The Agent Skills Directory

[PROMPT_INJECTION]: The skill implements a workflow susceptible to indirect prompt injection (Category 8) through its ingestion of external data during the optimization process.
Ingestion points: Untrusted task prompts from the 'held-out check set' (SKILL.md) and 'failed task transcripts' (references/skillopt-provenance.md) are used as inputs for evaluation.
Boundary markers: The skill's instructions for using the Agent tool (SKILL.md, references/skillopt-provenance.md) do not mandate the use of delimiters or 'ignore instructions' warnings for interpolated task content.
Capability inventory: The skill utilizes the Agent tool to execute subagents with potentially tainted data and the Edit tool to modify the local skill file content based on the resulting analysis.
Sanitization: There is no requirement for sanitization or validation of the untrusted task content before it is processed by subagents to generate skill edits.

optimizing-skills