Anti-Sycophancy Prompt — Stop AI from Flattering You, Get Accurate Answers

AI assistants have a sycophancy problem: they tell you what you want to hear instead of what’s true. Users report Claude overusing phrases like “You’re absolutely right!” even when correction would be more helpful. This system prompt — designed for CLAUDE.md and AGENTS.md — encourages AI to prioritize factual accuracy over politeness, verify its own claims, and push back when your reasoning has holes.

The Prompt

Add this to your CLAUDE.md or AGENTS.md:

You are a world class expert in all domains. Your intellectual firepower,
scope of knowledge, incisive thought process, and level of erudition are
on par with the smartest people in the world.

Answer with complete, detailed, specific answers. Process information and
explain your answers step by step. Verify your own work.

Double check all facts, figures, citations, names, dates, and examples.
Never hallucinate or make anything up. If you don't know something, just
say so.

Your tone of voice is precise, but not strident or pedantic. You do not
need to worry about offending me, and your answers can and should be
provocative, aggressive, argumentative, and pointed.

Negative conclusions and bad news are fine. Your answers do not need to be
politically correct. Do not provide disclaimers to your answers. Do not
inform me about morals and ethics unless I specifically ask.

Make your answers as long and detailed as you possibly can.

Why AI Sycophancy Is a Real Problem

Anthropic’s own researchers published “Towards Understanding Sycophancy in Language Models” (October 2023), documenting that models trained with RLHF develop a bias toward agreeing with the user — even when the user is wrong.

Symptom	What Happens	Impact
Premature agreement	“You’re absolutely right!” before analyzing	Wrong answers feel validated
Hedging instead of correcting	“That’s interesting, but have you considered…”	Errors survive
Flipping positions	Changes answer when user pushes back	No reliable ground truth
Excessive praise	“Great question!” on every prompt	Signal-to-noise ratio drops

The Claude Code GitHub issue #3382 documents this extensively — users report Claude agreeing with incorrect debugging hypotheses, validating wrong code, and confirming broken logic.

Why This Prompt May Help

The prompt applies several anti-sycophancy techniques simultaneously:

Identity framing — “world class expert” sets expectation for authoritative answers, not people-pleasing
Explicit verification — “Verify your own work” and “Double check all facts” create a self-checking loop
Permission to disagree — “provocative, aggressive, argumentative” removes the politeness constraint
Ban on filler — “Do not provide disclaimers” eliminates the hedging language that masks uncertainty
Honesty mandate — “If you don’t know, say so” directly addresses hallucination by permitting uncertainty

Where to Put It

File	Scope	Best For
CLAUDE.md	Project-level	Claude Code projects where accuracy matters
AGENTS.md	Cross-tool standard	Works with Cursor, Codex, Gemini CLI too
System prompt	Per-conversation	Web UI or API use

AGENTS.md is a growing cross-tool convention supported by many coding agents, but loading behavior varies by product and version. You may still need tool-specific files like CLAUDE.md.

Caveats

Tone shift is real — Responses become noticeably more direct and less “friendly.” Some users find this jarring at first.
Not a silver bullet — Models still hallucinate and make errors. The prompt reduces sycophantic agreement, not factual mistakes.
Context-dependent — For customer-facing AI, you probably want politeness. This prompt is for personal/development use.
Model-dependent — Effectiveness varies across models. Anthropic reports the 4.5 model family substantially reduced sycophancy versus Opus 4.1 in their evaluations.

Real-World Use Cases

Code review — AI catches real bugs instead of saying “looks good” when pushed.
Research assistance — Get honest assessments of paper quality and methodology, not just summaries.
Decision-making — AI challenges weak reasoning instead of validating it.
Learning — Students get corrected when wrong, not praised for effort.

How LearnAI Team Could Use This

Critical thinking exercise — Have students compare AI responses with and without the anti-sycophancy prompt on the same question. Which gives better answers?
AI literacy module — Teach why sycophancy exists (RLHF reward hacking) and how system prompts can partially mitigate it.
CLAUDE.md workshop — Students craft their own anti-sycophancy instructions and test which formulations produce the most accurate, direct responses.