The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests and analyzes untrusted data from logs/state.json and logs/events.jsonl, which record the inputs and outputs of previous workflow executions.
Ingestion points: logs/state.json and logs/events.jsonl located within the user-provided run_dir path.
Boundary markers: No explicit delimiters or instructions are provided to the agent to treat the log content as data rather than instructions, nor to ignore embedded commands within those logs.
Capability inventory: The skill has read access to historical run data and the capability to write improvement suggestions to improvements.md and propose/apply patches to SKILL.md files.
Sanitization: The process described does not include steps to sanitize, filter, or validate the log data before it is processed by the agent.

skill-improver