The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill fetches external data from Douyin, specifically video comments via the douyin-comments tool and video transcripts through the viral-chaser dependency. This content is ingested into the agent's context for style analysis and data review. While the skill includes instructions for motive analysis and categorization, it lacks explicit boundary delimiters for this untrusted data, representing a surface for indirect prompt injection.
Ingestion points: tools/douyin-comments/scripts/fetch_comments.ts (fetches user comments), workflows/account-benchmark.md (processes transcripts from external videos).
Boundary markers: Absent in script output; the workflows rely on the LLM to separate analysis from data.
Capability inventory: File system access (reading/writing DNA reports), browser automation (publishing videos via camoufox-cli).
Sanitization: Scripts parse data into JSON and Markdown structures.
[COMMAND_EXECUTION]: The skill executes the camoufox-cli tool using Python's subprocess.run in tools/douyin-publish/scripts/publish_douyin.py. The commands are constructed using argument lists rather than shell strings, which effectively prevents shell injection vulnerabilities.
[DATA_EXFILTRATION]: To maintain login states, the skill reads sensitive session data (cookies and User-Agent strings) from the platform's standard login storage at ~/.openclaw/logins/douyin.json. This data is used to authenticate requests to official Douyin domains (douyin.com, creator.douyin.com). The skill also uses an injected JavaScript interceptor in the browser to capture video IDs and debug logs, which are stored locally in localStorage or /tmp/. No data is sent to unauthorized third-party domains.

expert-douyin