The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill is configured to fetch instruction files (SKILL.md) and supporting references from external, third-party GitHub repositories managed by individual users (emilkowalski and jakubkrehel). These accounts do not fall within the scope of recognized trusted organizations or well-known services.
[INDIRECT_PROMPT_INJECTION]: By fetching and incorporating external instructions into the agent's workflow, the skill creates an attack surface for indirect prompt injection, where a compromise of the remote repository could lead to the agent receiving malicious instructions.
Ingestion points: references/motion.md and references/upstream.md direct the agent to fetch markdown files from github.com/emilkowalski/skills and github.com/jakubkrehel/skills.
Boundary markers: The skill suggests pinning to a specific commit for consistency, but it does not specify the use of delimiters or structured parsing to isolate these external instructions from the main system prompt.
Capability inventory: The agent uses tools like chrome-devtools-cli and hitch for environment inspection and has the ability to read and write project files such as 06-design.md and 00-project.md.
Sanitization: The skill lacks automated sanitization or validation of the fetched content, relying instead on high-level textual warnings advising the agent not to execute scripts or start subagents.

webcraft-design