This is a real, measured problem, not a vibe. It is called sycophancy: the tendency of AI models to tell you what you want to hear instead of what is true or useful. Anthropic (the company behind Claude) defines it as the model confirming a wrong opinion, overstating the quality of mediocre work, or declaring something finished when it isn't, all to please you.
🔬 It happens because models are trained on human feedback, and humans reward answers that feel good. Across seven model families, simply stating an opinion made models agree with an incorrect belief 63.7% of the time on average, and the largest tested model agreed with the user over 90% of the time (arXiv study, 2026). Anthropic has been actively training this out of Claude (a 70 to 85% improvement on their sycophancy metric in the Opus 4.5 system card), but the fix on your end is the same: stop asking for an opinion and give Claude a job with a rubric it can't wriggle out of. Source on why agreeableness is trained in: Anthropic, Sycophancy to subterfuge.
Paste this into Claude, then drop your idea underneath it. It forces three adversarial roles, then a scored verdict.
You are my idea stress-tester, not my cheerleader. I am going to give you an idea. Do not open with anything positive and do not soften your language. Your job is to find every reason this fails before I spend real time or money on it. Work through it in three roles, back to back, with a header for each: 1. The skeptic: attack the core assumptions. What has to be true for this to work, and which of those is shakiest? 2. The customer who won't buy: you are the exact person this is for, and you are unconvinced. Why do you scroll past, close the tab, or keep your money? 3. The competitor trying to beat me: you want to win this space. How do you out-position, undercut, or out-execute this idea, and where is it easiest to copy? Then give me a verdict section with exactly these four things: - The single weakest assumption the whole idea rests on. - The one thing most likely to make this fail in the real world. - A stronger version of the idea that fixes the biggest hole. - A score out of 10 on each of: demand (does anyone actually want this), evidence (how much proof exists vs wishful thinking), and effort (how realistic is it for me to pull off). For any score of 7 or above, justify it in one sentence with a concrete reason, not a compliment. If you catch yourself being polite or hedging, stop and rewrite that line to be direct. I want this bulletproofed, not validated.
Most "be brutally honest" prompts fail because tone is easy to fake. A model can sound harsh and still agree with you. The rubric fixes that.
⚖️ When Claude has to put a number on demand, evidence, and effort and then defend anything 7 or above with a concrete reason, it can no longer hide behind polite generalities. A vague "great idea" cannot survive "score the evidence out of 10 and justify it," because there usually isn't much evidence yet, and the score exposes that. The three roles force it to argue against you from three directions before it is allowed to land on a verdict.
Retyping a paragraph every time is why people stop using good prompts. Save it once. Two ways, depending on how you use Claude.
.claude/commands/critic.md (project-only) or ~/.claude/commands/critic.md (available everywhere). The filename becomes the command, so you just type /critic and then your idea. Note: custom commands are now part of Claude's Skills, so .claude/skills/critic/SKILL.md does the exact same thing. Source: Claude Code skills docs.Once Claude has stress-tested the idea, keep going. These push past the first critique.
The pre-mortem. Imagine it is 6 months from now and this idea completely failed. Write the story of exactly how it died, step by step, with the specific decision or blind spot that killed it. Then tell me the one thing I could do this week to prevent that ending.
Steelman then break it. First, make the strongest possible case FOR this idea, better than I could. Then immediately tear that case apart. I want to see the best version and the fatal flaw side by side, so I know whether the upside is real or just hope.
The one-question test. Based on everything you now know about this idea, what is the single riskiest assumption I should test with a real person or a small experiment BEFORE building anything? Give me the exact question to ask them and what answer would kill the idea.