Thibaut

Notion Developer Platform is live and it is actually a big deal. What this unlocks: 👉 Agent reads brief, writes back the draft 👉 Deal closes in CRM, Notion project spins up automatically 👉 Sync any data source, agents query it like a database 👉 Workers run code on their infra, no external hosting 👉 External agents can read, write, and act inside Notion Notion is becoming the workspace the whole agent stack can talk to. That is worth getting excited about. One request though: please make it faster. The app still loads like it is apologizing for something.

You can now create more with Claude Design. We've doubled token limits across every plan.

I spent 6 hours reading the open-sourced X algorithm. Then I made myself the cheatsheet I wish I had before reading it. The weirdest takeaway: It does not really reward "good content." It rewards posts that make the right person do several small things in a row: Stop. Read. Expand. Reply. Click the profile. Send it to someone. Come back and read the replies. The model is basically asking: "Based on this person's history, how likely are they to take a positive action on this post?" Then it stacks those probabilities. That means the boring advice is still true, but for a more precise reason: - a strong hook buys the first stop - a useful idea buys dwell time - a specific claim buys replies - a clear point of view buys reposts - consistency teaches the model who should see you So I turned it into a simple creator checklist: 1. Be original enough that it does not feel recycled. 2. Write for one specific reader. 3. Make the first line earn the stop. 4. Give people something worth saving, arguing with, or sending. 5. Avoid the stuff that creates negative signals: spam, ragebait, vague controversy, fake authority. The part most people will probably get wrong is trying to optimize for the algorithm directly. That usually turns into engagement bait, vague controversy, and recycled threads. My takeaway is almost the opposite: Write for a very specific reader so clearly that the model can recognize the match. Cheatsheet attached if you want the condensed version.

@SaimmCRO What are you using to setup AI Product Recommendation Quizzes ?

We've seen insane CVR lifts using AI product recommendation quizzes. It's simple to set up and so freaking powerful. Shoppers land on your site and have no idea where to start. Too many products. Too many choices. Decision paralysis kills the sale before it even begins. The fix: A quiz that asks 1-2 questions about their goals and recommends the perfect product. We place it everywhere: → Homepage hero → Mobile menu → PDP ("Not sure if this is right for you?") Higher CVR. Faster time to purchase. Less bouncing around the site. Why? Because you stopped making customers do the hard work. Huel does this perfectly. "Which Huel is right for you?" sits right in the hero.

I built a small agent for my own X account. It pulls my recent posts, replies, content plan, and the newly open-sourced X ranking logic. Then it suggests what I should post next, or which thread is worth replying to, based on the signals X seems to care about: - reply probability - dwell - originality - niche relevance - negative feedback risk - early engagement potential The wild part is not "AI writes tweets." That's the boring version. The interesting version is: your distribution system can become a feedback loop. Content -> engagement signals -> learned taste -> next post/reply -> better audience fit. Like + Repost to receive it (must follow)

🚨 Mistral just confirmed a supply chain attack. 450 internal repos for sale at $25k on a criminal forum. One week before they reportedly go free. How it happened: hackers didn't touch Mistral directly. They compromised TanStack, a tool Mistral devs were using. From there, infected npm and PyPI packages were published between May 11 22:45 and May 12 01:53 UTC. ~5 GB of code. Model training, fine-tuning, finance agents, KYC, dashboards, future projects. Mistral's advisory (MAI-2026-002) confirms an infected developer device. They say no infrastructure was compromised. That may be accurate. But the code is already out. Priced. Countdown running. ⏳ The part worth sitting with: this wasn't a brute force attack on Mistral's servers. It was a dependency. One infected open-source tool opened the door to 5 GB of IP. Supply chain attacks keep getting cheaper to run and more expensive to absorb. Mistral probably isn't the last AI company this happens to.

@ShopifyDevs imagine spending last week building: → custom metering → billing periods → usage-based extra charges on top of subscriptions → idempotency keys so you don't double-charge on retries → reconciliation for when charges drift just imagine...

Introducing App Events Send custom and billable events through one API and see all your app activity in one place

Shopify App Pricing launched and it's a bigger deal than it looks 👀 Before this, if you wanted usage-based billing on your Shopify app, you were building: → your own plan pages → your own metering and aggregation → your own invoicing → your own refund/correction logic when a job retried or a charge was wrong Now you configure a meter in Partner Dashboard, send events via the App Events API, and Shopify handles the rest. Aggregation, invoicing, charge approvals, corrections — all platform. The correction part is kinda wild. If a message bounced or a transaction reversed, Shopify auto-adjusts the charge. No refund processing. No support ticket. Just negative event reporting and it's handled. That's a lot of stuff that used to be your problem.

OpenAI just raised $4B for a deployment company. Not for a new model. Not for compute. For forward deployed engineers who go into businesses and redesign workflows around AI. That's a striking signal about where the value actually lives now. The model is table stakes. The hard part is figuring out where it fits, what it needs to know, and how to make the gains stick. OpenAI is betting that takes humans embedded in organizations — not just API access. For ecommerce: the same bottleneck exists. Agents can query your Shopify store today. But they're working with behavior data — what got clicked, what got added, what got abandoned. That's not the same as understanding what the shopper was actually trying to solve. Deployment gets you to the door. Customer intent is what the agent needs to do something useful on the other side.

🚨 How to Avoid Supply chain attack - Shai-Hulud echo -e "minimum-release-age=4320\nmin-release-age=3" >> ~/.npmrc This oneliner setting forces npm and pnpm to reject any package version published less than 3 days ago.

@nicklaunches Let’s go

If you're building with AI, let's connect

@tytustytus Spot on. More people = exponential communication tax. That’s why small teams ship 10x faster. Well done.

Building products has fundamentally changed. Most founders haven't realized it yet. 3 months ago at Tidio (160-person company), we pulled 4 people into a small team and gave them one job: build our next product from scratch. Here's what they shipped: → 1 new product (announcing soon) → 486,437 lines of code I wrote 56,309 of them myself.

@ohad too few people understood that! customer feedback is crucial to build something people actually want

Your product update is the only time users expect to hear from you. Most founders waste it on a release note. Split the update into two ships. Two emails. Two chances to ask: "what's broken for you?" That question is worth more than any release.

being out of token is one of the worst thing that can happen to you

🚨 Supply-chain attack hit TanStack today. 84 versions across 42 @tanstack npm packages were compromised between 19:20–19:26 UTC on May 11 as part of the "Mini Shai-Hulud" campaign. The attack vector: a git-resolved optionalDependency with an obfuscated router_init.js prepare script that silently exfiltrates: AWS, GCP, Kubernetes & Vault credentials, GitHub tokens, .npmrc files, SSH keys Basically anything useful in a CI or dev environment. If you use TanStack packages, act now: ✅ Rotate all credentials immediately ✅ Audit recent cloud access logs ✅ Pin to pre-compromise versions ✅ Wipe and reinstall from a clean lockfile TanStack has deprecated the affected packages and is working with npm security to restrict the malicious tarballs. Stay sharp out there.

New in Claude Code: agent view. One list of all your sessions, available today as a research preview.

Videom neden tutmadı diye ağlama devri bitti. Higgsfield yeni çıkardığı Virality Predictor ile "bu video kesin patlar" dediğiniz ama 3 izlenmede kalan o acı tabloyu daha videoyu paylaşmadan size gösteriyor. Klibi yüklüyorsun, sana "kardeşim bu video beynin sadece uykudan sorumlu bölgesini tetikliyor" diye ısı haritası çıkarıyor. Sadece skor vermekle kalmıyor; kancan zayıf mı, millet videonun neresinde sıkılıp kaçıyor hepsini tek tek yüzüne vuruyor. Üstelik Ad Reference ile başkasının tutmuş reklamını alıp kendi ürününle saniyeler içinde kopyalayabiliyorsun; yani "esinlenmek" artık profesyonel bir yapay zeka özelliği oldu. Algoritma bizi sevsin diye takla atmaktan yorulanlar için ilaç gibi çözüm gelmiş.

@trq212 wrote the post I kept meaning to write. Markdown made sense when humans were the ones typing and reading. Simple, portable, editable. Fine. But once agents are doing most of the generation — specs, plans, reports, workflows — Markdown starts to feel like the wrong container. You can't put a diagram in it. You can't make a table sortable. You can't add a slider so the person reading can actually try something instead of just nodding along. I switched to prompting for HTML artifacts a few months ago and the difference is not subtle. More context fits in one view. The people I share it with actually read it. And a link is easier to forward than a .md file no one knows how to open. The argument I hear against it: "harder to diff, harder to version control." That was true when humans were editing every line. Agents handle most of the editing now. The argument is getting weaker fast. The line from the post that stuck with me: "almost any data Claude can understand can be represented efficiently in HTML." That's a pretty big permission slip for how you think about agent outputs. Good read if you're spending real time building with agents: thariqs.github.io/html-effective…

In Claude Managed Agents, we’ve added multiagent orchestration, an outcomes loop for rubric-driven self-improvement, dreaming for self-learning, & webhooks.