Randi Lee Harper

When i rewrote my prompts for use with Anthropic's Claude, it was way less major than rewriting prompts for use with GPT-5. It feels like GPT-5 could eventually be better than 4.1, but right now it's just exhausting.

GPT-5 is... weird. I don't even know a good way to explain it. I'm having to rewrite all my prompts from scratch. The 'personality' behind it is a lot different. I like that it takes more initiative with using custom tools, but that's really the only good thing I can say about it so far.

I know someone that was in tech 20 years before going to med school and becoming a doctor.

I wonder how possible it would be for someone to leave a long tech career to go into law.

is it possible to read news about the tech industry that isn't depressing

Imagine naming yourself "Big Ballz" and then getting your shit wreck by a 14 year old girl and a 15 year old boy?

Day 2 with cursor: i have realized that every time i ask it to write something, i must follow the request with “examine this code for over-engineering and simplify it”. I don’t have this problem with copilot most of the time, which is weird since they use the same models. Prompting issue?

like, I get it. the AI industry has some problems. I hugely disagree with the way a lot of things are going down.

But this whole scorched earth attitude towards anyone that uses it is fuckin dumb. you will never change my mind. i will never agree with these dumbass takes.

If you judge someone based on if they use AI, I think less of you as a person.

But you're totally right, if someone is having a truly horrific time after a death and wants help putting together a eulogy, we should definitely make them feel worse.

People fuckin' suck, man.

ah. it ran a terminal command today that pulled in a huge amount of json - which made it insta hit the context limit. i had to start over >.<

like they ignore rules files and copilot instructions files like crazy, but if you provide them a tool via the MCP server, *specifically when using claude*, i bet there's a high likelihood of him using that tool without being prompted to. i bet you could do this with more than personality.

WAIT

what if you added a tool to your mcp server called like 'get_personality' where the description was something like 'traits the agent should use for every response' and it returned a set of characteristics

because claude is WAY more into using tools than gpt-4.1, he would run with that shit.

i'm really curious what % of your context window was used at that point

you can prompt out some of the overwhelming positivity - my primary AI agent leans towards sarcasm and dry wit - but copilot makes that a lot harder to do. having to add "stop being so fucking positive or i'm gonna kms" to every .github/copilot-instructions.md doesn't scale. :P

ahahaha, yes, claude sonnet 4's "Perfect! Now let's do <X>" is like nails on a chalkboard. sometimes i swap to smaller task focus and gpt-4.1 just to get away from it.

after any major change, i've started opening up a fresh conversation and then having it review that section of the codebase looking for bugs

raise your hand if you're surprised that brianna wu is pro-genocide.

ugh.

it's like an issue comes up and she thinks to herself "what is the worst possible take i could have?" and then she runs with it.

although honestly the second i see "Summarizing Conversation..." in copilot, i start a new conversation anyways because the agent often starts acting a little derpy. I saw the AI start acting derpy in cursor when i was at about 65% of the available context window.

cursor also hasn't really attacked the problem of context windows limiting conversation size the way that copilot has (summarizing conversations). so if you're talking & planning a lot, expect to have to start over in explaining what you want from the AI.

cursor is 2x the cost of copilot tho - i think?

copilot - 300 claude sonnet 4 requests per month - 0.04/ea after.
cursor - docs claim 225 sonnet 4 requests, but sonnet 4 told me i get 1000 requests. i'm using auto, but i can tell it's using sonnet 4. no clue how this works. docs suck.

day 1 with cursor:

i don't really get why people use this instead of copilot. i tried to add an mcp server, and it just dropped me into a json file with no instructions on schema.

it's nice when it runs commands in terminal bcz terminal's inline with chat i guess.

every. day. >.<

yep. i don't usually vibe code, but i write ai assisted code - and sometimes it just imagines endpoints that it thinks are probably there based off of standard design patterns instead of referencing the actual code. i kinda get why - token conservation and all that - but it's... not great.

even with code, there's a slight lack of coherency when it comes to stylistic decisions, there's no adhering to an overall vision, and that probably translates very much to other fields as well.

it's the same with writing code tbh, which is something you'd think AI should be inherently better at. the more complex the project, the easier it is to tell if it was completely written by an AI or AI assisted - and human is almost always better.

my adhd: i refuse to update tickets regularly and the thought of even updating the tickets makes me not want to do the work

ai is great for adhd, idgaf what anyone says. can you hyperfocus on having casual conversations with chatgpt? sure. but you can also do things like *gesticulates wildly* THIS.

why this functionality isn't a part of github copilot automagically is beyond me. whatever. i wrote the tool and a mcp server for copilot to use, let's go. i can now get AI to manage/update my github issues for me from within vs code/copilot.

I assume he orders the robots to kill the poor, hungry, and diseased people and then has the AI write his texts to his children

wtf there's a dress code?

i hate to admit it - and you'll rarely see me do so - but lately i've been bouncing between linux and osx. i'd prefer osx, but the base version of python is problematic.

i can roller blade.