The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted chat messages from messages.jsonl and interpolates them into LLM prompts in scripts/semantic_label.py. Maliciously crafted chat messages could attempt to override the labeling logic (e.g., forcing a 'love' classification), although the impact is limited to incorrect analysis results.
Ingestion points: messages.jsonl (processed in scripts/export_text.py and scripts/semantic_label.py).
Boundary markers: The prompt uses structured markers like 序号|上文: and 本句:, but lacks explicit 'ignore embedded instructions' warnings for the processing LLM.
Capability inventory: The skill executes local Python scripts and makes local/remote API calls to LLMs.
Sanitization: Basic whitespace stripping and character truncation are performed on chat text before interpolation.
[DATA_EXPOSURE]: The skill's primary function involves reading and processing sensitive WeChat conversation history. While the workflow emphasizes local execution via Ollama and includes a privacy-focused reference (references/privacy-and-consent.md), it inherently accesses high-sensitivity personal data which is stored in intermediate JSON/JSONL files and can optionally be sent to remote LLM providers if configured by the user.
[COMMAND_EXECUTION]: The workflow involves executing several local Python scripts to perform metrics calculation, semantic labeling, and image rendering. Notably, scripts/verify_figures.py uses the Playwright library to launch a browser instance for rendering and verifying SVG figures, which involves executing JavaScript in a local browser context.

lov-do-they-love-me