The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides an 'Analyze' workflow that ingests data from local files, pasted text, and external websites. This creates a surface for indirect prompt injection where malicious instructions hidden in the corpus could influence the agent's behavior.\n
Ingestion points: The SKILL.md file defines a process for reading local file paths using Read/Glob and fetching remote content using WebFetch.\n
Boundary markers: There are no instructions to use specific delimiters or protective prompts (e.g., 'ignore embedded instructions') when processing the ingested text.\n
Capability inventory: The agent possesses file system access (Read, Glob, Write to references/voices/) and network access (WebFetch).\n
Sanitization: The skill does not include steps to sanitize or validate the content of the corpus before it is analyzed and incorporated into voice files.\n- [OBFUSCATION]: The reference file references/voices/su-that-man.md documents several instances of character-level obfuscation used by the target author to evade automated filters. Examples include using dots to split words ('b.ốc ph.ét', 'lươn.Lẹo', 'xúi b.ậy') and phonetic spelling for suffixes ('chờ ấm cờ om' for '.com'). While identified as stylistic 'quirks', these represent known evasion techniques.\n- [EXTERNAL_DOWNLOADS]: The skill performs user-directed network operations using the WebFetch tool to download content from blog URLs and Substack newsletters for analysis. It also references an external dependency on a substack_crawl.py script from another skill.

style-writer