The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill documents the use of dspy.ProgramOfThought, which is designed to generate and execute Python code at runtime to perform calculations and data transformations. While a core feature of the library, executing code generated by a language model based on user input carries inherent risks of arbitrary code execution if the model is manipulated.
[INDIRECT_PROMPT_INJECTION]: The skill patterns for RAG (Retrieval-Augmented Generation) and agent-based reasoning (e.g., ReAct) involve ingesting data from untrusted external sources like Wikipedia, web searches, or document databases. These inputs are interpolated into prompts, creating a surface for indirect prompt injection where malicious instructions in the retrieved data could influence agent behavior.
Ingestion points: Data enters the context via dspy.Retrieve and tool outputs in the ReAct and Multi-Hop RAG modules.
Boundary markers: The provided code snippets do not explicitly show the use of XML tags or other delimiters to isolate untrusted context from instructions.
Capability inventory: The documented modules have capabilities including file system writes (save/load methods), tool execution (Wikipedia/Web search), and dynamic Python execution via ProgramOfThought.
Sanitization: No explicit sanitization or validation of retrieved context is described before interpolation into the prompt.
[EXTERNAL_DOWNLOADS]: The documentation includes instructions to install the DSPy library directly from its official development repository on GitHub (stanfordnlp/dspy).
[COMMAND_EXECUTION]: The skill provides standard installation commands using pip to set up the environment and its dependencies.

dspy