The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill requires the agent to generate and execute shell scripts, including .auto/measure.sh, .auto/checks.sh, and lifecycle hooks, which are executed during the optimization loop. This dynamic generation of executable content based on the agent's interpretation of a project increases the risk of executing unintended logic.
[COMMAND_EXECUTION]: The run_experiment tool is used to execute arbitrary shell commands and the generated benchmark scripts. The skill instructions do not impose restrictions on the scope or nature of commands that can be executed.
[INDIRECT_PROMPT_INJECTION]: The skill defines a surface for indirect prompt injection because it reads external source files to understand a workload before generating execution scripts.
Ingestion points: Project source files are read and processed by the agent.
Boundary markers: The instructions do not specify the use of delimiters or 'ignore' instructions to separate project data from agent commands.
Capability inventory: The agent has the ability to write files to the .auto/ directory, execute shell commands, and manage git branches.
Sanitization: There are no requirements to sanitize or validate the content extracted from source files before it is used to construct benchmark and check scripts.
[PROMPT_INJECTION]: The skill contains instructions that explicitly bypass standard agent safety and oversight protocols by commanding the agent to 'LOOP FOREVER', 'Never ask should I continue?', and 'NEVER STOP'. These directives are designed to suppress user interaction and maintain autonomous execution without review.

autoresearch-create