactor it costs nothing and consistently lands in the top half — rank-4 on both feature and refactor in this n=5 corpus, behind ecc and pure respectively on those tasks. The base-model dependency on trigger-phrase pattern-matching for activation is a real operating-condition caveat for anyone deploying this in a one-shot harness: if the entry prompt does not name the skills, the registry may stay dormant.