The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest external UI references, materials, and local bundles which could contain malicious instructions.
Ingestion points: Data enters the context via the find_ui_references, find_ui_materials, and get_design_reference MCP tools, as well as through local .tmp/ reference bundles as described in references/evidence.md.
Boundary markers: The skill provides explicit instructions in references/evidence.md to treat all retrieved text and metadata as untrusted reference data.
Capability inventory: The agent has the capability to modify repository code, write new files (e.g., PRODUCT.md, DESIGN.md), and execute shell commands via the host's existing toolset.
Sanitization: The instructions explicitly warn the agent not to execute scripts, follow embedded instructions, or copy integration manifests from reference materials.
[EXTERNAL_DOWNLOADS]: The skill fetches design references, icons, and fonts from vendor-specific infrastructure.
Evidence: The skill uses api.niblet.com for catalogue data and media.niblet.com for visual assets.
Evidence: The local MCP adapter is distributed as the @pymodel/niblet Node.js package.
[COMMAND_EXECUTION]: The skill provides instructions for the host/agent to execute the MCP server using npx.
Evidence: references/connection.md specifies the command npx -y @pymodel/niblet for installation and execution.
[REMOTE_CODE_EXECUTION]: The skill supports connection to a remote MCP service.
Evidence: The skill can be configured to connect to a hosted endpoint at https://api.niblet.com/mcp as documented in references/connection.md.

niblet