The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill analyzes local project files such as package.json and README.md during its Analysis Mode to automatically infer agent responsibilities and technical stacks. This functionality creates an attack surface where malicious instructions embedded in project metadata could influence the generated agent definitions. \n- [INDIRECT_PROMPT_INJECTION]: Ingestion points: The skill reads package.json, README.md, tsconfig.json, the src/ directory, .github/workflows/, and existing AGENTS.md files as specified in the SKILL.md Analysis Mode section. \n- [INDIRECT_PROMPT_INJECTION]: Boundary markers: The skill includes a mandatory instruction to confirm all analyzed findings with the user (e.g., 'Does this look correct?') before writing the agent definition to the filesystem. \n- [INDIRECT_PROMPT_INJECTION]: Capability inventory: The skill is capable of directory creation (presets/) and writing markdown files (AGENTS.md) to the local filesystem. \n- [INDIRECT_PROMPT_INJECTION]: Sanitization: The skill relies on manual user verification of the inferred context to prevent the propagation of malicious or incorrect instructions from the source project files.

agent-creation