The Agent Skills Directory

[PROMPT_INJECTION]: Indirect prompt injection surface. The skill is designed to ingest and process potentially untrusted external data such as error logs, raw data dumps, and user-provided reports to perform diagnostics.
Ingestion points: Processes error messages, raw data, and log files in Step 1 and Step 3.
Boundary markers: Instructions do not explicitly require the use of delimiters or 'ignore embedded instructions' markers when handling this data.
Capability inventory: Capability to write files to the .issues/ directory and execute shell commands via git.
Sanitization: No explicit sanitization or filtering of external content is mentioned before the agent processes or interprets the data.
[COMMAND_EXECUTION]: Modifies local environment configuration. The skill executes shell commands to check git status and modifies the local git exclusion file.
Evidence: Executes git check-ignore .issues and appends .issues/ to .git/info/exclude in the SKILL.md instructions.
Context: This is used to ensure that internal diagnostic drafts are not accidentally committed to a repository, which is a benign functional behavior but involves modifying environment metadata.

dig