The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill has an ingestion surface as it processes untrusted user-provided text and files to generate refined prompts. However, it lacks exploitable capabilities and includes robust mitigation strategies.
Ingestion points: User request text and explicitly supplied files (referenced in SKILL.md).
Boundary markers: The references/improvement-method.md file specifically instructs the agent to 'Delimit inputs when instructions, source text, examples, or data could be confused with one another' in the output.
Capability inventory: The skill is restricted to text generation only. It explicitly forbids browsing, searching, fetching URLs, or calling retrieval tools. No subprocess execution or file system writes are permitted.
Sanitization: The methodology includes a mandatory self-review step to ensure intent preservation and to check that the resulting prompt is 'isolated' from meta-commentary.
[SAFE]: The skill enforces a strict 'no-execution' policy for the generated content, requiring the agent to return the artifact in a code block rather than executing it. This prevents the immediate exploitation of any malicious instructions that might be embedded in the user's input.
[SAFE]: The documentation explicitly prohibits external research or retrieval, ensuring that the skill remains localized and does not leak context to external domains.

improve-prompt