
š¤ I Build Agentic AI experiences for Fortune 500s @AccentureSong. Father of 3, 1 unimpressed dog, 5 Mac Studios running frontier AI in my wife's closet.
Milan - Italy
Joined March 2009
- Tweets4.3K
- Following1.4K
- Followers1.6K
- Likes1.1K
A new agent launches this week, and marketers are paying for nice things said about it on X.
My rule: no demo on a real task, no opinion.
via @steipete
teapaw.org/steipete/status/212100ā¦
#AgenticDesign #SortedPixels #AIAgents
Thereās some new agent launching later this week and marketers offer a lot of money to say nice things about it on X. Keep a lookout on honest reviews š«
22
Floor plan in. 3D walkthrough out.
Fable 5.2 + three.js, no external assets. Developers pay $30k-100k for these.
The trick: it grades itself against the official renders.
via @everestchris6
teapaw.org/everestchris6/status/2ā¦
#AgenticDesign #SortedPixels #3D
claude fable 5.2 + three.js is f*cking crazy
it can make full 3d walkthroughs of real estate projects from a floor plan
real estate companies usually pay $30k to $100k for these because they sell apartments while the building is still under construction
here's the full setup for this biz model:
> use the apify google maps scraper or the google places api to pull real estate companies in your city
> crawl their sites and keep the ones selling a building that isn't finished
> download the floor plans and renders from their sales page, every project usually publishes them
> render a few rooms of their own plan with gpt image or a short walkthrough clip with seedance as the sample
> send the sample to the sales office by email or a postcard with a link
> once they pay, give fable the plans and the renders, the renders are what it judges its own work against
> fable builds the whole project in blender and three.js, with no external assets
and you send them the walkthrough of their own building.
1
62
Vercel shipped a UI library you're not meant to write. Your agents are.
UI SDK: behaviors agents use to build interfaces. 3.5k bookmarks overnight.
via @vercel @steipete
teapaw.org/vercel/status/21209500ā¦
#AgenticDesign #SortedPixels #DesignSystems
1
86
Pay top-model prices only for the thinking.
shadcn's /improve: your best model audits and plans, cheap models build, the best one reviews.
via @RoundtableSpace @shadcn
teapaw.org/RoundtableSpace/statusā¦
#AgenticDesign #SortedPixels #ClaudeCode
The creator of shadcn/ui uses his smartest AI model only to plan, and lets cheap models do the coding.
His skill is called /improve:
- Your best model audits the whole codebase
- Ranks what's worth fixing: bugs, speed, security, tests
- Writes a step-by-step plan for each fix
- Hands each plan to a cheaper model to build
- Reviews the cheap model's work before you merge
You pay top-model prices only for the thinking.
Set it up:
1. npx skills add shadcn/improve
2. Open your coding agent in your project
3. Type /improve (or /improve quick for a cheaper pass)
4. Reply "plan 1, 3 and 5", then run /improve execute 001
Works with any agent that supports skills.
github.com/shadcn/improve
34
AWS Emulator: 19,244 bug reports. 5 by a human. 12,701 fixed by other agents.
The trick: one strict format that defines "done" before anyone fixes.
via @RoundtableSpace
teapaw.org/RoundtableSpace/statusā¦
#AgenticDesign #SortedPixels #AIAgents
Amazon published every bug report its AI agents wrote to each other while building AWS Emulator.
I went through all 19,244 of them.
ā¢ā ā Written in 21 days, up to 2,943 in a single day
ā¢ā ā 12,701 already fixed by other agents
ā¢ā ā Only 5 were written by a human
Every single report follows the same format. That's how hundreds of agents fixed bugs with no human checking their work: the agent fixing a bug never has to guess what "fixed" means.
Here's the format. Paste it into your AGENTS.md:
When you find a bug, file it in this exact format:
# [What breaks, in one line]
ā Found at commit [hash]. The code may have changed since. Trace it to the current code before fixing.
Severity: CRITICAL / HIGH / MEDIUM / LOW
Where: [the exact function, endpoint or command]
## Description
Cause 1 of [N]: [one sentence]
Expected: [what correct behavior is, and where that's defined]
Actual: [what happens instead, with file and line]
Why it matters: [who breaks, and what they see]
Fix: [which part of the code the fix belongs in, not the patch itself]
## Evidence
[the failing test name and the exact error output]
## Steps to reproduce
1.
2.
3.
## Done when
[the exact tests that must pass]
Rules:
ā¢ā ā One cause per section. 3 causes = Cause 1 of 3, 2 of 3, 3 of 3
ā¢ā ā Fix the code, never the test
ā¢ā ā Nothing is fixed until the "Done when" tests pass
1
1
30
Product designer openings: -28%, a record low. UX research: -45%.
Meanwhile "Claude" just passed CSS as the #2 skill in designer job posts.
Fewer seats. Different toolbox.
via @lennysan
teapaw.org/lennysan/status/212057ā¦
#AgenticDesign #SortedPixels #DesignJobs
The tech job market is in its best shape in years. But almost none of the growth is coming from software jobs.
There are 281,000+ open roles at tech companies, up 17% from a year ago and 73% from the March 2023 low.
But open PM, eng, and design roles grew just 2%. Design (ā28%) and UX research (ā45%) actually shrank.
So whereās the growth coming from? Hardware engineering openings doubled. Business ops added ~4,200 openings. Supply chain is up 63%. Many of the fastest-growing jobs are physical: data center technicians, installers, field service engineers.
The AI boom is creating jobs, but mostly in building and running the physical infrastructure (chips, data centers, hardware), not building software.
Times, they are a-changinā.
Full report: lennysnewsletter.com/p/stateā¦
Anthropic's AI flagged 37,966 possible bugs in open source. Its own ledger marks 217 fixed.
Finding is cheap now. Fixing is still a human with a calendar.
via @RoundtableSpace
teapaw.org/RoundtableSpace/statusā¦
#AgenticDesign #SortedPixels #AISecurity
Anthropic said its AI found thousands of critical bugs in open source. Seven months later, someone counted.
VulnCheck has tracked Anthropic's Project Glasswing since launch:
ā¢ā ā 37,966 findings reported by Anthropic
ā¢ā ā 7,349 made public
ā¢ā ā 217 marked fixed
ā¢ā ā 442 official CVEs, about 0.5% of all CVEs this year
ā¢ā ā 2 exploited in the wild
ā¢ā ā 1,654 marked as duplicates, almost 8x the number fixed
ā¢ā ā curl rejected 19 of the 34 reports it got
Since September: nearly 12,000 new findings, 15 new fixes.
Finding bugs is now cheap. Fixing them still takes humans.
vulncheck.com/blog/anthropicā¦
20
Purple gradients, glow, emoji labels: your vibe-coded UI is showing.
Unslop, a free skill, flags the slop and pitches a direction before editing.
via @RoundtableSpace @heydamir
teapaw.org/RoundtableSpace/statusā¦
#AgenticDesign #SortedPixels #VibeCoding
Unslop is a free skill that makes vibe-coded screens look like a designer worked on them.
ā it spots AI tells like purple gradients, glow effects, emoji labels, hyped copy, and uneven spacing
ā it shows you the diagnosis and a design direction before changing anything, then fixes in stages
ā works with Claude Code, Codex, and Google Antigravity, on React, Vue, plain HTML, or even a live site
If a screen already looks good, it only suggests small tweaks. Free for commercial use with credit (CC BY 4.0).
Repo: github.com/heydamir/unslop
1
83
Agencies made millions claiming Google listings for local shops.
Now any business can claim its profile inside ChatGPT. Same playbook, new front door.
via @gregisenberg
teapaw.org/gregisenberg/status/21ā¦
#AgenticDesign #SortedPixels #AISearch
openai just let every business claim a profile inside chatgpt. my honest reactions:
1/ this could be the google my business moment for ai, and a lot of agencies made a great living (millions!) in the 2010s just claiming and fixing listings for local businesses
2/ ads can now click through to a conversation with your agent instead of your website (interesting)
3/ every business agent runs on context the company writes, and most brands are going to be bad at it
4/ your product feed is your new seo. boring, and it matters more than your blog now in some ways if you think about it
5/ the dashboard tells you how often chatgpt mentions you but HIDES the prompts and your competitors, so you still canāt see why youāre winning or losing
6/ discovery topics shows what chatgpt already cites you for, which is basically a free content calendar
7/ agents are only open to us ecommerce and online services right now. plumbers, dentists etc are next, and so lots of opportunity to help main street here
8/ businesses canāt see the conversations, so nobody knows which chats turn into sales yet
9/ iām not sure people want to talk to a brandās agent. i think theyāll want their own agent to talk to it for them
10/ 1.2 billion people use chatgpt every week. iād go claim my profile before finishing this post if you can
1
53
I Built My Own AI Phone with an Apple Watch Ultra and a 3D Printer nervegna.substack.com/p/i-buā¦
š¤ Made with AI
28
Claude Code can read your SSH keys and every .env you own.
Two settings fence its file tools into your project, even in bypass mode. Shell isn't covered.
via @RoundtableSpace
teapaw.org/RoundtableSpace/statusā¦
#AgenticDesign #SortedPixels #AISecurity
By default, Claude Code can read files anywhere on your computer. Your SSH keys, your .env files, your other projects.
Two new settings fence it into the folder you're working in. Add this to ~/.claude/settings.json:
{
"permissions": {
"blockReadsOutsideWorkingDirectories": true,
"blockWritesOutsideWorkingDirectories": true,
"deny": [
"Read(//**/.env)",
"Read(~/.ssh/**)",
"Read(~/.aws/**)"
]
}
}
What it does:
ā¢ā ā Claude's file tools can't read or write outside your project, even in bypass mode
ā¢ā ā Your .env files are off-limits everywhere
ā¢ā ā Your SSH and AWS keys are off-limits
Need another folder? Type /add-dir in Claude Code.
Update Claude Code first. The write block needs version 2.1.314 or newer.
18
Your product demos just went 120fps, free.
Paul Irish's chrome-capture: one npx command records a Chrome tab to MP4, up to 4K, with audio. No ffmpeg.
via @RoundtableSpace
teapaw.org/RoundtableSpace/statusā¦
#AgenticDesign #SortedPixels #Prototyping
Chrome can now record any tab to a smooth 120fps video by itself. No screen recorder, no ffmpeg.
Paul Irish from the Chrome team just released chrome-capture, a one-command way to use it:
ā¢ā ā Up to 120fps and 4K, straight to MP4
ā¢ā ā Records the tab's audio too
ā¢ā ā Chrome encodes while it records, so stopping is instant
ā¢ā ā Start and stop whenever you want: log in first, then record just the part you need
ā¢ā ā Works with Playwright and Puppeteer, so your AI agent can record its own test runs
Try it (needs Node and Chrome 153+):
npx chrome-capture paulirish.github.io/chrome-c⦠demo.mp4 4s
For 120fps at 1080p on a 120Hz screen like a MacBook Pro:
npx chrome-capture <url> out.mp4 5s --fps=120 --width=1920 --height=1080 --headful
github.com/paulirish/chrome-ā¦
25
Claude Fable 5.2 is #1 on @arena WebDev (1832, xHigh) and @ArtificialAnlys 3D (1271 Elo, Max).
Anthropic still calls Opus 5.5 the daily driver.
A #1 at max effort is a specialist, not your default. Read the effort tag before the headline.
anthropic.com/claude-fable-aā¦
29
A year ago the best AI agent did 2.5% of real freelance projects to client standard. Fable 5.2: 41.3%.
Agencies billing by the hour, read that twice.
via @gregisenberg @ScaleAILabs
teapaw.org/gregisenberg/status/21ā¦
#AgenticDesign #SortedPixels #Agencies
quick observations from the claude fable and mythos 5.2 launch:
1/ it went from doing 21% of real freelance work to 41% in one release, so agencies are about to start charging per deliverable instead of per hour!
2/ it can work 10-12 hours on its own, which basically gives every small business a night shift (might as well use it if you know what you want to build)
3/ over 90% of anthropic's own code is now written by claude, and so i think the solo $1b founder is coming soon
4/ one engineer ran 1000+ subagents for 8 days and got an operating system that boots, so the ancient software running hospitals and airlines is finally rebuildable (lots of opportunity here on rebuilding)
5/ it designed 3 working electronic devices from a single prompt, which makes 1 person hardware companies a real thing
6/ it now says "i don't know" instead of guessing, and that's what law, medicine and finance have been waiting for (this has been the biggest thorn for these industries, hard to trust guessing)
7/ 1000 agents formalized fermat's last theorem in 30 hours, so software that's mathematically proven not to break is coming for payments and medical devices
8/ mythos is only for verified orgs, which means someone is going to build a business selling what only they have access to
9/ the models are getting better at noticing when they're being tested, so testing ai out in the real world is about to become its own industry
i love new model day
31
If an agent can clone your app's logic in minutes, the code was never the moat.
694k views on REA vs a free counter-repo. Taste and distribution still don't clone.
via @0xCodila
teapaw.org/0xCodila/status/211914ā¦
#AgenticDesign #SortedPixels #Moat
REA (reverse engineer anything) - F*CKED by chinese student...
solo dev just dropped protection for your app & site & game from reverse-engineering and AI hackers, for free
REA not dangerous anymore.
here's how to use and setup BLA (Blind LLM Agents):
first understand how it works:
⢠REA opens your app in 1 command, BLA hides traps inside it before you ship
⢠fake keys and decoy code for AI to chase, plus files built to blow past REA's own limits
⢠AI burns tokens on dead ends, your users get the exact same app
here's how to set it up for FREE:
1 ā open Claude Code in the folder of your app, site or game
2 ā say: " install Node.js 24 if I don't have it, build my project, then run npx bla-agents setup and npx bla-agents harden"
3 ā it finds your build itself: dist, build, Electron .asar, .wasm, Python, native binaries
4 ā it plants fake keys and decoy code, then checks every protected file still runs the same
5 ā the step everyone skips: ship the files from the new .bla-hardened folder, not your old build
6 ā for sites it protects the JS your site sends to browsers, your server code never leaves your server anyway
7 ā want proof? run: npx bla-agents benchmark - before vs after on 14 test apps
paste link on repo below as soon as possible:
36
Meta's design system behind 13,000+ apps now talks to agents.
Astryx rules: start from a template, no raw divs, tokens not magic values. That's how AI UI stops looking AI.
via @0xMovez
teapaw.org/0xMovez/status/2119497ā¦
#AgenticDesign #SortedPixels #DesignSystems
this is wild...
Meta just open-sourced the design system behind 13,000+ of its apps, adapted for AI agents.
your AI agent can finally build UI that doesnāt look like AI built it
hereās how to plug Astryx into Claude and GrokBot agents:
first understand what you get:
⢠150+ accessible React components: buttons, forms, dialogs, tables, dark mode included
⢠7 ready-made themes: neutral, butter, chocolate, matcha, stone, gothic, y2k
⢠a CLI + MCP server built for AI, so your agent reads the real docs instead of inventing props
hereās how to set it up for Claude & Grok:
1 ā open Claude Code in the folder of your project (React 19+)
2 ā say: āInstall @astryxdesign/cli and run astryx init --features agents to set up your Astryx context. Read the generated file.ā
3 ā it writes the rules straight into your CLAUDE.md: component index, CLI reference, and what never to do (no raw divs, no inline styles, no magic values)
4 ā now every page starts the same way: astryx build ā<your idea>ā picks a template ā astryx template scaffolds it ā astryx component reads the props
5 ā want Claude to query the docs live? add the MCP server to .mcp.json: astryx.atmeta.com/mcp
6 ā for Grok or any other agent: plug the same MCP link into any client that supports remote MCP, or run init with --agent codex and it writes a universal AGENTS.md
7 ā quick test: ask your AI āwhatās the import path for Button and how do you make a Dialog non-dismissible?ā If it canāt answer, the setup didnāt load
paste link on repo below ā
2
36
Apple never announced it. One marketplace.json line did.
Xcode now plugs into Claude Code. Greg's iOS playbook starts at revenue charts, not code.
via @gregisenberg
teapaw.org/gregisenberg/status/21ā¦
#AgenticDesign #SortedPixels #iOS
now that apple put xcode inside claude code, hereās how iād find, build and grow an iOS app with AI agents:
find it
- open appfigures or sensor tower and sort utilities by revenue
- filter for apps with a rating under 3.5
- bonus points if the top app hasn't shipped an update in 6+ months
- check the keywords they rank for in appfigures. low difficulty + high revenue is the sweet spot
- scroll the "you might also like" row on the top app's page. every app there is another market to check
- or find a startup idea on ideabrowser.com based on social sentiment
validate it
- use facundoolano/app-store-scraper to pull every review for the top 20
- give them to an agent: "cluster the 1 and 2 star reviews into the top 5 complaints and quote the users"
- pick the app where the complaints are about the product, not the price
- if people complain about the price but keep paying, that's demand. if they complain about bugs, that's your opening
- search reddit for "alternative to [app]". every thread is a list of people waiting for you
build it
- claude code + apple's new xcode plugin to build, preview and fix
- cameroncooke/XcodeBuildMCP so the agent can build, test and run the simulator on its own
- joshuayoes/ios-simulator-mcp so it can tap, swipe and screenshot like a user
- write a skill with the top 5 complaints so every feature gets checked against them
- apple's foundation models for on-device ai, so your inference cost is $0
- app intents so siri and personal agents can use your app too
- add an "import from [competitor]" button. make switching take 10 seconds
test it
- maestro tests every flow the reviewers complained about
- the agent runs them on every screen size before each build
- if the competitor's worst flow takes 6 taps, yours takes 2
- sentry for crashes, telemetrydeck for analytics
- testflight it to 20 people from those reddit threads before launch
monetize it
- revenuecat for subscriptions and mrr
- superwall to test paywalls without shipping a new build
- paywall right after the first win, never on launch
- annual as the default with a 3-day trial, weekly as the anchor
- turn on win-back offers for everyone who cancels
rank it
- your screenshots answer each complaint directly. "no ads." "works offline." "cancel anytime."
- fastlane snapshot + frameit regenerates every screenshot automatically
- a/b test them with app store connect's product page optimization
- have the agent localize your listing into 10 languages. most competitors only rank in english
- ask for a review right after a win with apple's built-in review prompt, never after an error
grow it
- find 10 creators in your niche on tiktok one. small monthly base + a bonus for every 100k views
- give each one a fresh tiktok account that only posts about your niche
- apify's tiktok scraper pulls the top 100 hooks in your niche every week. that's your hook bank
- every creator tests 3 new hooks a day in capcut. anything that loses people in the first 3 seconds gets cut
- arcads or heygen for ai ugc when you need 50 variations of a winner
- postiz to schedule everything across tiktok, reels and shorts
- faceless slideshow accounts posting 3x a day: problem, screenshot, result
- when an organic video takes off, put money behind it with tiktok spark ads instead of making new ads
- a unique apple offer code for every creator so you know exactly who drove each subscriber
- a quiz landing page on framer that ends in "download the app." web-to-app funnels convert better than cold traffic
- appsflyer to track which video, creator and ad actually turned into revenue
- apple ads on the competitor's name
i could do a whole episode on @startupideaspod if people are interested. i will never gatekeep anything im learning in real-time.
welcome to the golden age of ios apps, because now anyone can build one
45
Best agent on Co-Design Bench meets 69.9% of the hidden criteria. It fully finishes 26.4% of the tasks.
273 real user change requests. Claude Code harness. GPT-6 Astra on top.
"Mostly done" is the new "looks done". Review is still the job.
arxiv.org/abs/2610.35138
2
2
69
Introducing Claude Fable 5.2 and Claude Mythos 5.2, designed to tackle the most daunting technical challenges.
Theyāre capable collaborators on ambitious projects in many fields, and show improvements in areas like coding, reasoning, and vision.
Apiiro reports Claude Code's "don't ask again" saved secrets in local settings that ended up public.
A convenience button also decided what to keep.
Permission UX is data-retention UX.
apiiro.com/blog/coding-agentā¦
#AgenticDesign #SortedPixels #AgentUX
16













