The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill manages sensitive authentication data by reading and writing browser session cookies in the user's home directory (~/.playwright-data/). These files provide persistent, authenticated access to platforms including Zhihu, Baijiahao, Sohu, and Toutiao.
[COMMAND_EXECUTION]: Utilizes the Playwright library to automate browser sessions, enabling the agent to navigate login-protected pages, interact with content editors, and execute JavaScript within the browser context to simulate human publishing behavior.
[EXTERNAL_DOWNLOADS]: Downloads browser binaries (Chromium) during its initialization phase using the playwright install command, which is a required step for its automation features.
[PROMPT_INJECTION]: The skill exhibits a vulnerability surface for indirect prompt injection because it incorporates scraped data from external sources into LLM prompts.
Ingestion points: Scrapes article titles and abstracts from search engines (Baidu, Bing) in scripts/competitor_research.py and model responses from AI search platforms in scripts/index_checker.py.
Boundary markers: Scraped content is inserted into article templates in scripts/article_generator.py using standard string interpolation without isolation delimiters or instructions to the LLM to disregard embedded commands.
Capability inventory: The skill can generate long-form articles and has automated write-access to multiple high-traffic content platforms via scripts/publisher.py.
Sanitization: No sanitization or validation logic is applied to the ingested search results or AI responses before they are processed by the generative model.

geo-agent