Giovanni Rosati

I am pleased to share that I have successfully passed the re-certification process for my board-certified clinical specialties in orthopaedic and neurologic physical therapy.

I recently stopped using AI primarily as chat and autocomplete after watching this YouTube video by Nick Chapsas on agentic development - https://lnkd.in/gwgAy5rk. I just turned 62, and I have about four decades of coding behind me. I’ve been clinging to the craft of writing code by hand. That’s fine — until it becomes the thing that causes me to get left behind. I tried vibe coding. I do not have the emotional fortitude to handle the chaos of a constantly changing blast radius from the latest AI hallucination. A few weeks ago, I pivoted to Specification Driven Development (SDD). There are many ways to approach SDD, including GitHub’s Spec Kit and others. What has worked best for me so far is simple: three markdown files and project skeletons. PRODUCT.md — What the product does, who uses it, and what it must not do. Roles with explicit Can/Cannot lists. Goals, non-goals, success criteria, and open questions. MODEL.md — Every entity the system stores: fields, types, relationships, lifecycle states, and business rules. No implementation detail — just what exists and what constraints govern it. WORKFLOW.md — Every meaningful user action, written as a sequence of steps: who does it, what must be true first, what happens, and what can go wrong. Then I provide project skeletons so the agent can adopt my preferred architecture, folder structure, coding conventions, naming conventions, libraries, NuGet packages, and related patterns. I keep cloud resources, security, infrastructure, and non-dev code branches under human control. That has been critical for keeping agents adequately sandboxed. At first, I almost slipped back into waterfall thinking while creating the specification files, and I got overwhelmed. Focusing on the MVP saved me. I’m not claiming this is more correct than any other process, but it is working well for me so far. I estimate at least a 4x velocity improvement over coding manually, with hopes of further improvement. I still review all of the code in every PR because the artifact is still my responsibility. If you’ve found a better way to keep agents productive without letting them run wild, I’d love to hear what’s working for you.

Organizations grow when people are willing to challenge assumptions, rethink systems, and help others see possibilities differently. At Krause Group, Heather Schott brings that kind of leadership to her role as Director of Diversity, Equity & Inclusion. Her work connects people, business strategy, and long-term growth across our businesses in the U.S. and Italy — helping teams think differently about culture, access, talent development, and organizational accountability. Heather’s approach is grounded in problem solving, not optics. From shaping workplace systems to navigating complex conversations across soccer, wine, and real estate, she consistently pushes organizations to operate with greater awareness, adaptability, and intention. As a gender non-conforming leader, Heather also brings perspective that challenges conventional thinking around leadership and inclusion while embedding behavioral changes in business spaces — strengthening conversations around innovation, equity, and growth. Her work continues to support Krause Group’s broader commitments to Gender Equality and Reduced Inequalities through the United Nations Sustainable Development Goals. 

Andrej Karpathy broke down exactly what Claude Code keeps doing wrong. A creator created a repo that hit 110K GitHub stars in under 3 months and stayed at number 1 trending for 28 days straight. Here is every principle inside it: ☑️ The Origin Story → Karpathy coined vibe code, ex-Tesla AI Director, and OpenAI co-founder. → Forrest Chang distilled his X thread into one CLAUDE. MD file → It got 5,828 stars in a single day. 110,000+ in under 3 months. ↳ Why it went viral: The 4 principles are not novel. Every developer recognised their own frustrations instantly. ☑️ The 4 Karpathy Failure Patterns → Models make wrong assumptions on your behalf → They overcomplicate code, build 1000 lines when 100 would do. → They change code they do not understand, even when unrelated to the task. → LLMs loop until they meet goals, but need those goals explicitly defined. ☑️ The 4 Principles Inside the File ✦ 1. Think Before Coding: State assumptions before implementing. If uncertain, ask. Never pick silently. ✦ 2. Simplicity First: Minimum code that solves the problem. Nothing speculative. If you wrote 200 lines and it could be 50, rewrite it. ✦ 3. Surgical Changes: Touch only what you must. Clean up only your own mess. Every changed line must trace directly to the request. ✦ 4. Goal-Driven Execution: Define success criteria before starting any multi-step task. Strong criteria mean Claude loops independently. Weak criteria mean constant clarification. ☑️ How to Use It Today: ✦ Claude code Users: → Drop CLAUDE. md into your project root folder. → Claude reads it automatically at every session start. ✦ Cursor users: → The repo includes .cursor/rules/karpathy-guidelines.mdc and works across any project without re-adding the file. ✦ Claude. ai users: → Paste the 4 principles into your project instructions and add them to your SKILL as a negative instructions block. ↳ Get it at: https://lnkd.in/g5gnxdHu ☑️ The Dos and Don'ts → Do: State assumptions before writing a single line of code. → Do: Write the minimum code that solves the problem. Then stop. → Do: Define success criteria before starting any multi-step task. → Do: Ask clarifying questions before implementing, not after making mistakes. → Never: Make silent assumptions. It is the number 1 failure mode. → Never: Add features, flexibility, or abstractions nobody asked for. → Never: Refactor or improve adjacent code. It creates collateral damage. → Never: Use weak goals like "make it work." Claude will guess endlessly. For more AI systems like this: 👇 → Go to https://lnkd.in/gKeWjXUN → Subscribe to my free newsletter (don't pay anything) → Get more free and daily cheatsheets Which of the 4 failure patterns are you seeing most in your own Claude sessions right now? Drop it below. ♻️ Repost to give your network an unfair advantage.

Everyone’s panicking about AI taking their job. But here’s what they’re missing: AI can write code. AI can analyze data. AI can generate images. But AI cannot hold your hand after a diagnosis. AI cannot inspire a room full of students who’ve given up. AI cannot look a client in the eye and build trust. The jobs that will outlast AI aren’t the ones with the most technical skills. They’re the ones with the most human skills. Empathy. Judgment. Creativity. Leadership. Intuition. These aren’t soft skills. They’re survival skills. See below for the 10 roles that will thrive (not just survive) in the age of AI. (Image credit to Daniel Hartweg) Save this. Share it with someone who’s worried about their future. And if you’re job searching right now, ask yourself: “Which of these human qualities am I leading with on my resume and in interviews?” Because that’s the real answer to AI-proofing your career. ♻️ Repost to help someone shift their perspective today. 🔔 Follow me for job search strategies that actually work in 2026.

What I’m seeing a lot lately are AI generated websites and vibe coded sites that fall completely flat. Flat on brand, flat on structure, flat on accessibility, flat on SEO, and flat on compelling content and design. And I can generally pick them out at first glance. Immediate giveaways are incorrect H tag structure, incorrect or no schema, and absolutely no soul. They’re quick(ish), cheap, and easy, but are they really serving you or the user? Probably not. Powerful tools with untrained guidance are just powerful tools.

It's been a year of stepping outside of comfort for me. I read a poem in February at an event this group put on. I found out, I enjoy learning from the form of art! Maybe you will too if you spend Saturday night experiencing the excitement and energy of a live poetry slam! Iowa Poetry Slam Championship: On Saturday night, witness the top Iowa spoken word poets compete for the title of 2025 IPA Poetry Slam Champion. The winner will receive cash prizes, publication opportunities, and the chance to represent Iowa at the national competition. If you haven’t been to a slam, this will be a loud, lively, unbelievable creative event – inspiring and energizing. #FindYourJoy #KeepExpandingYourMind

A few months back a customer came to us after blowing their entire annual OpenAI budget by several million dollars. The worst part was that it was only Q1. Thankfully, they were already spinning up on-prem GPU infrastructure to get off the expensive per-token pricing, but the question that kept coming up was was the same: "How do we know that our on-prem models are as good what we're getting in the cloud?" Nobody wanted to be the person who cut the cloud API bill in half, but frustrated thousands of users in the process. They needed real numbers, human feedback, and strong metrics. That's the problem that this reference architecture solves. you can log traces from both your cloud and local models, curate evaluation datasets from real production traffic, run the same prompts against your local models, and compare scores side by side. We built it with Comet on their Opik LLM observability tool, and it runs on an Everpure (formerly Pure Storage) FlashBlade. If you're thinking about this migration (or already in the middle of it), take a look. https://lnkd.in/gjkKF7TD Huge thanks to Chase Fortier, Nikolas Laskaris, Sarah Ostermeier and their team in bringing this together! 🤘

I recently served as editor on a published article for Revolver Gallery. The article revisits Andy Warhol’s Time Capsules—the more than 600 boxes of receipts, invitations, and everyday ephemera he sealed between 1974 and 1987—and asks a contemporary question: have screenshots become the modern equivalent? And what gets lost when we save everything? Within our monthly publishing cycle, I managed developmental and line edits through final proof. Grateful to continue shaping writing at the intersection of art, media, and archive. I hope you enjoy reading it as much as I enjoyed working on it. 🔗 https://lnkd.in/ggEJrsvz