The Agent Skills Directory

Indirect Prompt Injection Surface: The guide describes how to create agents that ingest user prompts and orchestrate tools (e.g., using tools: ['*'] or VS Code's tool calling API). This creates a potential surface where untrusted user input could influence tool execution.
Ingestion points: User input enters the agent context via request.prompt in the TypeScript handler or through direct interaction with declarative agents described in Step 2.
Boundary markers: The skill suggests using system prompt instructions to define behaviors, but does not explicitly document the use of delimiters or specific boundary markers for external data processing in the provided templates.
Capability inventory: Agents created using this guide are shown to have capabilities for file system interaction (via codebase toolset), network operations (via fetch and search), and potential full editing capabilities (tools: ['*']).
Sanitization: The guide includes a validation checklist that advises developers to ensure instructions don't include secrets or internal URLs, but does not detail specific methods for sanitizing or escaping user-provided input before it reaches the language model.
Security Best Practices: The skill promotes positive security patterns by including a validation step to verify that agent instructions are free of hardcoded credentials, tokens, or internal infrastructure details.

make-custom-agent