The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's primary function is to audit and refactor prompts and instructions, which involves ingesting untrusted data from the repository or user requests. * Ingestion points: The instructions in SKILL.md and instruction-design.md specify that the agent should read 'target artifacts' and 'instruction stacks' which may contain malicious commands. * Boundary markers: The documentation suggests the use of 'completion boundaries' and 'selective reading' to manage context, but no programmatic delimiters are implemented in the skill body. * Capability inventory: The skill does not define or use any tools; it relies on the agent's pre-configured toolset. * Sanitization: No sanitization, escaping, or filtering mechanisms are described for the ingested text.
[PROMPT_INJECTION]: The prompt-patterns.md file includes a template for 'Conflicting Loaded Instructions' that directs the agent to 'identify the file, quote the relevant passage, and explain' the guidance if it prevents an action. While presented as a tool for resolving developer conflicts, this pattern is structurally identical to prompt extraction techniques used to reveal hidden instructions.
[NO_CODE]: The skill consists entirely of Markdown documentation, templates, and reference guides; it does not distribute any executable binaries, scripts, or package manager files.

gpt6-prompting-guide