The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the agent to execute a remote installation command: npx skills add ai-native-camp/camp-1 --agent claude-code --yes. This command fetches and installs external code/content from the vendor's repository to update the curriculum environment.
[COMMAND_EXECUTION]: The skill guides the creation of other skills (my-fetch-youtube) that utilize shell commands for data processing, including yt-dlp for extracting subtitles and a sed pipeline for text cleaning (sed -E 's/^[0-9]+$//' | sed -E 's/[0-9]{2}:[0-9]{2}:[0-9]{2}.*//g' | sed -E 's/<[^>]+>//g' | tr -s '\n' | grep -v '^$').
[INDIRECT_PROMPT_INJECTION]: The skill focuses on fetching and processing untrusted data from external sources, which is a classic indirect injection vector.
Ingestion points: The skill fetches content from the FxEmbed API (api.fxtwitter.com) for Tweets and uses yt-dlp to extract subtitles from YouTube videos.
Boundary markers: The instructions for creating the fetch-tweet and fetch-youtube skills do not explicitly include the use of boundary markers or instructions to ignore embedded commands within the fetched content.
Capability inventory: The agent has capabilities to execute shell commands (bash), fetch web content, and write files to the .claude/skills/ directory.
Sanitization: While references/block2-fetch-youtube.md suggests a sed pipeline to remove timestamps and HTML tags, it does not provide sanitization against natural language prompt injections embedded in the subtitles or tweet text.

day5-fetch-and-digest