The Agent Skills Directory

[COMMAND_EXECUTION]: The skill orchestrates the development lifecycle by executing local Python utilities and the claude CLI via the subprocess module to perform evaluation runs, analyze transcripts, and package skills into distributable formats.
[EXTERNAL_DOWNLOADS]: Retrieves required client-side assets for the evaluation viewer from well-known technology services, including the SheetJS library (cdn.sheetjs.com) and Google Fonts.
[PROMPT_INJECTION]: As a potential surface for indirect prompt injection, the skill processes untrusted test prompts and feedback data which are used to influence the iteration and grading loop.
Ingestion points: Reads task prompts from evals/evals.json and human feedback from feedback.json during the evaluation and improvement cycles.
Boundary markers: Employs YAML frontmatter delimiters and block scalar formatting in temporary command files to isolate instructions from data.
Capability inventory: Possesses the ability to execute shell commands via the claude CLI and performs extensive local filesystem operations.
Sanitization: Implements YAML validation for metadata and uses standard HTML escaping for data rendered in the review interface.
[DATA_EXFILTRATION]: Reads and displays workspace file content within a local environment. The evaluation viewer initiates a local server on 127.0.0.1 for human inspection of outputs and does not exhibit patterns of unauthorized remote data transmission.

skill-creator