The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes and iterates on external artifacts and configuration files, creating a surface for potential instruction injection.
Ingestion points: The agent reads the experiment definition from CARD.md and the target content from a "mutable surface" file (SKILL.md).
Boundary markers: The skill references an "autonomy gate" to verify the safety of the environment, but does not define specific technical delimiters or instruction isolation for the content being optimized (SKILL.md).
Capability inventory: The skill performs file writes (editing the mutable surface), version control operations (git checkout, git commit, git reset), and executes a local script (tools/experiment_log.py) (SKILL.md).
Sanitization: There are no explicit instructions for sanitizing or escaping the content of the artifacts being modified (SKILL.md).
[DYNAMIC_EXECUTION]: The skill is designed to have the agent iteratively modify a "mutable surface" (which may be code) and then execute a "read-only harness" to evaluate the changes.
Evidence: The workflow involves editing a target file and running it via a harness to generate a results.tsv log (SKILL.md).
[COMMAND_EXECUTION]: The skill utilizes shell commands to manage the experimentation lifecycle and logging.
Evidence: The instructions explicitly direct the agent to use git checkout, git commit, git reset, and to execute a local Python script via tools/experiment_log.py add <results.tsv> (SKILL.md).

experiment