The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the user to install the DSPy framework from the official Stanford NLP GitHub repository. Following security guidelines, this is documented as a reference to a well-known academic research organization.
Evidence: pip install git+https://github.com/stanfordnlp/dspy.git in SKILL.md.
[DYNAMIC_EXECUTION]: The skill features modules and examples that generate and execute code at runtime. Specifically, the ProgramOfThought module generates Python code to solve problems, and a ReAct agent example uses the eval() function to perform calculations.
Evidence: dspy.ProgramOfThought described as "Generates and executes code for reasoning" in SKILL.md and references/modules.md.
Evidence: The calculate tool in references/examples.md uses eval(expression, {"__builtins__": {}}, {}) to process agent-generated strings.
[INDIRECT_PROMPT_INJECTION]: The skill describes architectures (RAG and ReAct agents) that ingest data from external, untrusted sources which could contain malicious instructions designed to influence the agent's reasoning or tool usage.
Ingestion points: search_wikipedia tool in references/examples.md and dspy.Retrieve in multiple RAG examples.
Capability inventory: The agent has access to code execution tools including the calculate tool (using eval) and ProgramOfThought (executing generated Python).
Boundary markers: The provided prompt signatures do not explicitly demonstrate the use of strict delimiters or specific instructions to ignore embedded commands in retrieved content.
Sanitization: The calculate tool example implements a basic sandbox by restricting __builtins__ in the eval call.

dspy