The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill represents a surface for indirect prompt injection attacks by ingesting untrusted data from external URLs (YouTube transcripts, X/Twitter posts, and web links) and processing it through an AI subagent to generate drafts.
Ingestion points: SKILL.md defines a routing matrix that fetches content from YouTube, X, and general web links using auxiliary skills like chromium, twitter, and transcript.
Boundary markers: There are no explicit instructions or delimiters defined to isolate external content or prevent the subagent from following embedded instructions within the fetched data.
Capability inventory: The skill uses a custom storage-cli tool to read/write ideas, drafts, and persona data, and utilizes an "Opus subagent" for drafting.
Sanitization: The "Fact-check" step in the pipeline (Section 3, Step 7) is designed for data accuracy regarding numbers and quotes but does not include sanitization for malicious instructions or prompt injection.
[COMMAND_EXECUTION]: The skill utilizes a custom tool, storage-cli, and the jq utility to manage persistent data through shell-like commands.
Evidence: references/orgbus-commands.md documents the use of storage-cli for get, push, patch, and put operations across various storage paths like content/telegram/ideas and content/telegram/stance/core. It also demonstrates the use of jq for parsing JSON responses in shell snippets.

content-engine