The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted text, comments, and images from a public social media platform (Xiaohongshu) and feed them directly into the agent's analysis context.
Ingestion points: The agent is explicitly instructed in SKILL.md to read the generated [keyword]_raw_data.md and use vision capabilities to interpret the semantic content of downloaded .webp and .jpg image files.
Boundary markers: The skill lacks delimiters or "ignore embedded instructions" warnings to separate user-generated content from the agent's core instructions, making it vulnerable to malicious payloads embedded in social media posts.
Capability inventory: The agent possesses file reading, file writing, and vision-based image analysis capabilities which are applied to the scraped data.
Sanitization: The scripts/parse.py script performs only basic text formatting and does not sanitize or filter the content for potential injection attempts.
[EXTERNAL_DOWNLOADS]: The skill fetches image artifacts from Xiaohongshu's official CDN domains.
Evidence: The scripts/parse.py script validates that image URLs belong to .xhscdn.com or .xiaohongshu.com and use the HTTPS protocol before downloading them to the local output directory.
[DYNAMIC_EXECUTION]: The skill generates and executes browser automation scripts at runtime to bypass anti-scraping measures.
Evidence: The scripts/run.sh script creates a temporary JavaScript file in /tmp containing site-specific extraction logic, which is then executed by the playwright-cli tool.
[COMMAND_EXECUTION]: The skill uses shell and Python scripts to orchestrate the collection and parsing workflow.
Evidence: It relies on scripts/run.sh to manage environment variables and directory structures, and invokes python3 to execute the data parsing and image downloading script.

xiaohongshu-search-summarizer