The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: As a meta-skill designed to generate other skills, this skill naturally ingests user-provided intent and descriptions to produce instructional content.
Ingestion points: User responses during the 'Create new skill' and 'Create domain expertise skill' workflows (e.g., in workflows/create-new-skill.md).
Boundary markers: The skill heavily emphasizes the use of XML tags to delimit content and structure, which helps the agent distinguish between different sections of the generated skill.
Capability inventory: The skill uses file system operations (mkdir, cat, ls) and file writes (>) to create and manage skill directories and files, including creating slash commands in ~/.claude/commands/.
Sanitization: While it doesn't use programmatic sanitization (like regex), it provides clear guidelines on how Claude should structure generated content to avoid common pitfalls.
[DYNAMIC_CONTEXT_INJECTION]: The skill provides explicit guidance on how to avoid accidental execution of shell commands or file reads during skill loading. In references/common-patterns.md, it warns against the dynamic syntax and demonstrates safe ways to present these examples by adding a space.
[COMMAND_EXECUTION]: The skill utilizes standard shell commands for managing its own environment and the skills it creates (e.g., chmod +x for scripts, mkdir for directory structures, and cat for reading/writing configuration files). These are all within the expected scope of a developer-oriented skill.

create-agent-skills