The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructions direct the agent to fetch and load methodology from an external GitHub repository if the local version is unavailable. This creates a dependency on untrusted remote content.
Evidence: references/empirical-tuning-gate.md references https://github.com/mizchi/skills/blob/main/empirical-prompt-tuning/SKILL.md as an external methodology source to be loaded by the agent.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process and analyze a wide array of untrusted project artifacts, including logs, code, and history, which provides a significant surface for adversarial content to influence the agent's behavior.
Ingestion points: The skill analyzes instructions, source, tests, scripts, commands, logs, errors, reviews, history, and outputs from potentially untrusted projects.
Boundary markers: The instructions lack specific delimiters or guardrails to prevent the agent from following instructions embedded within the data being analyzed.
Capability inventory: The skill produces and validates new agent capabilities, which may involve file writing and execution via a 'fresh executor'.
Sanitization: There is no process defined for filtering or sanitizing the input data to remove potential prompt injections.

extracting-agent-skills