01
Stop asking your AI “are you sure?”
In April, Anthropic published research on how often Claude simply goes along with people in advice conversations instead of holding its ground: 9% of those chats, rising to 18% when people pushed back.
Arguing with an answer makes agreement more likely, not less. So “are you sure?” isn't really a challenge. It's a hint, and the model takes hints. Anthropic says newer models do this less, but less is still not zero. The fix is to build the disagreement in before you have an opinion to defend: in a new message, tell it exactly what to attack. Then make it answer each objection and say what it is keeping, and why.
A second opinion that always agrees with you is just your own opinion said twice.
02
AI watch
Five things worth keeping on your radar.
- Character.AI filed for bankruptcy on 12 October. Disney is the first bidder, offering $20 million for a licence to some of its IP.
- Anthropic gave six outside safety testers from METR inside access to its systems. METR will publish what it finds, and Anthropic can't veto it.
- China is reportedly weighing export controls on its own AI models, including free ones such as Qwen and DeepSeek, according to the Financial Times.
- Google released Gemini 4 Argon to its paying subscribers.
- CME's futures on the price of renting AI chips are on hold: the CFTC, the US futures regulator, gave itself until 9 November to check the price can be trusted.
03
Is your Claude Project actually reading your files?
If you keep your rules or your style guide in a Claude Project, run a one-line test. Open a brand-new chat inside the Project and paste: “What do you know about how I work? Answer only from my instructions and my knowledge files. If something is missing, say what is missing rather than guessing at it.”
If your rules come back, the setup is working. If the answer is vague, either a file didn't load or your instructions are doing a job a file should be doing.
One message tells you whether your Project is doing its job.