What are golden prompts in AI development?

Golden prompts are standardized test phrases—including direct, indirect, and negative examples—used to verify if an AI model triggers the correct tools or actions based on user input.

What is the difference between an indirect and a negative prompt?

An indirect prompt describes a goal without naming a tool (e.g., 'Remind me to buy milk'), while a negative prompt is a similar-sounding phrase that should NOT trigger a tool (e.g., 'How do I write a reminder?').

How does this skill improve my ChatGPT App?

It automates the tedious process of writing test cases, ensuring you have sufficient coverage to identify where your app's tool-calling logic might fail or trigger incorrectly.

Where are the generated prompts stored?

The prompts are saved in a structured format in the '.chatgpt-app/golden-prompts.json' file within your project directory for easy integration with your test suite.

ChatGPT App Golden Prompt Generator

Name: ChatGPT App Golden Prompt Generator
Author: hollaugo

byhollaugo

•

セキュリティとテスト

Generates comprehensive sets of test prompts to validate tool invocation and intent recognition for ChatGPT Apps.

This skill automates the creation of 'golden prompts'—a curated set of direct, indirect, and negative test cases—to ensure AI models correctly trigger specific tools within a ChatGPT application. By analyzing existing tool definitions from your state or server code, it generates high-quality test data that helps developers refine their app's natural language understanding and avoid false positives. This is essential for building robust, production-grade custom GPTs and AI agents that respond reliably to diverse user inputs.

主な機能

01Creation of negative prompts to prevent false-positive tool triggers

02Structured JSON output compatible with automated testing workflows

03Automated tool-aware analysis of application code

04Generation of 5+ direct prompts per tool for explicit validation

05Generation of 5+ indirect prompts per tool to test intent recognition

061 GitHub stars

ユースケース

01Validating that an AI agent correctly interprets nuanced or indirect user requests

02Reducing model hallucinations by testing phrases that should not trigger tools

03Scaling test coverage for complex ChatGPT Apps with multiple interdependent tools

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add hollaugo/prompt-circle-marketplace golden-prompts

For use in Claude.ai and ChatGPT

Download Skill