The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external sources to generate the core logic and instructions for new AI personas. There is a risk that source material (URLs, files, or pasted text) could contain hidden instructions that influence the behavior of the resulting AI soul.
Ingestion points: Processes content from YouTube transcripts, blog URLs, social media exports, and local file paths provided by the user in Step 2.
Boundary markers: Absent. The instructions do not define delimiters or provide guidance to the agent to ignore potential commands embedded within the raw source material.
Capability inventory: The skill performs file writes to create a new directory structure (Step 3) and generates an executable SKILL.md (Step 4).
Sanitization: Absent. The skill explicitly favors "original quotes" (Step 3/Pass 1) and "hard boundaries" distilled from the text, which could inadvertently include malicious constraints placed by an attacker in the source material.
[COMMAND_EXECUTION]: The skill instructions direct the agent to utilize specific Python scripts (collectors/youtube_transcript.py, collectors/fetch_url.py, collectors/jike_export.py, and collectors/twitter_archive.py) to perform data collection. These scripts are not provided in the analyzed file and their internal safety cannot be verified.
[DYNAMIC_EXECUTION]: The primary purpose of the skill is to generate a functional SKILL.md file and supporting markdown instructions. This generated content is intended to be installed and executed as an AI agent skill, representing a path from untrusted input to executable instructions.
[EXTERNAL_DOWNLOADS]: The collection process involves fetching content from arbitrary URLs (Step 2), including third-party web content and media transcripts, which are then stored locally in the _raw/ directory.

create-soul