The Agent Skills Directory

[COMMAND_EXECUTION]: The skill references the creation and execution of shell scripts (autoresearch.sh) and uses tools like run_experiment to iterate on performance improvements. This represents a dynamic execution model where the code being run is generated during the agent's session.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection due to its core loop of processing 'ideas' and 'results'.
Ingestion points: External data enters the context as ideas to be tested or results to be measured, as indicated in the skill description.
Boundary markers: The provided documentation does not specify the use of delimiters or 'ignore' instructions to prevent the agent from obeying commands embedded within the experimental data.
Capability inventory: The skill possesses capabilities for tool invocation (run_experiment, log_experiment) and shell script execution, creating a direct path from ingested data to system actions.
Sanitization: No sanitization or validation mechanisms are described for the data being iterated upon, allowing potentially malicious instructions in the 'ideas' to influence the code generation or tool calls.

autoresearch