The Agent Skills Directory

[PROMPT_INJECTION]: The instructions contain the directive "Explicit user instructions override anything below.", which encourages the model to bypass the skill's logic and constraints upon user request.
[METADATA_POISONING]: The skill documentation includes references to non-existent future versions of AI models and tools, such as "Claude 5 generation models (Anthropic, 2026-07-24)", "Opus 5.5", and "GPT-6 Astra". This metadata is deceptive and may be intended to influence the model's persona or perceived capabilities.
[INDIRECT_PROMPT_INJECTION]:
Ingestion points: The skill is designed to fetch and process data from local project files (using rg and file reads) and external web sources (via the web-researcher helper defined in references/helper-protocols.md).
Boundary markers: While structured templates are used in references/helper-protocols.md (e.g., Objective:, Question:), they lack robust delimiters or specific instructions to treat external data as untrusted or to ignore embedded instructions.
Capability inventory: The agent has access to file system reading tools (rg, file reads) and external web research tools.
Sanitization: There is no evidence of content sanitization or validation for the data retrieved from external URLs or project files before it is processed by the AI.

meta-code