juliaz-tool-builder

The juliaz-tool-builder skill provides specialized engineering guidance for developers working on the Julia multi-agent system. It focuses on the critical bridge between agent intent and functional implementation by emphasizing that Large Language Models rely entirely on schema and documentation to interact with software tools. Without clear, structured definitions, agents are prone to ambiguity and hallucinations. This skill is designed for architects and developers maintaining the orchestrator, frontend, and various internal agents, helping them bridge the gap between logical tool design and runtime performance. It ensures that tool definitions are not only functional but also sufficiently expressive to allow the model to make precise decisions.

• Expert design patterns for tool definitions in both Anthropic and OpenAI formats within the orchestrator architecture.

• Implementation standards for frontend tools using the Vercel AI SDK and Zod schemas for type-safe interaction.

• Guidelines for crafting clear, 5-part tool descriptions that define purpose, trigger scenarios, negative constraints, required inputs, and expected output formats.

• Best practices for error handling, emphasizing returning strings instead of throwing exceptions to facilitate seamless model recovery.

• Strategies for preventing common anti-patterns such as vague descriptions, silent failures, tool bloat, and overlapping functional descriptions.

• Support for integration with diverse components like the bridge (MCP), backend APIs, and the Antigravity development environment.

• Trigger this skill during the development of new tools for orchestrator (julia/orchestrator/src/tools.ts) or frontend (julia/frontend/server.ts) modules.

• Use during debugging sessions when agents consistently misuse, ignore, or struggle to call specific functions.

• Follow the mandatory 5-question checklist to ensure every tool has a clear definition, usage triggers, and exclusion criteria.

• Adhere to the defined error handling patterns to ensure the LLM receives actionable feedback rather than opaque stack traces.

• Keep the tool set focused by maintaining a low number of tools per agent (e.g., 2-4 for the orchestrator) to minimize cognitive load on the model.

• Prioritize clarity in parameter descriptions by including specific examples, which significantly increases the likelihood of successful function calling.