The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is primarily designed to generate complex shell command strings (assertions) that are executed as part of an autonomous /goal loop. These assertions include file system checks, test suite executions, and CLI tool interactions.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from PRDs, GitHub issues, and local specification files to generate its output. A malicious document could attempt to inject shell commands into the acceptance criteria that the agent might then include in the generated /goal line.
Ingestion points: Arguments passed to the skill in SKILL.md, output from the gh issue view command, and contents of local files read via the Read tool.
Boundary markers: The instructions lack guidance on using delimiters or strict 'ignore' markers when processing the untrusted input content.
Capability inventory: The skill environment permits Bash, Write, Read, Grep, and Agent tools, enabling the execution of any generated shell strings.
Sanitization: No sanitization or validation logic is specified for the strings extracted from the input before they are formatted into the shell assertions.
[DYNAMIC_EXECUTION]: Section 8 of the documentation describes a pattern for the agent to spawn independent sub-agents using a shell command (claude -p --bare) that reads prompt content dynamically from a temporary file in /tmp.

prd-to-goal