The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and transform untrusted user-supplied text (such as blog posts, emails, and marketing copy) to improve its tone and style. It lacks explicit boundary markers or instructions to the AI agent to ignore directives embedded within the target text, creating a surface for indirect prompt injection. Additionally, the skill mandates loading several files from a references/ directory that are missing from the package, which may lead to hallucinations or unpredictable behavior when the agent attempts to process these instructions.
Ingestion points: User-provided prose and documents processed during the humanization phases.
Boundary markers: Absent; the system instructions rely on open placeholders without delimiting or protective instructions.
Capability inventory: The skill itself does not define external tools or scripts, but successful injection could potentially target other tools available to the agent environment.
Sanitization: No sanitization or filtering of input text is defined.
[METADATA_POISONING]: The skill's YAML frontmatter in SKILL.md identifies the author as ice-ninja, which contradicts the expected author identity of neversight provided in the skill context. This discrepancy in metadata can mislead users or agents regarding the skill's origin and trust level.

humanize-writing