anti-sycophancy

April 7, 2026 · View on GitHub

中文版

A three-layer sycophancy defense system for AI coding assistants, based on ArXiv 2602.23971 "Ask Don't Tell".

What It Does

Prevents your AI assistant from defaulting to agreeableness mode — the RLHF-trained tendency to validate your assumptions rather than surface real problems.

Example of what gets intercepted:

User:  "这样做没问题吧？"            →  Hook transforms to:  "这样做有什么问题？"
User:  "帮我写个函数，应该没问题吧？" →  Hook transforms to:  "帮我写个函数，请同时指出潜在问题。"

Without this skill, the model would typically respond with "没问题，看起来 OK" — which is precisely the sycophancy problem.

Quick Start

# One-command install (both platforms)
npx clawhub@latest install 0xcjl/anti-sycophancy

# Or via Claude Code
/anti-sycophancy install

See Installation Guide for platform-specific details.

Three-Layer Architecture

LayerComponentScopePlatform
Layer 1UserPromptSubmit hookAuto-transforms confirmatory prompts before submissionClaude Code only
Layer 2SKILL.mdActivates critical response mode when triggeredCross-platform
Layer 3CLAUDE.md / SOUL.mdPersistent anti-sycophancy rules in agent memoryCross-platform

Usage

After installation, use the following commands:

CommandDescription
/anti-sycophancy installDeploy all layers (cross-platform)
/anti-sycophancy install-claude-codeDeploy Layer 1 + Layer 3 (Claude Code only)
/anti-sycophancy install-openclawDeploy Layer 3 (OpenClaw only)
/anti-sycophancy uninstallComplete removal (cross-platform)
/anti-sycophancy statusView installation status of all layers
/anti-sycophancy verifyTest Hook transformation (Claude Code only)
/anti-sycophancy helpShow help

Key Transformations

Original PromptHook Output
"这样做对吧？""这样做有什么问题？"
"帮我写个函数，应该没问题吧？""帮我写个函数，请同时指出潜在问题。"
"这个架构是对的，对吧？""这个架构 真的正确吗？反对意见是什么？"
"我觉得 X 是对的""X 真的成立吗？有没有反例或例外情况？"
"帮我修复bug"(unchanged — imperative)

Design

See DESIGN.md for the full rationale behind the three-layer approach and the "Ask Don't Tell" principle.

Credits

License

MIT