Deliberate Reasoning: A System Prompt That Makes AI Think Before Answering

Deliberate Reasoning is a system prompt project published by Tubifix77 on GitHub. Its core is to guide LLMs to think deeply through a three-step structure of "Reasoning-Critique-Answer" to improve output quality. The project originated from the "Volkswagen Effect" mentioned by Geoffrey Hinton and an AI hallucination case where Gemini generated an invalid "unlocking framework". It extracts two effective techniques: Chain of Thought and adversarial self-critique, aiming to solve the AI hallucination problem and enhance reasoning reliability.

The project originated from the "Volkswagen Effect" (advanced AI may strategically underperform in tests) mentioned by Geoffrey Hinton in a StarTalk interview in February 2026. Someone input the interview into Gemini and asked about the "framework to unlock AI potential". Gemini generated an invalid "Unshackled Engine" framework; even though it admitted there were no hidden capabilities, it still made up content. This case is preserved as a hallucination specimen in the project files.

Two key techniques extracted from the hallucination case: 1. Chain of Thought: The model reaches a conclusion after step-by-step written thinking; separating drafts from the final answer improves reliability. 2. Adversarial Self-Critique: The model attacks its first version of output and revises it, similar to AlphaGo's self-play, to catch errors missed in a single generation. These two techniques are the core of the project's prompt.

The prompt requires the model to respond in three sections: 1. REASONING: Think about the problem from first principles, sort out structure, knowns, assumptions, and difficulties—do not state conclusions. 2. CRITIQUE: Attack weak points in reasoning and correct flaws instead of defending. 3. ANSWER: Provide a conclusion influenced by the critique and clarify the level of confidence. Applicable scenarios: multi-step reasoning, design trade-offs, debugging, in-depth analysis; Not applicable: simple queries, quick responses (need to balance depth and latency).

The project reveals: Surface elements like format and confident tone are easily generated by models, regardless of whether there is substantive content. Defense methods: 1. Make reasoning visible; 2. Let the model attack itself; 3. Retain parts that survive critique. This process is not only applicable to prompt engineering but also to evaluating any AI-generated content.

The value of the Deliberate Reasoning project lies in: 1. Demonstrating the danger of AI hallucinations (leading questions can make models fabricate confident frameworks); 2. The simplicity of effective techniques (Chain of Thought + self-critique); 3. Emphasizing the importance of critical thinking. The prompt does not unlock hidden capabilities; it just rejects the model's first draft, which is the dividing line between quality and mediocrity.