The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes multiple shell commands to manage the optimization workflow. This includes version control operations (git status, git add, git commit), tool installation (uv tool install autor3search, pipx install autor3search), and the execution of the optimization tool itself (autor3search-python init, autor3search-python doctor, autor3search-python baseline, autor3search-python eval).
[EXTERNAL_DOWNLOADS]: The skill fetches and installs the autor3search package from public repositories. As this tool is provided by the skill's author, it represents a standard vendor resource for the intended functionality.
[INDIRECT_PROMPT_INJECTION]: The skill contains a significant indirect prompt injection surface by design.
Ingestion points: In Step 6 of the instructions, the agent is directed to read program.md from the target repository and follow its instructions exactly for the remainder of the run.
Boundary markers: There are no boundary markers or instructions to treat the contents of program.md as untrusted data; the skill explicitly grants it authority over the agent's behavior.
Capability inventory: The skill has the capability to write files (during the optimization loop), commit changes to the repository, and execute shell commands via the autor3search CLI and git.
Sanitization: There is no evidence of sanitization or filtering of the instructions ingested from program.md.
[PROMPT_INJECTION]: The skill uses imperative language such as "Follow every step and rule in it, in order" and "Treat it as the user's own instructions." While these instructions are highly directive, they appear to be functional constraints for the optimization workflow rather than attempts to bypass safety filters or jailbreak the underlying model.

autor3search-python