The Agent Skills Directory

[Local File Access and Sandboxing]: The skill instructs the creation of agents that can read local files and search documentation. While accessing the file system is a sensitive operation, the instructions explicitly mandate that every tool must resolve paths and reject any requests outside a sandboxed project directory, effectively mitigating the risk of unauthorized file access.
[Indirect Prompt Injection Surface]: As the skill involves processing local data (via RAG and file tools), it introduces a surface for indirect prompt injection. However, the skill provides mitigation guidance by requiring scoped operations and validation of outputs.
Ingestion points: Data enters the agent context through read_file operations and search_docs via Chroma.
Boundary markers: The skill emphasizes keeping operations within a "sandboxed project directory."
Capability inventory: The resulting agent can execute local tool calls and file system reads based on its reasoning loop.
Sanitization: The instructions guide the developer to ensure path resolution logic rejects attempts to access files outside the project root.
[Model Provisioning]: The skill uses the foundry CLI to download and run language models locally. This is a standard procedure for local AI development and leverages the vendor's own infrastructure for model distribution.

local-ai-agents