Tommaso Nervegna (@Nervo) header
Tommaso Nervegna (@Nervo) profile picture

🤌 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
Photo by Tommaso Nervegna (@Nervo)Photo by Tommaso Nervegna (@Nervo)Photo by Tommaso Nervegna (@Nervo)Photo by Tommaso Nervegna (@Nervo)Photo by Tommaso Nervegna (@Nervo)Photo by Tommaso Nervegna (@Nervo)Photo by Tommaso Nervegna (@Nervo)Photo by Tommaso Nervegna (@Nervo)Photo by Tommaso Nervegna (@Nervo)Photo by Tommaso Nervegna (@Nervo)
What a day in London with @SpaceXSI 🇬🇧 Shared the stage at Grok Bot Galaxy with the one and only @nickwm. Thanks for having me to talk about how to run an AI-native #Design studio with Grok Bot. Plus, they had @bot cookies. How cool is that? 🍪 #grokbot
2
1
24
Claude's smallest model beat Opus with one text file. 84 real tasks: Haiku + skills 27.7%, Opus without 22.0%. The edge: writing the work down once. via @distortgeekin nitter.gq/distortgeekin/status/2… #AgenticDesign #SortedPixels #AgentSkills
Claude's smallest model just beat Claude Opus with one text file 84 real tasks: Haiku with skills 27.7%, Opus without them 22.0% The difference wasn't the model, it was writing the work down once 6 parts, 6 broken skills and a SKILL.md you can copy ↓
20
Open sourcing your SaaS is the new SaaS playbook. Turn your best workflows into skills, put them on GitHub, make them run in Claude Code, Codex and Cursor on day one. via @gregisenberg nitter.gq/gregisenberg/status/21… #AgenticDesign #SortedPixels #Startups
I’ve noticed open sourcing your saas is the new saas playbook 1. pick the job your customers hire you for 2. turn your best workflows for that job into skills 3. open source them on github 4. make them work in claude code, codex, cursor and openclaw on day one 5. ship an MCP so agents can use your product directly 6. post the repo with a list of every skill inside (like this one) 7. every install puts your brand inside someone's agent, and it gets used every day 8. the free skills do the work once. the paid product does it every week, with your data, hosting and support 9. upsell to done for you, team plans and the hosted version 10. keep shipping new skills. every one is a new launch Your free skills are the new free trial. saas has gone open source for real
39
User interviews are out. Evals are in. When your users are agents, you stop mocking journeys and run evals that feed the build loop. Who writes the evals on your team? via @rauchg nitter.gq/rauchg/status/21224155… #AgenticDesign #SortedPixels #UXResearch
The old way: users are human, you run user interviews, get inspired, mock up user journeys. The new way: users are agents, you run evals, evals inform the software engineering loop. Good read on how we do evals at Vercel.
1
1
22
A rough sketch to a working font in minutes. Gemini 4 Hydrogen builds the typeface and the design tool around it. Type designers: try it on your own glyphs. via @GeminiApp nitter.gq/GeminiApp/status/21223… #AgenticDesign #SortedPixels #Typography
Turn a rough sketch into a working custom font in minutes with Gemini 4 Hydrogen. Using Google @Antigravity in the Gemini app for Mac, Hydrogen handles the technical heavy lifting so you can build custom typefaces and interactive design tools from scratch.

1
53
A paywall skill that bans fake timers, fake reviews and hidden close buttons. It studies rival paywalls, then drafts 3 screens a critic scores until they pass. via @tobiadonadon_ nitter.gq/tobiadonadon_/status/2… #AgenticDesign #SortedPixels #UXDesign
the onboarding sequence is the most important thing to optimize in an app i built a skill that designs your onboarding and the paywall it researches all the main app paywalls/onboardings in the niche → your onboarding, redesigned so every screen has one job → your price, set against what your competitors charge → the doubts in their head when the paywall appears, and what happens if they say no → three directions, as real screens, scored by a fresh critic until they pass → the code: SwiftUI with Apple's purchase system, or Stripe for the web. no fake timers, no invented reviews, no hidden close button giving it away for free comment "PAYWALL" and i'll send it over (need to follow for dm)
28
Your coding agent can now learn Disney's 12 principles of animation. LottieFiles' motion-design-skill: timing and easing before code. 1 of 20 repos for agent motion work. via @0xMovez nitter.gq/0xMovez/status/2121312… #AgenticDesign #SortedPixels #MotionDesign
These 20 GitHub repos will turn your Fable 5.2 into a premier motion studio. this is f*cking gold for Fable 5.2 skills. 3D. FX. Motion. Sound. Export. 01 /motion-design-skill → Timing, easing and Disney's 12 principles. github.com/LottieFiles/motio… 02 ai-kit → CSS springs and motion audits for agents. github.com/motiondivision/ai… 03 Remotion Skills → Official video skills for coding agents. github.com/remotion-dev/skil… 04 /hairline → Interactive line figures from an idea or an SVG. hairline.lucasmarkes.com/ski… 05 /anatomy → Machines and systems as isometric figures. skills.wheresryan.sh/anatomy 06 /anatomy-reveal → Exploded product views and a reveal MP4. typingmind.com/skills/annngu… 07 /blender-agents → A full AI 3D team inside Blender via MCP. github.com/scenario-labs/ble… 08 Three.js → 3D scenes and camera moves. github.com/mrdoob/three.js 09 react-three-rapier → Real physics for React Three Fiber scenes. github.com/pmndrs/react-thre… 10 react-postprocessing → Bloom, depth of field and film effects. github.com/pmndrs/react-post… 11 three.quarks → Particle trails and VFX. github.com/Alchemist0823/thr… 12 Tweakpane → Live controls for your scene parameters. github.com/cocopon/tweakpane 13 GSAP → Animation timelines. Now fully free. github.com/greensock/GSAP 14 Remotion → React video engine. github.com/remotion-dev/remo… 15 Manim → Animated diagrams, equations and explainers. github.com/ManimCommunity/ma… 16 Lottie → Lightweight vector animation playback. github.com/airbnb/lottie-web 17 Openverse API → 5.2M+ free sound effects, ambience and music. api.openverse.org/v1/ 18 librosa → Beat detection for music-synced animation. github.com/librosa/librosa 19 WhisperX → Word-level timing for captions. github.com/m-bain/whisperX 20 FFmpeg → Encoding and delivery. github.com/FFmpeg/FFmpeg Skill → scene → motion → sound → final cut. Bookmark this then go build something worth watching with guide below ↓
2
3
119
A tool "wildly impractical" for a human engineer just shipped. xeet renders an X post to PNG in ~1ms. An oracle checks it against Chrome on 10k+ layout rules; agents get the PRs. via @rauchg nitter.gq/rauchg/status/21221847… #AgenticDesign #SortedPixels #WebDev
Introducing 𝚡𝚎𝚎𝚝: xeet-rs.vercel.app. A hermetic, pixel-perfect X post renderer in safe Rust / wasm. Zero browser, network, filesystem, system fonts. Renders a post to PNG in ~1ms (and 0.4 ms for @jack's "setting up my twttr" 😁). X is the global town square, and the medium is the message. Being able to preview what things will look like before posting is great for both humans and agents. Rendering with xeet is is 100x-1000x faster than booting up a browser and taking a screenshot, yet it's just as accurate. It reproduces Blink & HarfBuzz text shaping, 1/64px line breaking, RTL bidi, 28 Chirp font subsets, CJK/Math/Emoji fallbacks, "Show more" rules, etc. It's the kind of thing that would be wildly impractical, if not impossible, for a human engineer to pull off. Except 𝚡𝚎𝚎𝚝 is structured as an "autorepo." An oracle boots up Chrome, and verifies tens of thousands of text layout rules. A CI canary then watches the evolution of the original, and files signed PRs for agents to act on. It was conceived from the ground up around 'verification engineering', and the result speaks for itself. Enjoy!
1
1
54
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 nitter.gq/steipete/status/212100… #AgenticDesign #SortedPixels #AIAgents
There’s 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 🫠
33
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 nitter.gq/everestchris6/status/2… #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
79
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 nitter.gq/vercel/status/21209500… #AgenticDesign #SortedPixels #DesignSystems
Introducing the UI SDK.
1
100
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 nitter.gq/RoundtableSpace/status… #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
44
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 nitter.gq/RoundtableSpace/status… #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. •⁠ ⁠Written in 21 days, up to 2,943 in a single day •⁠ ⁠12,701 already fixed by other agents •⁠ ⁠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] ⁠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: •⁠ ⁠One cause per section. 3 causes = Cause 1 of 3, 2 of 3, 3 of 3 •⁠ ⁠Fix the code, never the test •⁠ ⁠Nothing is fixed until the "Done when" tests pass
1
1
34
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 nitter.gq/lennysan/status/212057… #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 (−28%) and UX research (−45%) actually shrank. So where’s 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-changin’. Full report: lennysnewsletter.com/p/state…
1
1
93
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 nitter.gq/RoundtableSpace/status… #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: •⁠ ⁠37,966 findings reported by Anthropic •⁠ ⁠7,349 made public •⁠ ⁠217 marked fixed •⁠ ⁠442 official CVEs, about 0.5% of all CVEs this year •⁠ ⁠2 exploited in the wild •⁠ ⁠1,654 marked as duplicates, almost 8x the number fixed •⁠ ⁠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/anthropic…
23
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 nitter.gq/RoundtableSpace/status… #AgenticDesign #SortedPixels #VibeCoding
Unslop is a free skill that makes vibe-coded screens look like a designer worked on them. → it spots AI tells like purple gradients, glow effects, emoji labels, hyped copy, and uneven spacing → it shows you the diagnosis and a design direction before changing anything, then fixes in stages → 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
110
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 nitter.gq/gregisenberg/status/21… #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 can’t see why you’re 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 can’t see the conversations, so nobody knows which chats turn into sales yet 9/ i’m not sure people want to talk to a brand’s agent. i think they’ll want their own agent to talk to it for them 10/ 1.2 billion people use chatgpt every week. i’d go claim my profile before finishing this post if you can
1
55
I Built My Own AI Phone with an Apple Watch Ultra and a 3D Printer nervegna.substack.com/p/i-bu…
🤖 Made with AI
30
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 nitter.gq/RoundtableSpace/status… #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: •⁠ ⁠Claude's file tools can't read or write outside your project, even in bypass mode •⁠ ⁠Your .env files are off-limits everywhere •⁠ ⁠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.
19
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 nitter.gq/RoundtableSpace/status… #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: •⁠ ⁠Up to 120fps and 4K, straight to MP4 •⁠ ⁠Records the tab's audio too •⁠ ⁠Chrome encodes while it records, so stopping is instant •⁠ ⁠Start and stop whenever you want: log in first, then record just the part you need •⁠ ⁠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-…
28