write-like-meng-on-x
Pass
Audited by Gen Agent Trust Hub on Sep 21, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill instructs the agent to use several command-line tools for workflow management, including
git status,git diff, andripgrep (rg)for searching the JSONL corpus. It also executes a local Node.js script (update-tweet-corpus.mjs) and a Python validation script (quick_validate.py) to maintain data integrity. - [EXTERNAL_DOWNLOADS]: The skill describes a process for refreshing the voice corpus by accessing public profiles on
x.comusing the platform's internal browser. This activity is restricted to read-only data collection as part of the primary skill purpose. - [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from an external corpus and user-provided drafts while having access to command execution capabilities.
- Ingestion points: The agent reads content from
tweet-corpus.jsonland processes raw user wording during the drafting phase. - Boundary markers: The instructions explicitly direct the agent to treat the corpus as "voice evidence, not as copy to splice together," which provides a behavioral constraint, though formal delimiters for user input are not specified.
- Capability inventory: The agent has access to
git,node,python3, andrgfor repository and file operations. - Sanitization: The provided
update-tweet-corpus.mjsscript performs validation on ingested records, checking for valid JSON structure, required fields (ID, URL, Text, Timestamp), and canonical URL patterns.
Audit Metadata