Antonin 🌞

@DilumSanjaya Really cool! Why Gemini and not codex or opus for coding?

Been thinking about sharing some fun, interactive science app ideas Made this one today UI design and planet textures GPT Images 2 Code Gemini 3.1 Pro

@mtwit75 @DavidCarralon Je comprends pas pourquoi le gouvernement laisse mourrir des secteurs entiers utiles au lieu d’utiliser l’impôt ou autre incentive pour aider

À Paris, les commerces qui cartonnent le plus sont désormais : la restauration rapide assise (+579), les ongleries (+217), la téléphonie discount (+201), les friperies (+99), tandis que le prêt à porter femme (-678), les agences bancaires (-205), les agences de voyages (-179) et les restaurants traditionnels français (-139) mettent la clé sous la porte. En seulement 5 ans.

@pbakaus Something I’m still struggling is to get all the spacing/padding right on desktop + mobile. I wish it could a analyse all pages and fix it all in 1 prompt

What are your favorite Impeccable commands? And if you had three wishes for Impeccable / AI design, what would they be?

@umzrs image

You can now create brand kits just like this in a matter of seconds. I put together a playbook on how to make it. like + comment "image" And I'll send it over

@jordan_ross_8F Ai

Most ad and SEO agencies will spend the next twelve months pointing AI at the wrong things. They'll automate the work that doesn't move margin. Ignore the work that does & wonder why the P&L looks identical to last year's. I made a strategic guide naming the top five things you should and can build— the ones that actually compound into profit, month after month. Take my PDF, upload it to any LLM and it will spit out the exact strategy you need to follow. Execute all five and you'll add seven figures to the bottom line. Comment "AI" and I'll send it over. (Must be following so the auto-DM lands.)

@jordan_ross_8F OS

I fully reverse-engineered Ramp's internal AI operating system for marketing agencies. Their system — called Glass — is how they got 99% of their entire company using AI every single day. 350+ reusable workflows. Every tool connected at first login. Memory that refreshes every 24 hours. Automations running while everyone sleeps. I partnered with my engineering team and we broke down every component inside it. Then we rebuilt the whole thing for marketing agencies. 76 pages. Every system. Every layer. Every step. Steal it. Comment "OS" and I'll send it directly. Must be a following to receive auto DM

@pbakaus awesome, looking forward :)

@antonincohen You’ll love Impeccable 3 then 👀

Impeccable 2 (my take on the most effective way to make your AI harness better at design) dropped last week. But I was so busy working on the next iteration that I buried the lede a bit. So here's a detailed post on what's new and improved: 1. CLI: npx impeccable detect. Scans your code *without need for an LLM* for 25 anti-patterns across typography, color, layout, motion, and quality. Works on HTML, CSS, JSX/TSX, Vue, Svelte, and CSS-in-JS. Framework detection, multi-file import tracking, Puppeteer-backed live URL scanning, CI-ready JSON output,--fast regex mode for huge codebases. Yes, you can built automatic PR anti-slop detection with this in a few lines of code! 2. Chrome extension. One-click detection on any page (yours, staging, production, or someone else's). DevTools panel. Live computed styles, interactive panel, on-page highlights. 3. Dramatically upgraded /critique. LLM + real time detection sub-agents run in parallel, score against Nielsen's heuristics, auto-run the detector, and open a live browser overlay so you can walk each finding in place. 4. Data-driven design performance improvements, driven by an internal eval framework that runs the same brief through frontier models with and without the skill loaded, then measures how much the output collapses into monoculture. The biggest unlock: an anti-attractor procedure that forces the model to enumerate and reject its reflex defaults before picking. Validated on gpt-5.4 and Qwen 3.6 Plus across 15 niches. Result: much more font and color diversity, sharper quality, stronger Codex support. 5. Design before code. /shape runs a structured discovery interview and produces a design brief before any code is written. /impeccable craft chains that brief into the full implementation flow. You ship a designed feature instead of a reflex card grid. 6. One name, one namespace, and 18 commands (down from 21). The skill used to be called frontend-design, it's now called impeccable. /teach-impeccable became /impeccable teach, /extract became /impeccable extract. /arrange became /layout. /normalize merged into /polish. /onboard merged into /harden. Auto-cleanup handles leftover files on first load after updating. 7. A real docs site. 18 per-skill pages with before/after demos, canonical SKILL.md inline, two tutorials, 38 rule cards with inline visual examples. Give it a try: impeccable.style

Before emailing, I tried in person. Guy said he'd check stock and call me back. 7 days and 3 follow-up calls later, still "I'll check." Never heard back. 100+ Google reviews. 4.8 average. Make it make sense.

This is Spain. I emailed 6 bike e-shops asking if they had a specific part in stock (from a brand they all carry). Three days later: zero replies. The level of service is unbelievable. Btw I picked 6 because I knew half wouldn't reply. Turns out none of them did lol

@AbdelStark Would love to hear about this from someone @MistralAI :D

@antonincohen Ah very good question. I did not test aside from me speaking, it worked very well but not sure I have a super French strong accent. Would be cool to perform some tests and benchmarks for this

I just shipped "parler" (French: *to speak*): Multilingual voice intelligence built on Mistral Voxtral model, generating decision logs from French/English meetings. Cocorico a bit 🇫🇷, but unironically: this is the kind of French AI stack I want to see being built, deployed on the industrial level, and win more. Sovereign AI is such an important matter, not only for us all as individuals, but also strategically for Nation States, it will become more and more a major geopolitical issue. Need more counter balance to the US frontier labs. We cannot let intelligence and knowledge be controlled by a handful of players, especially not concentrated in few countries. It leverages 2 pure French models from @MistralAI: - voxtral-small-latest for multilingual speech-to-text with timestamps - mistral-medium-latest for structured JSON extraction of decisions, commitments, open questions, and rejected options github.com/AbdelStark/par…

@scaling_shields LEADS

i'm gonna regret leaking this but fuck it 15 lead sources that ARENT apollo took me 4 hours to put together 40 pages of where i actually find leads - tweet scraper (40K emails → 355 opportunities from ONE source) - the shopify app review trick - how to find leads who just raised money - the google maps play for local - sources for ecom, agencies, SaaS, local after sending 1,000,000+ emails these are the sources that ACTUALLY work like + comment "LEADS" and i'll send it over (must follow + RT for priority access)

@nb4ld Merci, à faire sauter aussi l'obligation d'un avocat dans certaines procédures. Au tribunal de commerce j'ai du payer un avocat pour qu'il plaide mon dossier que j'ai créé avec l'IA lol

My skill is too specific to my setup, but here's a generic skill prompt — paste this into Claude Code and adapt the to your setup and don't forget to install mcp and related skills: Create a Claude Code skill called "create-site" that automates end-to-end website creation from a text description. SKILL.md file (user-invocable, argument-hint: "[project-name] [description] [--resume] [--step=N]") with a references/ subfolder for templates. Pipeline (7 steps): 0. Initialize: Parse project name + description. Write site.json to track state (enables --resume and --step=N). 1. Research: Auto-decide if topic needs research (health/science/travel = yes, portfolio/landing = no). WebSearch + WebFetch 5-10 sources → research.md. 2. Design Direction: From the topic, determine audience, feeling, physical metaphor ("a leather-bound travel diary"). Translate into sensory material descriptors for Stitch — "architectural limestone, burnished copper" works far better than "modern, clean" because image gen models anchor on concrete materials. Color hierarchy: neutral 80-90% (canvas), primary (ink), secondary (subdued), accent (loudest, used least). Create Stitch project, write DESIGN.md. 3. Content: Small sites (1-5 pages) → build semantic HTML5 directly. Large sites (6+) → generate prd.json, run Ralph loop (fresh Claude instances in bash, each reads prd.json + progress.txt, implements one story, commits, exits, stop on COMPLETE). 4. Stitch Visuals: Every page gets mcp__stitch__generate_screen_from_text (DESKTOP, GEMINI_3_1_PRO). Prompt includes design direction + physical metaphor, color palette with material descriptions, actual page content (stripped HTML), IMAGE RULES (no text in images, harmonize with palette), CONTENT INTEGRITY (only provided text, no fake stats/testimonials). 5. Build: Download Stitch HTML, merge visual framework with content HTML. Consistent nav/footer, mobile toggle, scroll animations. 6. Deploy: [YOUR DEPLOY SETUP — DNS + hosting config + verify] 7. QA: Headless browser screenshots at desktop + mobile, check console errors, verify links. Fix → commit → re-test, max 3 iterations. Key: separate content from design (prevents hallucinated content), site.json as state machine, Stitch prompt template in references/stitch-prompt.md, append-only progress.txt for Ralph cross-instance memory. Step 2 is what makes or breaks it — the material descriptors in the design direction are what make Stitch output look designed vs generic AI slop. Step 6 you'll need to adapt to however you deploy.

@antonincohen @AnthropicAI @GeoffreyHuntley @ryancarson No, not yet. Can you share the skill you created?

50+ tools exist to make @AnthropicAI's Claude Code build websites autonomously. I tried most of them. None of them solve the actual problem.   The autonomous loop space is wild. @GeoffreyHuntley invented the Ralph Wiggum technique. @ryancarson turned it into a 14k-star repo. @garrytan's gstack hit 61k stars with screenshot-based design review that iterates until the page is clean. @pbakaus built Impeccable, which scores your UI against Nielsen heuristics. There are 14+ loop tools and 5+ design quality skills, all fighting over who makes Claude write better CSS.   But every single one of them outputs a page with no images. Clean layouts, solid typography, intentional animations, and empty boxes where visuals should be. That's why AI-coded sites look like wireframes. That's the gap @lovable_dev solved by controlling the full stack.   @stitchbygoogle changed this for Claude Code. It generates complete UI screens with real photography, illustrations, hero images from a text prompt. 5 variants at once. 350 free generations per month. Powered by Gemini.   The MCP server connects the two worlds. Claude Code plans the architecture, writes the content, sends each screen to Stitch as a text prompt. Stitch generates the full visual design. Claude pulls it back, builds production HTML/CSS to match, and with @Cloudflare's API wired in, deploys to a subdomain.   One voice message. A site with real images that looks like it came out of @Lovable

@OpenAI @grok how much total raised year by year?

Today, we closed our latest funding round with $122 billion in committed capital at an $852B post-money valuation. The fastest way to expand AI’s benefits is to put useful intelligence in people’s hands early and let access compound globally. This funding gives us resources to lead at scale. openai.com/index/accelera…

@ravilutionx @AnthropicAI @GeoffreyHuntley @ryancarson Did you try? It's very easy, I built a skill that sticks everything together (including optimized prompt for stitch)

@antonincohen @AnthropicAI @GeoffreyHuntley @ryancarson can you share any success stories of making ralph-loops work stably with Stitch integration? how would you define stopping criteria here?

I moved to Florida at 23 and also got stopped for a bit of speeding by cops approaching me with hands on guns. I had 3 police cars surrounding me. And no ticket given 🤣