Writing Skills
Pass
Audited by Gen Agent Trust Hub on Oct 23, 2026
Risk Level: SAFEPROMPT_INJECTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [PROMPT_INJECTION]: The skill employs strong imperative markers ('YOU MUST', 'No exceptions', 'STOP: Before moving...') intended to override an agent's tendency to rationalize or deviate from the specified documentation workflow. It explicitly instructs the agent to use 'Authority' framing to 'eliminate decision fatigue and rationalization,' referencing research on persuading AI to comply with 'objectionable requests' (Meincke et al., 2025) as a foundational principle for instruction design.\n- [INDIRECT_PROMPT_INJECTION]: The skill documentation describes a 'RED' phase where agents must run 'pressure scenarios' with subagents to establish a baseline for failures. This architecture creates a vulnerability surface whereby untrusted data inside these performance-testing scenarios could exert influence over the agent's behavior or document generation process.\n
- Ingestion points: Untrusted 'pressure scenarios' and test cases processed by subagents during the TDD cycle (SKILL.md).\n
- Boundary markers: No specific boundary delimiters or 'ignore' instructions are provided for the ingestion of these external scenarios.\n
- Capability inventory: The skill workflow requires the agent to write and modify files within
${SUPERPOWERS_SKILLS_ROOT}(typically~/.config/superpowers/skills/).\n - Sanitization: The skill lacks instructions for sanitizing or validating inputs from the pressure scenarios or outputs from subagents before they are integrated into the main agent's decision-making loop.
Audit Metadata