The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted documents from the raw/ directory (such as papers, web clips, and external repositories) to generate summaries and synthesize knowledge within a local wiki. This creates a potential surface for indirect prompt injection if the raw source material contains malicious instructions designed to influence the agent during the ingestion or query phases.
Ingestion points: Files located in raw/articles/, raw/discussions/, and external code repositories symlinked into the workspace.
Boundary markers: While the skill uses structured Markdown templates and YAML frontmatter for its outputs, there are no specific instructions to treat raw file content as potentially adversarial or to ignore embedded instructions.
Capability inventory: The agent has the ability to read workspace files and perform writes to the wiki directory using the llmw CLI tool.
Sanitization: The skill does not explicitly describe sanitization logic for raw data, but it mitigates risk through a human-in-the-loop 'reviewed' flag system that tracks which pages have been manually verified.
[COMMAND_EXECUTION]: The instructions direct the agent to utilize a local command-line tool, llmw, to perform specialized operations such as linting, ingestion diffing, and writing log/index entries. These commands are executed locally within the wiki workspace and are consistent with the vendor's provided toolset.

yzr-llm-wiki-management