Aaron Peddle

you have 7 min to escape the permanent underclass

@borisjabes i've been meaning to give @mattshumer_'s new open source project unslop a try. x.com/mattshumer_/st…

what's the state of the art for folks who want their apps to have great design while being claude code maximalists?

@reach_vb environment cleanup / teardown script option for worktrees

We are expanding Codex’s toolkit - what tools would you love to see work with codex the most?

with agents entering the workforce remote first teams seem to be at an advantage

@maxinomics this seems very possible. agents can be very api request hungry so i think we’ll have to see new pricing models emerge to support agent “power users” though

What if AI actually expands the market for software companies that run core services like Slack, Notion, Atlassian, Figma, etc. What is stopping software companies from keeping how they charge per user and then switching to treating agents as a user. Agents are using more of the software than humans in some cases. Just like DeepSeek created the panic around Nvidia GPUs but was actually a huge tailwind for them instead, it’s not unreasonable to think software also has a Jevon’s paradox moment. The only way to immediately dismiss this possibility is if you think companies really are interested in building their own core software. I'm very skeptical of that idea. A well known venture capitalist was talking about this mega agent his team built, runs on it's own Mac. If Slack and all his other services said great you can hook us up to that but it’s an extra $250 per month is he going to say no? Is he going to say take a hike we’ll build a replacement ourselves?

@borisjabes m series macbooks are just so far ahead every other laptop on the market too. I miss my tilling window manager but going from macbook to an x86 laptop feels like going back a decade

i tried to take the omarchy/linux plunge and just the pain of getting vanilla machine up & running with built-in wifi was enough to kill the dream. we mostly talk about imessage/icloud but the windows/macos moat on drivers is underrated.

@JayD0ubleu @jeffreyhuber HNSW for MCP search? how many MCP servers do y'all have running 😬

@jeffreyhuber that's used in the new MCP tool search

i think an under rated factor in why developing with anthropic is easier is they don't block agents from reading their docs like openai does. if i'm picking between anthropic and openai for a new project i'll pick anthropic all else equal for this reason

@mattshumer_ @joshgonsalves_ @openrouter very curious to hear your review of GPT 5.2 in claude agent sdk harness. if it works better in some cases i wonder how an agent alloy would fair (xbow.com/blog/alloy-age…)

@joshgonsalves_ @openrouter You're definitely going to want something really capable to make good use of the harness. I'm personally starting by trying the GPT-5.2 series of models

FYI you can use ANY model with the Claude Agent SDK Just change these three environment variables!

just had an incredibly disorienting realization that o1 was released only 1 year ago

the fact that claude code treats '?' in an empty input bar as a hotkey for help is making it a lot harder for me to prefix messages with "??????" when clod tries to do something stupid

this is why i pay $200 a month

@erikdunteman Agreed. anything "new" i don't use AI for nearly as much. For my own enjoyment it's kinda great. the code that is the most boring and repetitive i can just have agents do and i get to focus on the new and interesting stuff

@AaronPeddle it's the core concepts I get most hung up on. Once some foundations are set and there are clear patterns to follow (especially trained-in patterns like React) the models do better.

I still don’t like coding agents. At light-medium prompting effort they produce garbage code. At full prompting effort, it’s passable, but at the time spent typing the perfect spec or code review, it’d just be faster to code it myself (+tab assist)

@erikdunteman also with claude code i find you can get significantly more mileage by explicitly asking to use sub agents and sub sub agents

i think it depends on the task. for react frontend i use coding agents for all components while I focus my attention on making sure the types are defined well (minimizing illegal states) and reducers are clean and well written. for backend i use a plan / implement pattern (you don't have to write the perfect spec, an agent writes it and you just give feedback until it's good) for experimental / core concepts - i still hand write ~80%+ of the code

@erikdunteman one reason is terraform's DSL is helpful to maintain code consistency across a team without having to have as strict of a code style review process

Other than compat/status quo, I don’t understand why someone would choose terraform over pulumi Every pulumi feature is self documenting through types, it’s just so easy to know what’s needed

AI Slop update: I've added an HTML rendering pipeline for my AI generated tiktok videos letting the AI make nicer looking titles.

test render of adding an AI avatar to my slop generator lol

I've been spending way too much money making these

POV you're scrolling ur fyp to chill but come across my ai slop instead