The Agent Skills Directory

[PROMPT_INJECTION]: The file skill-report.json contains deceptive metadata, including an author name ("Dami-Randle") that contradicts the expected author context ("aiskillstore") and references to non-existent AI models ("gpt-5.5"). It also makes self-referential safety claims ("risk_level": "safe", "Verdict: SAFE") which attempt to override independent security analysis.
[INDIRECT_PROMPT_INJECTION]: The skill facilitates the generation of other skills based on untrusted user input, creating a surface for indirect prompt injection.
Ingestion points: User-provided skill descriptions, requirements, and trigger words used in Step 1 and Step 6 of the SKILL.md instructions.
Boundary markers: Absent; the skill does not instruct the agent to use delimiters or specific warnings to ignore embedded instructions when incorporating user data into generated skill files.
Capability inventory: The skill provides instructions for performing file system operations (like mkdir) and generating functional agent instructions.
Sanitization: Absent; there is no guidance on validating, filtering, or escaping user-provided strings before they are inserted into the frontmatter or body of a new SKILL.md file.
[COMMAND_EXECUTION]: The documentation includes several examples of shell commands for directory management (mkdir -p ~/.claude/skills/), file inspection (ls, cat), and dependency management (pip install package-name). These are legitimate educational examples provided within the context of the skill's purpose as an authoring guide.

skill-writer