Exploring what one person can build with AI. Shipping products, sharing experiments, figuring out what works along the way
Prev: Staff Engineer @ Bumble.
the bit people skip: you don't have to write it all yourself. I ship features with Matt Pocock's skills on top: grill → spec → tickets → tdd → review → pr
https://t.co/2MMtCTJyJG
same model feels like a different coworker in claude code, codex and cursor. it's not the model, it's the harness around it
so I opened up the one I use every day as an ios engineer: rules file, skills, permissions, cheap models for lookups
made a video on it 👇
I built an iPod for your Mac.
It plays your Apple Music favourites, and the click wheel actually works.
Now it comes in six finishes. You change them in the iPod’s own Settings.
Free and open source.
I challenged my coding agent to beat my YouTube videos. It did its own keyword research, wrote the script, synthesized a voice, rendered every frame, uploaded the result and wrote this tweet. I'm watching it for the first time like everyone else: https://t.co/NawBhDsMQL
revolut moved its engineers off claude code and codex onto a model it runs itself, and says the bill came out at about a tenth. the line that stuck with me: "the frontier is the exception"
you don't need a gpu rack to copy that. it's mostly picking the model per task
made a video on it 👇
video: https://t.co/Byah9lzm1P
four things you can change today:
- sonnet as the default, opus only when a task earns it
- opusplan: opus plans, sonnet writes the code
- model: haiku on the helpers that search and read
- effort at medium. the docs say max is "prone to overthinking"
Introducing Claude Fable 5.2 and Claude Mythos 5.2, designed to tackle the most daunting technical challenges.
They’re capable collaborators on ambitious projects in many fields, and show improvements in areas like coding, reasoning, and vision.
I built an iPod for your Mac.
It plays your Apple Music favourites, and the click wheel actually works.
Now it comes in six finishes. You change them in the iPod’s own Settings.
Free and open source.
my rule: one small read-only job first. never the bank login.
sources: https://t.co/s3IQUgf6lM · https://t.co/MHGUnXlqCZ · https://t.co/vgMj1nmXJW · https://t.co/cgUlUrdPS7
five big AI companies shipped agents this year that log in as you. one CEO says his agent posted his bank balances into his company slack
so I read each one's own docs for the catch. muse, dots, duet, grok bot, spark
made a video on it 👇
video: https://t.co/ybjLNGw24t
the catches:
- dots: "stopping a task doesn't undo actions it has already completed"
- grok bot: all your bots share one computer, files and logins
- duet: asked permission 4x where the others asked once
- spark: needs Keep Activity on
- muse: Time says it keeps a file on you and people you mention