Jonny Buchanan

Local folder access working in Chat Gippety

I love the smell of free inference you can actually do stuff with in the morning

Using free Claude dot ai to edit my local files, making it fix the folder picker I've added to pick the folder it has read and write access to 👀

How do you find out what's staged? You look in the unstaged area and try to figure it out. WTF. These must always be separate Back to the cherry-pick, checkout and hard reset mines for me

Tried GitButler because I'm 4 branches deep on something, and the UX for creating commits is so very weird Select some lines you want to stage and right click? That's a de-selectin' - they want you to click individual lines or click and drag in the gutters. WTF

git-gui is the perfect UI for crafting commits, and if your GUI tool doesn't let you do everything git-gui does, it's no bueno

Assistant on my phone just got replaced with Gemini without warning. Free inference! *slaps roof of car* this badboy can fit so much TypeScript in it

My excitement when I discovered @cloudflare/vite-plugin and my favourite Vite + Hono combo came to the app I'm making and worked first time

@kenwheeler This is the way, literally working on this yesterday Run the same code every time working with the data set from hell, go to the agent with no-matches - it found someone whose name doesn't match their email (married) or their…, uh, name? (name in employee DB is a nickname lol)

agent driven workflows with a sprinkling of deterministic tools are a fools errand. especially with subsidization fading, expect to see deterministic workflows with a sprinkling of agent nodes, doing what they’re actually good at.

@luciascarlet Sadly, part of any kind of moat they have is that GPT-5.x suucks

@ryanflorence Short horror story: it was by GPT-5.X

Just merged my own 23k line PR AMA

90 minutes of everyone sitting in a room and coming up with the best course of action can be completely invalidated by someone taking 2 hours to build a prototype (_genuinely_ throw-away ones are now both fast and cheap!) The last time I experienced this personally? Wednesday

Plan → Build doesn't feel like the most effective way to use agents to me, for the same reason it's never felt like the best way to write software: the process of building can invalidate the plan _so_ quickly

Not me thinking about implementing my own browser-based agent for personal use because the internal one I made for work is a better experience than ChatGPT and Claude

I added this to our internal multi-model LLM chat app last week and as soon as it could edit its own source to add more tools it became a pretty effective coding agent I added branded clientside Word, PDF and PPT generation today, just with the basic edit tools it wrote on day 1

Browser-based agents you can give read/write access to an entire directory to via <input type="file" webkitdirectory> are so good, is that why vanilla Claude and ChatGPT don't offer that feature (that I can find)?

GPT-5.4's whole schtick is blabbing to give the impression of competence while generating the worst code you've ever seen

If you need me I'll be listening to this on loop until I get it out of my head youtube.com/watch?v=3c3QfD…

I work on an internal browser chat app, <input type="file" webkitdirectory> and tools which take a directory handle are effective as hell

Yesterday at work I had The Experience everyone who's ever written their own LLM harness has the first time you make it capable of editing its own source code while it's running Fuckers _still_ can't write a useEffect for shit, still a huge source of you're absolutely right