The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest untrusted data from user requests and external research sources to generate new prompts.
Ingestion points: Processes user queries, local project files, and external search results (GitHub, Reddit, Official Docs) as defined in the Fetch and Deep Research stages in SKILL.md.
Boundary markers: The skill explicitly instructs the agent to generate output containing Stop conditions, Decision standards, and Verification fields to bound future execution.
Capability inventory: While the skill itself only generates text, the instructions it produces guide other agents to perform file system operations, command execution, and network tasks.
Sanitization: The instructions mandate an Evidence Map and Intent Alignment to validate external data before incorporating it into strategic decisions. Given its role as a prompt-engineering utility, this surface is managed through structural guardrails.
[EXTERNAL_DOWNLOADS]: The skill references several external sources for information gathering and research purposes.
Trusted Sources: References official documentation and research from known entities such as OpenAI, Anthropic, and GitHub (e.g., github.com/yaojingang/yao-meta-skill, platform.claude.com).
Well-Known Services: Utilizes standard community platforms like Reddit and X (Twitter) for capturing 'failure modes' and 'practice trends' as metadata for prompt generation.
[SAFE]: The skill maintains a strict 'prompt-only' boundary. The SKILL.md file contains explicit instructions (e.g., 'DEFAULT ONLY OUTPUT PROMPTS, DO NOT EXECUTE') to prevent the agent from performing unauthorized actions or persistence mechanisms without specific user intent. No hardcoded credentials or obfuscated commands were found.

goalpro