tastemaker is slowly becoming one of the best ui skills there is
> 200+ github stars
> fixes ai slop
> works with the coding agents you already use
> builders are already shipping with it
you can ignore tastemaker at 200 stars
or use it now before everyone starts copying the taste
https://t.co/RkeWtS6ZpJ
an airline just turned away customers who were trying to give it money
united banned ai agents from booking its flights. one day after openai bought an ai travel agent
they didn't block search. they blocked the buy button
> united[dot]com, the app, travel agencies and aggregators: agents can't book
> google flights and kayak still show the fares
> reason 1: your login shared with an agent can leak
> reason 2: an agent can buy what you didn't mean, or spend your miles
> reason 3: at check-in, a real person has to confirm what's in the bag
the part most people miss: the ceo said it out loud. "i don't want to create another middleman"
whoever books the flight owns the customer
my agents draft my posts and dms. nothing goes out until i say yes
how it actually works, ep 14 ๐
every other day there's a new agentic video editor
that's the hype right now, so everyone's building one
i'm not
every episode of this series is edited end to end with claude code on the $20 plan
sync, color grade, captions, music, animations, one command
you don't need a more powerful agentic editor
i've already shared my workflow, follow it and you're good to go
meta just bought an app 100,000 developers use every day
and then shut its cloud down
they didn't want the product. they wanted the people who built it
> conductor runs claude code and codex side by side, each in its own copy of your code
> meta's own engineers were already using it
> meta has its own coding tool now, muse code
> sign-ups stopped nov 4, read-only nov 13, cloud off and data deleted nov 20
the part most people miss: a refund doesn't move your workflow
if your agents' work lived in that cloud, you had days to get it out
i run several claude code sessions at once for this series. every bit of that work sits in git on my own machine
how it actually works, ep 13 ๐
Goldman Sachs just took Claude Code away from most of its traders and bankers
Only ~12,000 engineers still have access
Interesting part?
Back in July, their tech chief was openly praising vibe coding by non-technical employees
So what changed?
> A tool vibe-coded on a Tuesday could end up touching client data or real positions
> Nobody reviewed the code
> Nobody owned what it shipped
> And when regulators come knocking, someone needs to be accountable
But here's what most people are missing:
Goldman didn't ban AI
Its in-house AI tools are still available to everyone
They just restricted Claude Code to people who can actually read, review and take responsibility for the code it generates
AI adoption isn't the problem. Uncontrolled AI adoption is.
measure usage and you get more ai, not better software
same thing happens with ui
that's the check tastemaker runs: it reviews ai-generated ui against real shipped sites before it goes live
proton sells privacy and security
its ceo called an engineer's security concern "loser thinking"
then planned to re-interview almost the whole company about how much they use ai
the day after it went public, he cancelled all of it
here's what he actually got wrong
> he wasn't wrong about speed. agents can write code fast
> but proton measured how much people used ai
> measure usage and you get more ai, not better software
measure what ships and what breaks, not how many prompts someone wrote
i cut my ai budget in half on purpose
turns out i'd been using the most expensive model to fix typos
with a limit in place, i started routing:
> open-source models for easy work
> the big model for hard problems
side projects let you ignore cost. production doesn't
@Apple just made the same move with its engineers. watch why ๐
apple was letting its engineers spend up to $300 a day on ai
now it's $200 and some teams are going down to $100
but most engineers weren't even using the full $300
so the cut barely saves money. here's what it's actually about
> when ai feels free, everything goes to the biggest model, even a one-line fix
> with a budget, you start choosing
> cheap or local model for easy work, the big one for hard problems
apple is telling some teams to do exactly that
the skill that matters now is knowing which model a task needs
you're worried ai will take your job
aws just showed what your job turns into
> agents wrote every line of a tool covering 200+ services
> engineers wrote the tests
> agents weren't done until they passed
so stop worrying and get good at writing the checks
broke it down in 90 seconds ๐
aws just shipped a copy of aws that runs on your laptop
200+ services. every line of code written by ai agents
so how do you trust code no human wrote?
> they started from the public docs of how each service behaves
> agents turned those into tests, then ran them against real aws until they passed
> then agents built the emulator until it passed the same tests
> security agents attack every change before it merges
the engineers didn't disappear
they decide what gets built and they build the checks
werner vogels: one rebuild went nowhere for 4 nights without a way to check the agent's work. 2 hours once it had one
an agent is only as good as the check you give it
write the test first