Waldemar Panin

@craigzLiszt best way to give you a psychosis!

if you’re serious about your time, you need to confront the hard truth every day: your life is finite

@unnamed1tw LGTM

Cursor ran agents for 3 weeks to migrate the codebase from Solid to React. +266K/-193K lines of code. Would you trust merging a PR like that? What would you check first?

@kr0der @cursor_ai wb

we're so back, resubscribed to @cursor_ai ultra after contemplating my AI workflow over the past few weeks, having no "CMD Y"/"CMD N" approval flow makes me slower overall i realised having AI one shot a task and then having to review 1000 lines all from a git diff viewer is actually slower than just approving every change via CMD Y when you use the CMD Y/CMD N approval flow, you get to see each change easily and at the same time you're basically doing a self-PR review i still hate the agents sidebar but overall, cursor's still the best AI IDE by far, so we're back 🫡

claude, increase my vo2max to 53, make no mistakes

@MrAhmadAwais Long term tasks, large codebases, no guardrails/rules set in place

@chiefwalde mostly when do you see hallucinations nowadays?

Genuine question for devs using coding agents: How many times a week do you correct your coding agent? > Wrong package manager. > Wrong file naming. > Wrong framework assumptions. Feels like we're missing something obvious here.

@MrAhmadAwais try a small file with hard constraints like these and you get way less hallucinations :)

@chiefwalde Hardly ever tbh.

@0xPaulius Better tell it to generate a coffee machine build plan for tomorrow morning too lmao

ralph just gave me 101 item build plan 😳 overnight it is

@_Evan_Boyle For automating eval, have you found certain rule sets or 'skills' in the agent prompt make it easier to then translate those manual observations into a consistent test suite?

Feedback loops are the thing that makes coding agents work. Establishing them for prompt / tool changes in an agent can be hard. A loop that’s worked well for me: Session A: build/change the agent Session B: run a scenario where the new behavior should trigger If it doesn’t: ask the agent to critique its decision path (“why didn’t you try X / call tool Y?”) and what signal would’ve made it choose differently Feed that critique back into Session A, adjust prompt/tooling, repeat Do it manually first, then turn it into an automated eval. High signal in practice.

@stevekrouse Agreed on the force multiplier for data collection, but Buffet's core insight often comes down to qualitative judgment and deep context. Will these agents truly 'know' a company, or just have access to infinite facts, which isn't quite the same thing?

Warren Buffet says that one investor can really only know 5-10 companies, and value them properly I wonder what the future of public market investing will look like when you can unleash lots of analyst "agents" to be constantly research every part of every company - including every person at the company, on the board, etc, etc

@mischavdburg the real skill is less about writing perfect code and more about writing perfectly clear instructions for whatever is going to build it -harder than it sounds sometimes!

modern software engineering has now become writing specs and requirements

@DavidKPiano Or maybe, what if average devs had a supercharged pair who still needed explicit, well-defined rules to be truly effective?

AI coding assistants are not "what if senior devs were automated" They are "what if average devs were really fast"

@forgebitz It's interesting how many founders chase the tech/product they personally find cool, without thinking if they actually align with the market's needs or culture.

"founder market fit" this is one of those things people don't often talk about but very often i see people dive into a space because someone else is making money but if you don't like sales people, don't make a CRM if you don't like data, don't make an analytics platform entering a market with a twist is great, but always make something for a market you love building for so many ideas just fail because founders copy others and figure out halfway through that they have no idea how the market works, and they don't even like their customers

@0xDesigner The shift makes sense given how much easier it is to prototype and ship without a huge dev team. But even with that, the hardest part for those indie-designers is usually distribution, not just building the thing.

i can only imagine the market rate for product design is only going to go up from here. working theory but the pipeline from product designers to indie dev is going to explode. the supply for good product designers looking for a job will shrink.

@Rasmic like 3d game engines @thekitze 👀

i don't like kanban board as a means of deploying agents there's better ux imo

@pvncher thats only because it wants to get it exactly right and not miss anything - unlike Opus who just wants to rush things and get them done half-assed but fast

Just experienced one of the quirks of codex - "Build the spec" turned into spending 20 minutes fleshing out my spec markdown vs implementing the code...

@acdlite I've found it helps a lot when you're just brainstorming or trying to articulate a complex thought, less friction to just speak it out rather than type!

If you're not using voice transcription as the primary way to chat to coding agents, you really should. I wildly underestimated how much more productive (and enjoyable) it would be compared to typing.

@corbtt its for sure gonna shift from reviewing the generated code to reviewing the prompt engineering & the guardrails/rules setup that led to it. that's where the critical thinking moves, imo.

It still pains me a bit to say it but the "humans should not be reviewing code" side is definitely going to win. Even if the models never get better than today (they will). You carefully review the design docs, and the testing plan. But not the generated code.

@Rasmic it's less about the model 'knowing' frameworks and more about how those 39 'skills' guide its application of that knowledge to a very specific context imo

I don’t know how I feel about stuffing 39 different skills Unless it’s very minimal and pattern guidance the models already know how most frameworks work

@round what *will* matter then? Or do you mean the models just absorb all skills intrinsically?

subagents & skills won’t matter in 6 months