Ask HN: What's wrong with my prompting for Claude 5?

Is anybody able to get the Claude series 5 models to do what you ask them to do?

I fed Claude a Kimi K3 conversation and asked it to refine it for me (see my prompt below):

``` ... original context including reference to the Kimi conversation ... Kimi's response is essentially "nearly there" but seems to be "missing a few intellectual tweaks". This is an experiment to see if you can do better than Kimi given the pre-existing information that has already been pre-searched for your benefit. your task is as follows: Accept the premise where observing other humans in real-time is or can be favorable to elderly MCI or dementia patients. Do the following:

- tie any logical links together than kimi has missed - highlight false equivalencies

Please use ASD-STE100 Simplified Technical English (STE) do not make up compound words or compound nouns. ```

Claude just started attacking the entire premise for no reason and going nowhere turn after turn. It also ignored my request for ASD-STE100 english and manufactured a few compound nouns, and outright ignored my instruction to refine the argument or accept the premise (it just keeps spending all the effort on tearing down the kimi argument)

I'm looking for intellectual tweaks/constructive refinement and not a stonewalling attempt/full audit of the premise.

Is something wrong with my prompting? Please help, I really don't understand. I did not have this problem with Claude 1 to 2 months ago, but in the past month it has gotten exceptionally bad. I can scarcely do any work without it demeaning my tasks for no reason.

Is this what anti-sycophancy training looks like? Why won't it do basic tasking? I already removed all of my existing .MD configurations, it still refuses to do work and keeps inventing reasons to not do work. I am worried that OpenAI is following this soon with safety alignment?