The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is designed to generate shell-executable assertions for verifying project state. These include file system operations like test -f, checking line counts with wc, searching files with grep, and performing Git operations with gh. It also utilizes rm to manage internal state files for verification loops.- [DYNAMIC_EXECUTION]: Section 8 defines a pattern for a "Post-timeout assertion grader" that dynamically spawns a sub-agent by executing the claude CLI through the Bash tool. The command CLAUDE_CODE_FORK_SUBAGENT=1 claude -p --bare "$(cat /tmp/grader-prompt.txt)" reads a prompt from a temporary file created at runtime and executes it in a fresh context.- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data to generate instructions for subsequent agent loops, creating a surface for injection attacks.
Ingestion points: External data is ingested from user-provided PRD text, GitHub issue bodies via gh issue view, and local specification files via the Read tool.
Boundary markers: The skill lacks explicit boundary markers or isolation instructions to prevent the agent from adopting commands or logic embedded within the source text.
Capability inventory: The environment allows the use of Bash, Write, Read, and Agent tools across all functional scripts.
Sanitization: There is no evidence of sanitization or input validation to filter potentially malicious instructions from the source documents before they are transformed into goal assertions.

prd-to-goal