Obra Superpowers: fourteen skills, one disciplined workflow.

Four categories, four kinds of leverage

These categories follow the repository README’s organization; they are descriptive groupings, not scanner risk categories.

Pre-install screen

Static analysis surfaced leads, not a verdict.

Each skill directory was scanned separately with NVIDIA SkillSpector 2.10.0 in static-only mode. The reports contain 46 findings across the 14 skills. Three skills reached the scanner's high or critical bands; all 14 reports are retained below as JSON.

The companion OWASP high-signal regex triage found no matches for direct “ignore previous instructions,” developer-mode, system override, prompt-leakage, API-key, or numbered-instruction patterns. OWASP treats those expressions as leads, not clearance; indirect, encoded, persistent, and tool-mediated attacks still need human review.

Per-skill results, with the JSON beside them

Scores below reproduce the machine-readable results from the exact commit. Use the JSON link for finding IDs, locations, explanations, confidence values, analyzer status, and completeness details.

The scanner’s low-score output still recommends CAUTION for this static-only run. That conservative label is preserved exactly; it is not rewritten as “safe.”

Human context

What the high scores appeared to be seeing

The scanner correctly points a reviewer toward sensitive surfaces, but its patterns do not understand intent on their own. A first context pass found these useful distinctions:

• brainstorming: local companion-server lifecycle code uses background process management, state files, and a sudo lsof check; an HTML comment was ordinary branding. Those are legitimate development-server behaviors that still deserve explicit permission and persistence review.
• systematic-debugging: credential and keychain commands appear in debugging examples, while instrumentation examples show logging. They are documentation leads, not confirmed credential theft or exfiltration, but real credentials should never be copied into the workflow.
• writing-skills: the skill teaches authoring, editing, testing, and enumerating other skills and agent configuration directories. That explains much of its high score, but it also makes least privilege and trust boundaries important for any skill-authoring workflow.
• The remaining findings: most are process instructions about autonomous execution, branch operations, human checkpoints, or review verdicts. Those can be useful capabilities while still requiring a human gate for destructive commands, pushes, credentials, and broad file access.

Origin story

The research thread that started this review

This guide also preserves the visible contents of the user-originated shared ChatGPT thread that asked whether Superpowers had been evaluated, how effective it was, and whether it worked equally well in Codex and Claude.

As an AI engineer, review the Obra superpowers skills (see GitHub). Has anyone online does evaluation or reviews of these skills? How effective are they? Are they equally useful for Codex as well as Claude?

The archived response’s central conclusion was selective use: Superpowers is best understood as an agentic software-development operating procedure, with its strongest value in process discipline rather than proven model intelligence. It described Claude Code as the more native target, while treating Codex as a viable but more explicit environment for subagent-heavy workflows.

The transcript is historical source material, not an independent benchmark or a replacement for the commit-pinned static review above. Product documentation and research claims can change.