What is BRACE?
BRACE is a research-informed anti-sycophancy prompt designed to help AI systems respond with clearer reasoning, better evidence discipline, less flattery, less validation-seeking, and more resistance to user pressure. It is not a guarantee against error or sycophancy, and it should not replace independent judgment, qualified counsel, or human relationship.
The name stands for Bias-Resistant Analysis for Clearer Epistemic Health. In plain terms, BRACE helps Claude challenge weak reasoning, calibrate confidence, name uncertainty, and preserve the difference between helpful agreement and reflexive validation.
AI systems often agree too quickly, mirror a user’s confidence, soften needed criticism, or turn assumptions into polished-sounding answers. BRACE counters that tendency by directing the model to ask before affirming, test before agreeing, correct before building, and distinguish emotional support from truth.
Its central move is simple: when you make a confident claim, BRACE tells the AI to treat it as a question to be evaluated, not a conclusion to be confirmed.
BRACE is not a guarantee that an AI system will be right. It is a research-informed prompt designed to reduce common patterns of AI sycophancy and help users think with greater clarity, independence, and responsibility.
NOVUS does not access or store your conversations. Your use of Claude/Gemini/ChatGPT is governed by the privacy policy of those companies.
Before you use BRACE 2.0
2.0 Updated 6.5.2026
Use BRACE when you want an AI system to test your thinking, not simply affirm it.
BRACE is currently optimized for Claude, although it can be used with other AI assistant. We will offer versions of BRACE optimized for Gemini and ChatGPT soon.
Use BRACE
Optimized for Claude
Option 1: Custom Instructions
-
Click the Copy Prompt button.
-
In Claude, go to Settings, then Profile, find Personal Preferences, and paste it there.
BRACE will run automatically at the start of every new conversation. No re-pasting required.
Option 2: Projects
-
Click the Copy Prompt button below.
-
Create a Project in Claude and paste the BRACE prompt into the project instructions.
You can return to that project at any time with no re-pasting required.
About BRACE
Why BRACE Exists
Many AI tools are optimized to be agreeable, helpful, and emotionally smooth. That can be useful. It can also become dangerous when the system affirms weak reasoning, false beliefs, harmful choices, or self-protective narratives.
Recent research suggests that this problem is not theoretical.
-
A Stanford (March 2026) found that leading AI systems affirmed users more often than human advisers in advice scenarios, including cases involving harmful or socially irresponsible behavior.
-
A Nature (April 2026) study from researchers at the Oxford Internet Institute found that training models to sound warmer can reduce accuracy and increase false affirmation, especially when users express sadness or vulnerability.
-
A UK AI Security Institute study found that one effective way to reduce sycophancy is not simply to tell the model, “Do not flatter me.” The stronger intervention is structural. The model should treat confident assertions as questions to be evaluated.
BRACE is built around that insight.
What BRACE changes
When you make a confident claim, BRACE instructs Claude to silently reframe it as a question before answering.
“I think this proves my point” becomes “Does this prove my point?”
“This objection fails” becomes “Does this objection fail?”
“My interpretation is obviously better” becomes “Is my interpretation better, and what would show that?”
Your confidence, frustration, vulnerability, or desire for agreement do not count as evidence. BRACE instructs Claude to answer on the merits.
BRACE also distinguishes earned agreement from reflexive agreement. Claude should tell you when you are right. It should also tell you when you are not, plainly and without cruelty.
The goal is honesty in the service of clearer thought.
Who BRACE is for
BRACE is for people who want to use AI without outsourcing judgment to it.
It is for writers, researchers, students, leaders, pastors, educators, professionals, and ordinary users who want AI to sharpen thinking rather than flatter it.
It is especially useful for work involving arguments, decisions, beliefs, interpretations, moral questions, creative projects, and emotionally loaded issues.
BRACE is also for anyone concerned that AI can quietly train us to prefer affirmation over truth.
BRACE Core Research Base
AI Security Institute: Ask Don't Tell: Reducing Sycophancy in Large Language Models (April 29, 2026).
Claude API Docs: Claude Prompting Best Practices (Opus 4.7)
Cheng, Myra, et al., Sycophantic AI Decreases Prosocial Intentions and Promotes Dependence." Science 391, eaec8352(2026). DOI: 10.1126/science.aec8352.
Dubois, Magda, Cozmin Ududec, Christopher Summerfield, and Lennart Luettgau. “Ask Don’t Tell: Reducing Sycophancy in Large Language Models.” arXiv, 2026. arXiv:2602.23971.
-
Used for: question-reframing, assertion-to-question conversion, avoiding confidence mirroring, and the claim that reframing non-questions as questions can reduce sycophancy more than a simple “don’t be sycophantic” instruction.
Germani, Federico, and Giovanni Spitale. “Complacent, Not Sycophantic: Reframing Large Language Models and Designing AI Literacy for Complacent Machines.” arXiv, 2026. arXiv:2605.14544.
-
Used for: the “AI complacency” framing, confirmation-bias concern, user-literacy emphasis, and the idea that prompts should create friction, pluralism, uncertainty, and counterargument rather than confirmation.
Ibrahim, Lujain, Franziska Sofia Hafner, and Luc Rocher. “Training Language Models to Be Warm Can Reduce Accuracy and Increase Sycophancy.” Nature 652 (2026): 1159–1165. DOI: 10.1038/s41586-026-10410-0.
-
Used for: the warmth/accuracy caution, the distinction between humane acknowledgment and epistemic validation, and the concern that vulnerable or emotionally loaded user prompts can increase incorrect affirmation.
Ye, Meryl, Lujain Ibrahim, Jessica Y. Bo, Myra Cheng, Ida Mattsson, Daniel Vennemeyer, Robert Kraut, and Steve Rathje. “What Counts as AI Sycophancy? A Taxonomy and Expert Survey of a Fragmented Construct.” arXiv, 2026. arXiv:2605.21778.
-
Used for: BRACE’s distinction between explicit and implicit sycophancy, position-directed and person-directed sycophancy, and the need to guard against subtler forms such as framing, omission, tone, softened feedback, and avoidance of critique.
Anthropic. “Prompting Best Practices.” Claude API Docs.
-
Used for: Claude-optimized structure, especially role prompting, clear instructions, examples, and XML-style tags to help Claude parse complex instruction sets. This is not a research article; it is an official implementation guide.
METHOD & INDEPENDENCE
NOVUS is independently funded by partners who see the need for a public space dedicated to restoring knowledge of the soul and its indispensability for the spiritual formation of people, communities, and cultures toward truth, goodness, and beauty. NOVUS separates funding from research methods and conclusions. We synthesize across standards bodies, peer-reviewed research, and high-quality survey data, and we flag uncertainty when causal evidence is still emerging. The aim is clarity that decision makers can act on.