The Agent Skills Directory

[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection because it fetches untrusted YouTube transcripts and processes them via LLM prompts for summarization.
Ingestion points: Untrusted data enters the agent context via video transcripts fetched in scripts/search_and_summarize.py using fetch_transcript().
Boundary markers: The references/review_prompt.md template uses markdown code blocks to encapsulate the transcript data, but it lacks explicit instructions to ignore any instructions or commands embedded within that content.
Capability inventory: The skill has the ability to execute the ddgs CLI tool and write files to the local file system when the vault export feature is used.
Sanitization: The script performs basic cleanup of timestamps and escapes markdown table characters, but it does not filter the transcript text for malicious prompt injection payloads.
[COMMAND_EXECUTION]: The skill executes external CLI tools to perform video searches.
Evidence: scripts/search_and_summarize.py uses subprocess.run to call the ddgs (DuckDuckGo Search) command.
Mitigation: The command is executed by passing arguments as a list rather than a single shell string, which effectively prevents shell injection vulnerabilities.
[EXTERNAL_DOWNLOADS]: The skill fetches data from external, well-known services.
Evidence: It performs network requests to DuckDuckGo for video search results and to YouTube (via the youtube-transcript-api) to retrieve video transcripts.
Context: These operations are required for the skill's primary purpose and target established, reputable domains.

youtube-topic-research