When building apps with AI agents in Cursor or Claude Code, LLMs often default to loose TypeScript types (any) and silent exception swallowing inside try-catch blocks.
By default, an LLM prioritizes code that executes without breaking the immediate compiler scope, even if it creates severe production runtime bugs down the road.
The Problem: Loose Types & Silent Failure
typescript
// AI Default Output: Loose types and swallowed errors
const fetchUserProfile = async (userId: any) => {
try {
const res = await fetch("/api/users/" + userId);
const data = await res.json();
return data;
} catch (err) {
console.log("Error fetching profile", err);
return null; // Silent failure swallowed!
}
};
The Solution: Explicit Prompt Directives
To force Cursor or Claude Code to output production-grade error boundaries and strict TypeScript interfaces, add these explicit directives to your system prompt:
•Never use loose type assertions or any. Always define explicit interfaces or Zod schemas.
•Do not swallow errors in catch blocks. Return explicit HTTP 500 error responses or throw custom Error classes.
•Ensure all async promises have explicit error handlers.
Refactored Production Pattern
typescript
import { z } from 'zod';
const UserProfileSchema = z.object({
id: z.string(),
email: z.string().email(),
subscriptionStatus: z.enum(['active', 'canceled', 'trialing']),
});
export type UserProfile = z.infer<typeof UserProfileSchema>;
export async function fetchUserProfile(userId: string): Promise<UserProfile> {
const res = await fetch("/api/users/" + userId);
if (!res.ok) {
throw new Error("Failed to fetch user profile: " + res.statusText);
}
const json = await res.json();
return UserProfileSchema.parse(json);
}
Audit Recommendation: Don't rely solely on manual code reviews. Run an automated static AST scan to verify missing error boundaries before deploying to production.