Claude Agrees With You Too Much. Put It on a Council.
A Stanford study in Science found AI assistants validate users 49% more often than humans do. Karpathy's answer is a council of models with anonymous peer review — here's how I run the same mechanism in a single Claude chat, and when to upgrade to the real thing.
- #prompting
- #sycophancy
- #llm-council
- #decision-making
Sign in to keep reading
Claude Agrees With You Too Much. Put It on a Council. — A Stanford study in Science found AI assistants validate users 49% more often than humans do. Karpathy's answer is a council of models with anonymous peer review — here's how I run the same mechanism in a single Claude chat, and when to upgrade to the real thing.
Your first guide is free. Sign in with Google or LinkedIn to read the whole library and get new guides as they ship — no cost.
Keep reading
More guides like this one.
Intent Engineering: Stop Writing Steps, Start Writing Done
Anthropic's own docs now say a hand-written step-by-step plan often reasons worse than the words 'think thoroughly' — and that on Claude Opus 5 you should delete the 'verify your answer' line you added last year. Here's the shift, a real before/after, and a four-slot intent template.
The Anatomy of a Prompt That Actually Works
Stop collecting magic phrases. A reliable prompt has five parts, and once you can name them you can debug any bad output in seconds.
Your Knowledge File Fails on Contradictions, Not Gaps
One curated file makes Claude an expert in your domain — until two lines disagree and nothing says which wins. Claude can't detect that, and it can't tell which of your facts expired in March. Here's the skeleton I use, with dates, a precedence block, and the prompt that proves the file was read.