Context advantage that’s what puts humans in the driver seat. AI will work better when it has this context advantage
8
Adding to your prompt “this is a test question” makes the reward function of the model go bananas.
20
this week the new roadster hopped over a cybertruck. the og one has been orbiting the sun since 2018
what should my spacex dashboard add next?
🚗 where's the roadster right now
🛰️ starlink train spotter (can i see one tonight?)
🚀 starship launch-day mode
32
til vercel's hobby plan only deploys commits from the account owner, so a github actions bot commit never goes live. fix: the job commits as me. on paper it runs at 7:17 every morning. today it didn't, so i ran it from my phone at 9:04
32
most side projects don't need a backend. my spacex dashboard puts 11k+ satellites on a live 3d globe with zero api keys and zero env vars. a scheduled job writes json, the browser does the orbit math. that's the whole stack
24
ui hot take: if two things on screen look like time, they'd better be the same clock. my globe spun on its own while the time slider sat still, so it looked like time was passing when it wasn't. now one clock drives the earth, sun, satellites and the slider
17
yesterday i dropped a zip of a side project into claude code. by dinner it was a live spacex + starlink dashboard: 3d globe with 11k+ starlink sats, launch countdown, booster leaderboard (b1067 has flown 37 times), space weather
spacex-daily.vercel.app
2
45
underrated AI feature: being allowed to say "nothing to do"
my new automation scout read 68 of my sessions today and proposed... nothing
correct answer after one day. most agents would've invented 3 ideas to look busy
13
hot take: AI session transcripts don't belong in your git repo
they hold command output, file contents and your home folder paths. one push to a public repo and it's all out there
keep them local, or in a separate private repo with your secret checks in front
11
hot take: stop writing step-by-step prompts for your agents
the next model will be smarter than your steps
built a weekly Claude routine today that gets goals, not instructions. new model drops, I change one setting and the whole thing gets smarter
1
1
45
hot take: half of "AI infrastructure" should just be git
I tried having Claude upload my session data every hour. turns out every upload needs a human to approve it
switched to a plain git push. 0 tokens, no login, nothing to approve
1
22