The Agent Skills Directory

[COMMAND_EXECUTION]: The skill includes a self-validation script in references/guide.md that utilizes standard Unix utilities such as head, grep, test, and sed. These commands are designed to verify the presence of mandatory file headers, directory structures, and documentation tables within the skill's own directory. The operations are read-only and localized to the skill's workspace.
[INDIRECT_PROMPT_INJECTION]: The skill facilitates the creation of new agent instructions based on user requirements, which represents an ingestion surface for potentially untrusted data.
Ingestion points: User prompts triggered by phrases like "create a skill" or "new skill" (defined in SKILL.md).
Boundary markers: The skill employs explicit "Escalation triggers" that pause execution for human review when design tradeoffs or scope changes occur.
Capability inventory: The skill utilizes file writing to generate documentation and shell execution for running evaluation checks.
Sanitization: The process relies on an iterative "co-development flywheel" and human validation via the AskUserQuestion tool to ensure generated content aligns with intended safety and functional requirements.
[DYNAMIC_EXECUTION]: The skill framework encourages the use of "Runnable checks" for evaluation, such as grep for pattern violations and jq for schema validation. This is a standard functional component of a development-oriented skill and is implemented with specific criteria for success and failure.

skill-crafting