Thanks for extracting this.

I always enjoy examples of prompt artists thinking they can beg their way of the LLM's janky behaviour.

> Critical UI Requirements

> Therefore, you SHOULD ALWAYS test your completion requests first in the analysis tool before building an artifact.

> To reiterate: ALWAYS TEST AND DEBUG YOUR PROMPTS AND ORCHESTRATION LOGIC IN THE ANALYSIS TOOL BEFORE BUILDING AN ARTIFACT THAT USES window.claude.complete.

Maybe if I repeat myself a third time it'll finally work since critical, ALL CAPS and "reiterating" didn't cut the mustard.

I really want this AI hype to work for me so I can enjoy all the benefits but I can only be told 'you need to write better prompts' so many times when I can't see how that's the answer to these problems.

Maybe if we added ALWAYS BE RIGHT NEVER BE WRONG it will work this time?

Can't argue with math!

here we have a glowing example of humans not injecting their baggage into every conversation

Not supporting White Supremacy and Holocaust denial is "baggage"? I'd hate to know what kind of baggage you're carrying around.

There are loads of things I don't support, mostly I don't bring them up in unrelated conversations.

And I don't get my panties in a bunch and tone police when somebody says something mean or funny about fascists.

Your behavior is not conducive to the desired environment.

https://news.ycombinator.com/newsguidelines.html#comments

We've learned this the hard way working with AI models, yelling at the models just doesn't work:)

I would think someone working for Anthropic would be quite aware of this too.

Either fix the prompt until it behaves consistently, or add conventional logic to ensure desired orchestration.

Totally agree. We’ve seen similar weirdness when trying to build deterministic behaviors around LLMs. It’s fun at first…. Until you’re debugging something that just needed a if/else. We’re now mixing prompts with conventional logic exactly for that reason, LLMs are powerful, but not magical.

if you hire someone are they going to always be right the first time you give them directions?

"Large language models don’t behave like people, even though we may expect them to"

https://news.mit.edu/2024/large-language-models-dont-behave-...

> You MUST include the ENTIRE conversation history in each prompt to Claude, not just the last message.

> ALWAYS include ALL messages from the beginning of the conversation up to the current point.

That doesn't seem very scalable.

If you're building full blown chat apps where that length limit matters using Claude Artifacts you should graduate to something less restrictive.

Can you explain how you come up with such a prompt, especially the part with underscores?

Take a look at my full transcript here: https://claude.ai/share/42b70567-8534-4080-9227-b834e8c13d6e

Claude tipped me off about the name of that section as part of its thinking process in answer to my first question.

Thanks!