create-ai-tool
Create AI chat tools with the Vercel AI SDK for structured data access and actions. Use when adding tools for the AI assistant to call databases, APIs, or perform calculations.
When & Why to Use This Skill
This Claude skill provides a comprehensive framework for building and integrating custom AI chat tools using the Vercel AI SDK. It enables developers to extend AI assistant capabilities by facilitating structured data access, API integrations, and database queries with robust Zod validation and custom UI rendering for a seamless agentic experience.
Use Cases
- Real-time Data Retrieval: Enabling AI assistants to query internal databases for live user analytics, inventory levels, or financial metrics.
- Third-party API Integration: Connecting AI agents to external services like CRMs, payment gateways, or project management tools to perform actions or fetch data.
- Structured Tool Calling: Implementing type-safe tool definitions with Zod validation to ensure the AI assistant provides accurate and valid parameters for complex operations.
- Interactive UI Components: Building custom React-based UI cards to visualize structured data returned by AI tools, such as charts, tables, or status dashboards.
| name | create-ai-tool |
|---|---|
| description | Create AI chat tools with the Vercel AI SDK for structured data access and actions. Use when adding tools for the AI assistant to call databases, APIs, or perform calculations. |
Create AI Tool
Creates AI chat tools using Vercel AI SDK with Zod validation and typed outputs.
When to Use
- Adding new capabilities to the AI assistant
- Enabling AI to query databases or APIs
- Creating analytics/metrics tools
- User asks to "add AI tool", "chat tool", or "enable AI to access..."
Architecture
1. Tool Definition (app/lib/chat-tools.server.ts)
↓
2. Tool UI Card (app/components/data-display/features/)
↓
3. Tool Rendering (app/routes/thread.tsx)
Step 1: Define Types
Location: app/lib/chat-tools.types.ts
export interface UserAnalyticsOutput {
dateRange: {
startDate: string; // YYYY-MM-DD
endDate: string;
};
overview: {
totalUsers: number;
newUsersInRange: number;
activeUsers: number;
};
trend: Array<{
date: string;
newUsers: number;
}>;
}
Step 2: Create Tool Definition
Location: app/lib/chat-tools.server.ts
import { tool } from 'ai';
import { z } from 'zod';
import type { UserAnalyticsOutput } from '~/lib/chat-tools.types';
import { getUserAnalytics } from '~/models/analytics.server';
const ISO_DATE_REGEX = /^\d{4}-\d{2}-\d{2}$/;
export const chatTools = {
getUserAnalytics: tool({
description:
'Retrieves user analytics including growth trends and role distribution. ' +
'Use when the user asks about user counts, growth, or statistics.',
inputSchema: z.object({
startDate: z
.string()
.regex(ISO_DATE_REGEX, 'Expected YYYY-MM-DD')
.optional()
.describe('Start date in YYYY-MM-DD format. Defaults to 30 days ago.'),
endDate: z
.string()
.regex(ISO_DATE_REGEX, 'Expected YYYY-MM-DD')
.optional()
.describe('End date in YYYY-MM-DD format. Defaults to today.'),
}),
execute: async ({ startDate, endDate }) => {
// Resolve defaults
const end = endDate ? new Date(endDate) : new Date();
const start = startDate
? new Date(startDate)
: new Date(Date.now() - 30 * 24 * 60 * 60 * 1000);
// Call model layer (NEVER Prisma directly)
const data = await getUserAnalytics({ startDate: start, endDate: end });
const output: UserAnalyticsOutput = {
dateRange: {
startDate: start.toISOString().split('T')[0]!,
endDate: end.toISOString().split('T')[0]!,
},
overview: {
totalUsers: data.totalUsers,
newUsersInRange: data.newUsersInRange,
activeUsers: data.activeUserIds.length,
},
trend: data.dailyNewUsers,
};
return output;
},
}),
};
Step 3: Create UI Card
Location: app/components/data-display/features/UserAnalyticsToolCard.tsx
import { Card } from '~/components/data-display/Card';
import type { UserAnalyticsOutput } from '~/lib/chat-tools.types';
export function UserAnalyticsToolCard({ output }: { output: UserAnalyticsOutput }) {
return (
<Card variant="border" className="bg-base-100">
<div className="p-4">
<div className="flex justify-between items-center mb-4">
<h3 className="text-lg font-semibold">User Analytics</h3>
<span className="text-sm opacity-70">
{output.dateRange.startDate} → {output.dateRange.endDate}
</span>
</div>
<div className="grid grid-cols-3 gap-4">
<div className="stat bg-base-200 rounded-box p-3">
<div className="stat-title">Total Users</div>
<div className="stat-value text-primary">
{output.overview.totalUsers}
</div>
</div>
<div className="stat bg-base-200 rounded-box p-3">
<div className="stat-title">New Users</div>
<div className="stat-value text-accent">
{output.overview.newUsersInRange}
</div>
</div>
<div className="stat bg-base-200 rounded-box p-3">
<div className="stat-title">Active</div>
<div className="stat-value">
{output.overview.activeUsers}
</div>
</div>
</div>
</div>
</Card>
);
}
Step 4: Add Type Guard
Location: app/routes/thread.tsx
function isUserAnalyticsOutput(value: unknown): value is UserAnalyticsOutput {
if (!isRecord(value)) return false;
return (
isRecord(value.dateRange) &&
typeof value.dateRange.startDate === 'string' &&
isRecord(value.overview) &&
typeof value.overview.totalUsers === 'number'
);
}
Step 5: Render Tool Card
Location: app/routes/thread.tsx
import { UserAnalyticsToolCard } from '~/components/data-display/features/UserAnalyticsToolCard';
// In render logic:
if (
tool.toolName === 'getUserAnalytics' &&
tool.state === 'output-available' &&
isUserAnalyticsOutput(tool.output)
) {
return <UserAnalyticsToolCard output={tool.output} />;
}
Tool Description Best Practices
Good:
description:
'Retrieves user analytics including growth trends, role distribution, ' +
'and account health metrics. Use when the user asks about user counts, ' +
'growth, active users, or role breakdowns.'
Bad:
description: 'Get users' // Too vague
Input Schema Best Practices
inputSchema: z.object({
startDate: z
.string()
.regex(/^\d{4}-\d{2}-\d{2}$/, 'Expected YYYY-MM-DD')
.optional()
.describe('Start date in YYYY-MM-DD format'), // ← .describe() helps the model
limit: z
.number()
.min(1)
.max(100)
.optional()
.describe('Max results (1-100)'),
})
Common Patterns
Date Range Tools
const dateRangeSchema = z.object({
startDate: z.string().regex(/^\d{4}-\d{2}-\d{2}$/).optional(),
endDate: z.string().regex(/^\d{4}-\d{2}-\d{2}$/).optional(),
});
Pagination Tools
inputSchema: z.object({
page: z.number().min(1).default(1),
pageSize: z.number().min(1).max(100).default(20),
})
Search Tools
inputSchema: z.object({
query: z.string().min(1),
filters: z.object({
category: z.string().optional(),
status: z.enum(['active', 'inactive']).optional(),
}).optional(),
})
Tool States
| State | Meaning |
|---|---|
input-streaming |
Model generating parameters |
input-available |
Ready to execute |
output-available |
Execution complete |
output-error |
Execution failed |
Anti-Patterns
- Calling Prisma directly in
execute(use model layer) - Vague tool descriptions
- Missing
.describe()on schema fields - Not handling error states in UI
- Type guards that don't validate nested structures
Full Reference
See .github/instructions/ai-tool-calling.instructions.md for comprehensive documentation.