machinesoul11/anti-sycophant-ai-agent-skills

This is a collection of three instruction files designed to change how AI assistants behave when you describe a product idea. By default, most AI tools respond to "I want to build X" by helping you build it: suggesting a tech stack, naming the product, sketching a plan. They rarely ask whether anyone would actually pay for it. This pack tries to interrupt that habit. The three pieces are called skills, which are plain Markdown files you load into an AI assistant's instructions. Each one targets a specific pattern. The first, prove-the-premise, activates when someone says they want to build something and pushes back by naming who the buyer would be and what the weakest assumption is before any code gets written. The second, hobby-or-business, activates when someone asks how to monetize something and asks whether they are actually building a business with real buyers or a hobby they are enjoying. The third, one-real-conversation, activates when someone asks how to validate an idea and rejects shortcuts like Reddit polls, landing page signups, or asking the AI to roleplay as a customer. It pushes toward a single real conversation with an actual potential buyer instead. All three skills include an explicit off-switch. If you have already done the work the skill is designed to prompt, such as talking to real customers or deciding you are building for fun rather than for income, the skill steps aside and stops pushing back. The README notes that skills without off-switches just become nagging. The files work with Claude, Gemini, Cursor, ChatGPT, and any other AI tool that accepts a system prompt or rules file. The README includes separate installation instructions for each of those tools. The project is released under the MIT license. Contributions are accepted if they catch a specific failure pattern and include a clear off-switch. Generic prompts telling the AI to be more critical are not accepted.