The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains instructions that attempt to override the agent's decision-making process, using phrases such as 'IF A SKILL APPLIES TO YOUR TASK, YOU DO NOT HAVE A CHOICE. YOU MUST USE IT.'
[METADATA_POISONING]: The skill's description uses authoritative directives to force the agent to prioritize this specific sequence, claiming it 'MUST run before any other Designpowers skill.'
[PERSISTENCE]: The skill records user design decisions and 'taste' in a local file (~/.designpowers/taste-profile.md) to influence agent behavior in future projects.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted data from external sources, creating a potential for indirect injection attacks.
Ingestion points: Processes URLs, screenshots, code snippets, and remote DESIGN.md files through the design-review and design-library skills.
Boundary markers: The instructions suggest treating DESIGN.md files as data rather than instructions and mentions an 'injection guardrail' for remote content.
Capability inventory: The skill manages a team of agents that can perform file operations (design-state.md) and generate application code (design-builder).
Sanitization: The skill explicitly mentions a 'fetched file injection guardrail' for external resources.
[DATA_EXFILTRATION]: The skill performs sensitive file path access by reading the user's design profile stored in the home directory (~/.designpowers/taste-profile.md).

using-designpowers