The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to execute shell commands to manage and invoke external CLI tools like Gemini and Codex. This includes environment checks such as which gemini, version validation using gemini --version, and complex command constructions involving stdin piping (e.g., codex exec ... < /tmp/doubt-prompt.md). These instructions involve interacting with the system shell and external binaries.- [DYNAMIC_EXECUTION]: The skill facilitates the execution of content through tools like codex exec. While the documentation emphasizes the use of a --sandbox read-only flag to mitigate risks, the core functionality involves passing potentially untrusted data to an execution environment for analysis.- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data referred to as ARTIFACT and CONTRACT to perform reviews. This introduces a risk where malicious instructions embedded in these inputs could attempt to influence the agent or the external reviewer tools. The skill notes this risk and suggests mitigations such as sandboxing and piping input to stdin to prevent shell metacharacter injection. Ingestion points: Data enters the context in Step 2 (EXTRACT) and Step 3 (DOUBT). Boundary markers: The instructions recommend explicit headers like ARTIFACT: and CONTRACT: to delimit data. Capability inventory: The skill utilizes shell execution, file system writes for temporary prompt files, and potential network access through AI service CLIs. Sanitization: The instructions advocate for read-only sandboxing and stdin piping to ensure input remains inert.

doubt-driven-development