The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The servers provided (Memory, Sequential Thinking, Structured Argumentation, and Analogical Reasoning) act as persistent storage for agent reasoning and potentially untrusted external data.
Ingestion points: Untrusted data enters the agent context through the thought input in sequential_thinking, entity and observation contents in memory, and claims/premises in structuredArgumentation.
Boundary markers: The skill guide (SKILL.md) explicitly defines the 'Critical Principle' that the server provides structure but does not perform reasoning. It warns that servers are persistence mechanisms and should be treated as organizational tools rather than trusted reasoning engines.
Capability inventory: The memory server has local file-write capabilities (memory.json) to persist state across sessions. Other servers provide context back to the model and log formatted output to stderr for human oversight.
Sanitization: Input is validated against JSON schemas ensuring expected types (strings, booleans, integers) are received, which prevents basic schema confusion.
[DATA_EXPOSURE]: The Knowledge Graph Memory server interacts with the local filesystem to provide persistent memory for Claude.
Evidence: example-servers/memory/index.ts uses the MEMORY_FILE_PATH environment variable (defaulting to memory.json) to read and write knowledge graph state.
Safety Context: This is the intended core functionality of the server. Instructions within the skill correctly advise users to store secrets in environment variables rather than hardcoding them into the memory graph.

model-enhancement-servers