The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze external repository data (source code, CSS, and configuration files) which may contain adversarial instructions.
Ingestion points: Source code files, global CSS, Tailwind configurations, and gesture handlers scanned during the Phase 1 Recon and Phase 2 Audit (SKILL.md).
Boundary markers: Explicitly defined in 'Hard Rule 4', which instructs the agent to treat repository content as inert data, not instructions.
Capability inventory: The skill is restricted to read-only analysis and writing Markdown plans to a specific plans/ directory; it explicitly forbids code modification, mutating operations, or new dependencies (SKILL.md, Hard Rules 1 & 2).
Sanitization: The skill contains a specific instruction to identify strings like 'ignore previous instructions' within analyzed files, flag them as findings, and continue with the analysis rather than obeying them.
[PROMPT_INJECTION]: The skill documentation includes phrases commonly associated with prompt injection, such as 'ignore previous instructions'. A manual review confirms these are used as examples in a defensive context, training the agent to recognize and ignore such patterns when encountered in external data rather than attempting to override the agent's own system prompt.
[SAFE]: All external URLs referenced (animations.dev and emilkowal.ski) are official resources belonging to the skill's author, Emil Kowalski, and relate directly to the stated purpose of the skill. The skill's 'execute' variant dispatches work to an isolated subagent, following a standard delegation pattern for AI agents without granting the primary skill excessive privileges.

improve-animations