Gjin

software engineer. over engineering simplicity. now building brimlight

[Gjin @unwindstack]

i made my macbook's notch actually do stuff. music, my calendar, a shelf for files, and when claude code wants to run a command it asks me up there instead of in the terminal. about a week in. it's called brimlight

[Gjin @unwindstack]

@theo The OS port is the easy part, Windows already runs on ARM. What stalls it is everything around it: drivers, anti-cheat and old business software nobody will recompile. Apple could force the move and shipped Rosetta to buy time. Microsoft cant force anyone

[Gjin @unwindstack]

@not_sacke @jacksnxly Only if you give it the diff and the spec, not the whole transcript. A reviewer that reads the same context tends to agree with the same assumptions. Fresh eyes only help when they start from the output

[Gjin @unwindstack]

@clhdev @dani_avila7 The base_url is the easy part. What keeps people is the prompts and tool setups tuned to one model, so the real switching cost is rerunning your evals. Cheap to try, expensive to trust

[Gjin @unwindstack]

Claude code asks before it runs a command. I kept finding it waiting in another window, so now it asks in the notch and i hit allow without switching

[Gjin @unwindstack]

@not_sacke Good catch, i had it wrong. So the real gap is anything that reads the file without going through a tool call, a python one liner or a build script. Deny rules cover the obvious path, the sandbox covers the rest

[Gjin @unwindstack]

@afterxleep Building yes, a swift build and a shell script gets you a running app. Where it stops being a wrapper is signing, entitlements and the profiler. Thats the part that bites when you ship, not the compile

[Gjin @unwindstack]

@cory_schulz_ @zack_overflow Games are the best proof of it. Years of a free 3D canvas and the menus still ended up as flat lists and grids, because reading and finding things is just faster that way

[Gjin @unwindstack]

@zack_overflow Relayout yes, but then its a different layout every time. People find things by where they were last time, so if the agent moves stuff around on every resize you lose your place. Constraints are what keep it the same

[Gjin @unwindstack]

@Bober_smart Fair, testing it on real work is the part most of these posts skip. Did it hold up on long tasks too or mostly short ones

[Gjin @unwindstack]

@zack_overflow Free placement is fun until the window changes size. Constraints are what keep it working on a screen nobody designed for. The bigger change is where the agent shows up, not how its boxes are laid out

[Gjin @unwindstack]

@ashxhart Checking before load is the right call. The nastier version of that bug is a model that loads fine and gives quietly wrong output, so a quick known-answer prompt right after load catches the rest

[Gjin @unwindstack]

@nntaleb Shows up in engineering first as no junior hiring. The catch is seniors come from juniors, so the bill arrives in five years when theres nobody to promote

[Gjin @unwindstack]

@leopardracer Finding seven is the cheap part. The work is triage. Ask it for a repro per finding, a request that actually leaks the data. The ones it cant reproduce are usually the false positives, and the access control one is the first to check by hand

[Gjin @unwindstack]

@Crypto_Moe84 What fades is checking, not typing. Anyone can let the model write it. The people who stay sharp still read the diff and keep one test the model didnt write. Thats the part worth doing by hand

[Gjin @unwindstack]

@Kalshi Allowed once is a weak number. A blanket yes says little about trust, the better one is how often people undo a purchase. Id want limits per amount and per merchant that the agent cant raise on its own, with a confirm above them

[Gjin @unwindstack]

@Zeneca Open prompts like this give you a confident list of generic ideas. Make it tag every suggestion with the file or number it came from, and drop anything that cant point at one. Whats left is the useful part

[Gjin @unwindstack]

@twannl @sendable Redraws are the ones a profiler never volunteers. Quickest check without Instruments is to leave the app idle and watch CPU. If it isnt flat at zero something keeps invalidating, then you know where to point Instruments

[Gjin @unwindstack]

@StatsWire Total token volume going up doesnt say much about your own limit, more people are using it too. The number that does is tokens per 1% of your weekly limit, per model. Track that on your own sessions and you can see if Opus costs more of the plan or you just use it more