The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes the claude CLI and the timeout (or gtimeout) utility via shell commands. It interpolates environment variables such as $MODEL, $SECS, and $PROMPT_FILE directly into the command execution string.
[INDIRECT_PROMPT_INJECTION]: The skill is explicitly designed to process external, caller-supplied prompt content from a file ($PROMPT_FILE), creating a surface for indirect prompt injection attacks.
Ingestion points: The content of the file specified by $PROMPT_FILE is passed as stdin to the claude -p process.
Boundary markers: The skill attempts to mitigate risks by disabling external settings (--setting-sources ""), slash commands (--disable-slash-commands), and restricting available tools (e.g., --tools "Read,Grep,Glob").
Capability inventory: The skill executes the claude process, which has the capability to read, write, and grep files, or execute shell commands if the Bash tool is enabled.
Sanitization: The skill relies on Claude's internal permission-mode (e.g., dontAsk, acceptEdits) and output format flags, but does not perform pre-processing or sanitization of the input prompt itself.

claude-exec