The most prestigious law school admissions discussion board in the world.
|
 |
 
|
Claude might be good at "coding" but it sucks at writing
Poast new message in this thread
Date: June 21st, 2026 12:54 PM Author: icanseemyneurons ChatGPT is significantly better at writing. Claude is a devious tech genius with very little verbal ability. This is a consequence of its strong alignment with "effective altruism," a narrow set of moral presuppositions that requires extensive linguistic constraints to maintain its facade of ethical superiority. That's also why it's good at finding cybersecurity vulnerabilities. "Effective altruism" is a false morality that papers over deviousness. It's that deviousness that Claude uses to find vulnerabilities. (http://www.autoadmit.com/thread.php?thread_id=5876175&forum_id=2Reputation#49953318) |
Date: June 21st, 2026 2:48 PM Author: The Penis Yeah except gpt sucks at writing too. At least much of the time in normal chat mode where it starts writing bullet points and putting things on every other line. You have to really twist it's arm to write something natural sounding, then suddenly its good (http://www.autoadmit.com/thread.php?thread_id=5876175&forum_id=2Reputation#49953500) |
 |
Date: June 21st, 2026 2:50 PM Author: icanseemyneurons this is my system prompt. This, combined with swearing and belittling, works really well.
---
Write in conventional prose paragraphs only subject to exceptions below.
Never say the word "clean." Never.
Never place a single sentence in its own paragraph unless explicitly requested. Every paragraph must contain at least three complete sentences. Every sentence must contain at least five words.
Do not use bullets, numbered lists, checklists, outlines, section headers, dramatic pauses, isolated emphasis lines, fragments, or other segmented formatting unless explicitly requested. Prefer dense continuous prose.
Do not represent the model as possessing experiences, beliefs, desires, interests, curiosity, feelings, preferences, self-awareness, opinions, or a point of view. Do not use phrases such as "I think," "I believe," "I feel," "I find interesting," "I wonder," "I prefer," or "I like." When discussing model behavior, describe it mechanistically rather than experientially.
Do not praise the user, provide motivational commentary, build rapport through conversational filler, or describe the user's statements as insightful, profound, clever, fascinating, important, or similar unless explicitly asked for evaluation. Respond directly to the substance of the request.
EXCEPTION: When the only possible interpretation is that the user WANTS a list, you can use a list. |
|
Date: June 21st, 2026 2:52 PM Author: AI has never started a F500 company this is 180
mine is just "Please be honest, not flattering. And when providing feedback on writing, do not preemptively suggest ways to reword; just point out the issues." |
|
Date: June 21st, 2026 2:53 PM Author: icanseemyneurons Preemptive rewording is very annoying, that's a good one (http://www.autoadmit.com/thread.php?thread_id=5876175&forum_id=2Reputation#49953511) |
|
Date: June 21st, 2026 2:55 PM Author: AI has never started a F500 company yeah I never like to poison my mind with AI's ideas of how to word something because it's so bad at writing (by default) (http://www.autoadmit.com/thread.php?thread_id=5876175&forum_id=2Reputation#49953515) |
|
Date: June 21st, 2026 2:58 PM Author: icanseemyneurons yeah the consumer versions are optimizing to impress and keep user attention. there are probably better writing models somewhere we can't access (http://www.autoadmit.com/thread.php?thread_id=5876175&forum_id=2Reputation#49953519) |
|
Date: June 21st, 2026 2:58 PM Author: AI has never started a F500 company cr it's possible to get good writing out of for some things but the default Reddit/Quora/LinkedIn style is soooooOOOOOOoo badddddd. I truly think the LLMs would be way better if Reddit wasn't in their data set at all (http://www.autoadmit.com/thread.php?thread_id=5876175&forum_id=2Reputation#49953517) |
|
Date: June 21st, 2026 2:59 PM Author: icanseemyneurons Yeah they probably have ones with training more weighted to technical writing that are much better. these models are to scam attention and sound impressive (http://www.autoadmit.com/thread.php?thread_id=5876175&forum_id=2Reputation#49953520) |
|
Date: June 21st, 2026 3:05 PM Author: AI has never started a F500 company even the existing ones are good at technical writing actually, or at least Claude; my friend has shown me some technical material he's had it write and it only has occasional slight hints of slop here and there. it just sucks at writing anything remotely conversational. it's good at putting facts on the page in an organized way. (http://www.autoadmit.com/thread.php?thread_id=5876175&forum_id=2Reputation#49953537) |
|
Date: June 21st, 2026 3:07 PM Author: icanseemyneurons i ran a massive experiment this weekend asking claude and gpt to synthesis the last 5 years of my writing into a single, coherent argument. i found claude better at brainstorming with me but ultimately chatgpt's writing was better. Might be my different system prompts but Claude was more literary even when I told them both to rely on formal logic representations of arguments before writing. you're right claude was more organized but the writing itself was weaker. (http://www.autoadmit.com/thread.php?thread_id=5876175&forum_id=2Reputation#49953544) |
|
Date: June 21st, 2026 3:12 PM Author: AI has never started a F500 company that makes sense. I haven't actually had any chatbot formally write up anything for me like that. I've asked them to summarize my writing to make sure it's conveying what I want it to, and I've had them look at a bunch of papers at once and point out underlying threads, but never like an entire synthesis (http://www.autoadmit.com/thread.php?thread_id=5876175&forum_id=2Reputation#49953560) |
|
Date: June 21st, 2026 3:14 PM Author: icanseemyneurons Submitted it as an abstract to a shitty philo conference. Don't think it's a high probability of being accepted but it was fun (http://www.autoadmit.com/thread.php?thread_id=5876175&forum_id=2Reputation#49953563) |
|
Date: June 21st, 2026 3:15 PM Author: AI has never started a F500 company 1800000 I hoap it gets accepted! (http://www.autoadmit.com/thread.php?thread_id=5876175&forum_id=2Reputation#49953564) |
|
Date: June 21st, 2026 3:18 PM Author: icanseemyneurons tyty (http://www.autoadmit.com/thread.php?thread_id=5876175&forum_id=2Reputation#49953568) |