The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes doc comments from user-supplied Go files, asking the agent to evaluate them "sentence by sentence" to classify and potentially delete them.
Ingestion points: Text content is extracted from files in the user-specified <files-or-dirs> via the scripts/scan.py utility.
Boundary markers: The instructions in SKILL.md do not specify delimiters or escaping mechanisms to separate the untrusted comment text from the agent's operational instructions, which could allow instructions hidden in comments to influence agent behavior.
Capability inventory: The agent utilizes an edit tool for file modifications and has the ability to run shell commands for scanning, building, testing, and git version control operations.
Sanitization: No filtering or sanitization of the comment content is performed before the agent analyzes it.
[COMMAND_EXECUTION]: The workflow relies on executing local shell commands, including python3 scripts/scan.py, git commit, and various build/test/lint tools. This exposes an execution surface that, while intended for development, could be abused if the agent's context is compromised via injection.

comment-trim