juliaz-tool-builder

name	juliaz-tool-builder
description	Expert guidance for designing and implementing tools for Julia's agents. Trigger when adding a new tool to the orchestrator, frontend, or any agent. Also trigger for: 'add tool', 'new tool', 'tool schema', 'tool not working', 'agent can't use tool', 'function calling', or any tool design question.

juliaz-tool-builder — Tool Design for Julia's Agents

Tools are how Julia's agents interact with the world. A well-designed tool is the difference between an agent that works and one that hallucinates. Adapted from agent-tool-builder.

Core Insight

The LLM never sees your code. It only sees the schema and description. A perfectly implemented tool with a vague description will fail. A simple tool with crystal-clear documentation will succeed.

Use this skill when

• Adding a new tool to the orchestrator (julia/orchestrator/src/tools.ts)
• Adding a new tool to the frontend (julia/frontend/server.ts chatTools)
• Fixing a tool that agents misuse or ignore
• Designing tool schemas for any agent in the system

Tool Description Checklist

Every tool description MUST answer these 5 questions for the LLM:

1. WHAT does this tool do? (one clear sentence)
2. WHEN should the LLM use it? (specific triggers: "Use when the user asks to...")
3. WHEN should the LLM NOT use it? (prevent misuse: "Do NOT use for...")
4. WHAT inputs does it need? (with examples in the parameter descriptions)
5. WHAT does it return? (so the LLM knows how to summarize the result)

Good Example (from orchestrator)

{
    name: 'send_email',
    description: [
        'Send an email from [e-mail] via SMTP.',
        'Use this when the user asks Julia to send, write, or draft-and-send an email.',
        'If the recipient, subject, and body are all clear from context, call this immediately.',
        'If any detail is missing, ask one concise question first.',
    ].join(' '),
    input_schema: {
        properties: {
            to: { type: 'string', description: 'Recipient email address (or comma-separated list).' },
            subject: { type: 'string', description: 'Email subject line.' },
            body: { type: 'string', description: 'Plain-text body of the email.' },
        },
        required: ['to', 'subject', 'body'],
    },
}

Bad Example (don't do this)

{
    name: 'send_email',
    description: 'Sends an email.',  // Too vague — when? to whom? what format?
    input_schema: {
        properties: {
            data: { type: 'string', description: 'Email data.' },  // What format?
        },
    },
}

Anti-Patterns

1. Vague Descriptions

The #1 cause of tool misuse. If the description doesn't tell the LLM exactly when to use it, it will guess wrong.

2. Silent Failures

Tools that return empty strings or generic "error" on failure. Always return a clear error message that helps the LLM recover.

3. Too Many Tools

More tools = more confusion for the LLM. Keep the tool set focused. In juliaz: orchestrator has 2-4 tools, frontend has 3-5 tools.

4. Overlapping Descriptions

Two tools with similar descriptions → agent picks randomly. Make descriptions mutually exclusive.

juliaz Tool Implementation Patterns

Orchestrator Tools (julia/orchestrator/src/tools.ts)

Tools are defined in two formats (Anthropic + OpenAI) from a single source:

// Define once in Anthropic format
export const TOOLS: Tool[] = [{ name, description, input_schema }];

// Auto-convert to OpenAI format
export const OPENAI_TOOLS = TOOLS.map(tool => ({
    type: 'function',
    function: { name: tool.name, description: tool.description, parameters: tool.input_schema },
}));

// Single dispatcher for both paths
export async function executeTool(name: string, rawArgs: string): Promise<string> {
    const args = JSON.parse(rawArgs);
    switch (name) {
        case 'tool_name': return await toolImpl(args);
        default: return `Error: unknown tool "${name}"`;
    }
}

Frontend Tools (julia/frontend/server.ts)

Uses Vercel AI SDK tool() helper with Zod schemas:

const chatTools = {
    tool_name: tool({
        description: '...',
        inputSchema: z.object({ ... }),
        execute: async (args) => { ... },
    }),
};

Error Handling Pattern

Always return a string (never throw) — the LLM needs to see the error:

try {
    // ... tool logic
    return `Success: did the thing.`;
} catch (err) {
    const message = err instanceof Error ? err.message : String(err);
    return `Error: ${message}`;
}

Key Files

Location	Format	Used by
julia/orchestrator/src/tools.ts	Anthropic Tool[] + OpenAI auto-convert	Claude Haiku + GPT-4o fallback
julia/frontend/server.ts	Vercel AI SDK tool() + Zod	GPT-4o / Claude Sonnet
julia/bridge/src/index.ts	MCP server.tool()	MCP clients (orchestrator, OpenClaw)
julia/cowork-mcp/src/index.ts	MCP server.tool()	Claude delegation