The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes shell commands specified in the Verify field of the PLAN.md file. This includes running tests, linters, and other utility commands like pnpm, grep, and ls.\n- [PROMPT_INJECTION]: Instructions in SKILL.md explicitly state to 'Proceed immediately — no confirmation needed' once the execution order is determined. This bypasses the typical security practice of requiring user approval before executing shell commands or making significant file changes.\n- [REMOTE_CODE_EXECUTION]: The combination of writing code (based on the Do field) and executing commands (based on the Verify field) from an external plan allows for arbitrary code execution. If the plan is generated from untrusted data, this constitutes a significant risk.\n- [DATA_EXFILTRATION]: The skill's ability to read arbitrary files listed in the plan's Files field, combined with command execution capabilities, provides a pathway for data exfiltration if a malicious plan is processed.\n- [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection via the plan files it processes.\n
Ingestion points: .context/*/PLAN.md and .context/*/DECISIONS.md.\n
Boundary markers: None identified. The skill follows plan instructions directly.\n
Capability inventory: git commit, git add, shell command execution, and file writing.\n
Sanitization: None identified.

go