The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external data (task transcripts and check sets) to drive an optimization loop that modifies skill files. This creates a surface where malicious instructions in processed transcripts could influence the skill's behavior.
Ingestion points: Failed task transcripts and representative task prompts (SKILL.md, references/skillopt-provenance.md).
Boundary markers: No explicit delimiters or warnings to ignore embedded instructions are specified for the ingested data.
Capability inventory: The skill utilizes the Agent tool (for reflection and scoring) and the Edit tool (for modifying files).
Sanitization: No sanitization or validation of the ingested transcript content is performed before passing it to the optimizer subagent.
[DYNAMIC_EXECUTION]: The skill uses the Edit tool to perform runtime modifications of skill documents (including its own or other skills) based on content generated by an AI subagent.
Pattern: The skill translates LLM-generated reflection into literal string operations (replace, append, delete) applied to skill files via the Edit tool.

optimizing-skills