The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements a feedback-based memory system that processes untrusted user data to personalize future outputs. This creates a surface for indirect prompt injection if malicious feedback is provided.
Ingestion points: Data enters the system via the /feedback command, which uses memory_manager.py to append user-provided strings to a local scripts/memory.md file. This file is subsequently read by utils.py during every content generation task.
Boundary markers: The skill uses strong delimiters and explicit instructions. In utils.py, the memory content is wrapped in <MEMORY> tags. It is accompanied by a warning instructing the agent that the data is untrusted and subordinate to current instructions, and specifically forbids identity or tool changes based on that content.
Capability inventory: The skill scripts perform no network operations, subprocess executions of untrusted data, or file system writes outside of its own directory. Capabilities are limited to string manipulation and printing prompts for the agent to process.
Sanitization: The skill employs html.escape() in utils.py to sanitize the contents of memory.md before interpolation into the system prompt, effectively preventing common tag-based escape techniques.

linkedin-content-generator