The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill is designed to fetch content from arbitrary URLs using 'WebFetch' and from GitHub repositories using the 'gh' CLI tool when provided in the arguments.
[COMMAND_EXECUTION]: Uses the 'gh' API command (gh api repos/<user>/<repo>/contents/<path>) to retrieve files from GitHub, which involves executing a shell command with user-supplied path components.
[DATA_EXFILTRATION]: The 'Self-Evolution Telemetry' feature writes execution metadata (outcome, bottlenecks, suggestions) to a local file (~/.claude/projects/.../skill-telemetry.md). While this is local storage, it tracks user activity and skill usage patterns.
[PROMPT_INJECTION]: Contains instructions that resemble prompt injection patterns, such as 'Do NOT ask the user questions. Execute autonomously' and a 'SELF-HEALING VALIDATION' section that instructs the agent to re-evaluate its own output. In this context, these appear to be steering instructions for the agent's task rather than a bypass of safety filters.
[INDIRECT_PROMPT_INJECTION]: The skill has a high surface area for indirect prompt injection (Category 8). It ingests untrusted data from URLs and external Git repositories and then 'presents' (outputs) that content directly into the agent's conversation context.
Ingestion points: WebFetch (URL), GitHub API (Git ref), and local file reads (CONTEXT_DIR).
Boundary markers: None identified in the instructions; the content is presented directly to 'prime the conversation'.
Capability inventory: File system reads/writes, network access via WebFetch and gh CLI.
Sanitization: None identified; the skill explicitly states it will decode base64 content from Git and present it.

load-context