A three-layer sycophancy defense system for AI coding assistants, based on ArXiv 2602.23971 "Ask Don't Tell".
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.
# One-command install (both platforms)
npx clawhub@latest install 0xcjl/anti-sycophancy
# Or via Claude Code
/anti-sycophancy installSee Installation Guide for platform-specific details.
| Layer | Component | Scope | Platform |
|---|---|---|---|
| Layer 1 | UserPromptSubmit hook |
Auto-transforms confirmatory prompts before submission | Claude Code only |
| Layer 2 | SKILL.md |
Activates critical response mode when triggered | Cross-platform |
| Layer 3 | CLAUDE.md / SOUL.md |
Persistent anti-sycophancy rules in agent memory | Cross-platform |
After installation, use the following commands:
| Command | Description |
|---|---|
/anti-sycophancy install |
Deploy all layers (cross-platform) |
/anti-sycophancy install-claude-code |
Deploy Layer 1 + Layer 3 (Claude Code only) |
/anti-sycophancy install-openclaw |
Deploy Layer 3 (OpenClaw only) |
/anti-sycophancy uninstall |
Complete removal (cross-platform) |
/anti-sycophancy status |
View installation status of all layers |
/anti-sycophancy verify |
Test Hook transformation (Claude Code only) |
/anti-sycophancy help |
Show help |
| Original Prompt | Hook 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) |
See DESIGN.md for the full rationale behind the three-layer approach and the "Ask Don't Tell" principle.
- Research: ArXiv 2602.23971 — "Ask Don't Tell: Reducing Sycophancy in Large Language Models" (Dubois, Ududec, Summerfield, Luettgau, 2026)
- Playbook: openclaw-playbook
- Author: 0xcjl
- Optimized via: cjl-autoresearch-cc — 40-round iterative optimization
MIT