About
I help build the GrowthX, an agentic compounding organic growth platform. My current work…
Articles by Matthew
-
Building a Power Loader for the Mind
Building a Power Loader for the Mind
First up, some news – just so we don’t bury the lede: I’ve joined GrowthX as Chief Content Officer. For the past…
24 Comments
Activity
2K followers
-
Matthew Panzarino reposted thisMatthew Panzarino reposted thisOur CEO has thoughts. About 10,000 hours of them! Marcel Santilli's workshop hit capacity on Day 1, so it's back for an encore: "What 10,000 Hours of AEO Taught Us" Friday, September 18 12:45 to 2:15 p.m. ET Room 104 If you missed it the first time, this is your shot. Come find us at Booth 67 until then!
-
Matthew Panzarino shared thisWe turned GrowthOS on for our first design partners this week. Launch posts on here tend to trend towards the grandiose, so I'll just talk about some of the cool stuff it does. GrowthOS is an agentic marketing stack for anyone trying to establish a source of truth about their business using organic content. You connect your site, kick off our Autopilot context agent, and a couple of hours later it has read everything you've published, worked out who your competitors are and who you're writing for, and built a profile of how you actually sound. I have some experience in understanding what a rich, accurate profile of a business looks like and the results of this process are as good as the context of an experienced researcher or reporter, or better. Assembling that business context used to take our team weeks of calls and shared docs. Now it's the starting point. GrowthOS does a few major things on a daily basis: - Watches every page you care about and tells you what's working and what's quietly dying. - Suggests what to write next based on where you're losing to competitors. - Deeply researches the opportunity you'd like to create and gives you a brief that is perfectly aligned to your goals and your audience. - Closely monitors your existing portfolio and the new content you just shipped to give you an honest view of how your strategy is performing. And the whole thing is wrapped in an agentic loop that keeps evolving and getting better the more you use it. Agents do the reading, structure content, and track it all. The humans are there for their editorial judgement and input. Every edit teaches the system more about how you sound, so the tenth piece or hundredth piece is more you, not less. Most AI content tooling runs the other direction: more volume, more sameness, then the traffic is penalized for that sameness and dies. If you run content programs, you're probably paying for the same lesson over and over. Every new writer, agency, or tool relearns your voice from zero. This stops with GrowthOS. The learning accumulates in one place and doesn't walk out the door with a contractor. And you get an honest picture of what your site is doing at the page level, every week, without commissioning an audit. If this sounds intriguing, I'll throw a link to book a demo in the comments. We're looking forward to talking with you!
-
Matthew Panzarino shared thisIn 2014, somewhere between another seed round writeup and a Series B announcement, I sketched out a plan to build automated support for writers of TechCrunch's funding posts. The problem was that we always had dozens of funding stories we would have loved to cover, really exciting companies we just weren't able to talk about, purely because of time and mechanical constraints. Back then, ML systems were still a transformer whitepaper away from being LLMs, but there were enough tools to see where we were headed. The idea I had to combat this was atomic units of content — small composable pieces you could recombine. A data layer pulling rounds, valuations, lead investors, sector and geography. A programmatic skeleton handling the scaffolding every funding post needs. And nested into this framework was the fun stuff for the writer. The juicy editorial core only a human reporter could actually provide. These core paragraphs offered the informed operator's point of view on the money, the product, the founders and their chances of success - it was what everyone scanned down to anyway. I jokingly called it Nucleus, based on the doomed software product in Silicon Valley, a show where TechCrunch was heavily featured (and parodied, hilariously.) Even though it was absolutely the right thing to do, we never built it. Because none of our corporate owners had the money or headcount to spare on the (at the time) heavy engineering lift needed to bring it to life. I've kept the sketch in my mind and I thought about it on and off for eleven years. The leverage problem has never gone away. The last great tools to ship for text content creators were arguably Wordpress and Google Docs. Then, when I got to talking to Marcel Santilli and Daniel Lopes a little over a year ago, I realized that we were all thinking about this problem in the same way. They were also working on content as composable units at GrowthX AI. Programmatic where it should be programmatic, human where it has to be human. It has always been a universal content team problem, no matter what kind of writer you are. It took a decade for the tools to emerge to be able to build the system that adds that engine to an editor's toolkit. Every content team has felt the pain of having to build and maintain deep business context, do deep research on a demanding schedule, find opportunities to create content, monitor the results of that content and collaborate across teams both internal and external to stay on track. We built GrowthOS for you.
-
Matthew Panzarino reposted thisMatthew Panzarino reposted thisThe AI dev ecosystem is missing its Postgres. Its Rails. Its React. That's the reason we open-sourced Output.ai. So many parts of the stack are products now: tools getting acquired, licensing shifts, your data being used in ways you didn't sign up for. We didn't want to build our company on top of that. We wanted a stable foundation that isn't someone else's business model. So we built our own. Output runs all of our AI infrastructure at GrowthX - our products and the agents we build for clients like Lovable, Webflow, Airbyte. We think other teams are dealing with the same problems and deserve more options. And selfishly, we want to meet and hire like-minded AI developers. Open source is the best way to find them. https://lnkd.in/grWTBPjsGitHub - growthxai/output: The open-source TypeScript framework for building AI workflows and agents. Designed for Claude Code describe what you want, Claude builds it, with all the best practices already in place.GitHub - growthxai/output: The open-source TypeScript framework for building AI workflows and agents. Designed for Claude Code describe what you want, Claude builds it, with all the best practices already in place.
-
Matthew Panzarino shared thisIf you're building agentic workflows in any shape, grab Output and give it a look, it's what we've built GrowthX on and it's radMatthew Panzarino shared thisAfter over a year of hard work we're open-sourcing our AI infrastructure. It's called Output [dot] ai. Here's why 👇 In 2024 we already had thousands of workflow runs a day at GrowthX tied to some unreliable dependencies. We wanted to rewrite the whole thing as code, we knew coding agents would eventually do a better job than any GUI. We'd be able to iterate in plain text, be 10x more productive, and handle 1000x the volume of runs we were doing with drag-n-drop tools. Today we have about 20 engineers building AI features into our platform or forward-deployed into companies like Lovable, Webflow, Airbyte to build custom AI agents for them. That means we start the same kinds of projects over and over - and we kept hitting the same challenges every time. We looked at the market and didn't feel like it was mature enough to bet on. Choosing your AI dev stack is a one-way door. You're going to grow a giant codebase in whatever pattern you pick. So many things felt like science projects that became successful because of lack of options or paid products - nothing truly OSS the way we liked it. We missed the kind of open source that powered the web - Django, React, Postgres, Rails. Tooling that exists to power companies, not to be a product. So we did what everyone is already doing - we rolled our own. But because of our forward-deployed model we got to see the same problems repeat across dozens of projects. - How do you iterate on a codebase packed with prompts? - How do you track what things actually cost? - How do you test non-deterministic code? - How do you create datasets from production data? Every team building AI is learning these skills right now. But because of our forward-deployed model we have to teach them over and over with every new project and every new team member. We needed the answers baked into the tooling, not in people's heads. We've been improving it for over a year now, focused on three things: 1. Making it easy for devs and coding agents to create and modify workflows in one or few shots (filesystem-first, TypeScript + Zod). 2. Self contained with minimal third party tooling sprawl (one framework instead of a dozen SaaS subscriptions). 3. A simple and relatively flat learning curve (conventions baked in so new engineers are productive from day one). We finally got to the point where we could finish the extraction and share it with the community. I'm really proud of what the team built. 👇
-
Matthew Panzarino reposted thisMatthew Panzarino reposted this🚨Signups to my newsletter are now open🚨 After more than a decade at TechCrunch, I’m starting my own tech publication. I’m calling it Hypertext. 💌 Click on the link in the first comment to sign up. So what is Hypertext exactly? It’s a hyperfocused newsletter about the most important stories shaping startups and venture capital in Europe — with more context, analysis, and opinion. My goal is simple: help people in the tech ecosystem make sense of what matters with better content than the average ChatGPT-fuelled misinformed BS that you can read in your newsfeeds. Have a good day ☀️ Romain #tech #startups #newsletter
-
Matthew Panzarino shared thisIf you have any sort of agent set up using Claude or Claude Code (or any others tbh) remember to regularly run conflict checks across all of your artifacts and files. Conflicts kill agent efficiency and will break their measurement rubrics, making it a coin toss whether you get amazing output or trash. If you have multiple rubrics for what is 'good' across multiple files, you basically end up with a multiple wave pattern as it tries to adhere to these potentially conflicting instructions. Your output then ends up at whatever quality level it landed on at the point in the wave where it hits max iterations. 'Hey claude, run a conflict check across all of my instructions to you. Tell me what the conflicts are and how you would resolve them.' Guarantee there is something and that your setup will work better after this check.
-
Matthew Panzarino shared thisIf you're engineering with Claude or building any SaaS tools with LLMs, you really should check out Starter - it's an engine for building your own tools that has been proven out by the crazy talented team here at GrowthX.Matthew Panzarino shared thisWe just open-sourced the starter kit we use to build all of our core products. (🔗 Link in comments👇) --- We use Next.js a lot for some client projects. It's great for smaller apps. But the backend holds most of the real complexity of any sizable product. All biz logic, data integrity, auth, jobs, migrations - that's where the hard problems live. And frameworks like Next.js have almost no opinion on any of it. You're stitching together community packages for everything. ORM choices, job queues, caching, project structure - all on you. Every dependency is a bet on someone else. And a bet on your team keeping it all well organized. And now with AI - one bad hire with Cursor and you're stuck with a giant mess. Frameworks like Ember/Redwood don't get even remotely close to the maturity of something like Rails/Django. For our core products we went with Rails + React + Shadcn - but without the need for an API layer (Inertia.js). Works so well. Best of both worlds - no SSR hydration, no state management, no routing drama. Coming from vanilla Rails/Turbolinks? The devx is basically the same, but with React's awesome ecosystem available. Stimulus is just not good enough anymore with Claude Code doing a lot of the frontend work. --- (🔗 Link in comments👇)
-
Matthew Panzarino reposted thisMatthew Panzarino reposted thisThere are two types of AI users right now. Compounders vs Skippers **I'm doing a 4-hour hands-on WORKSHOP 🤓 on how I use Claude Code/Cowork + Cursor to build my agent context OS. Details at the end.** ——— [COMPOUNDERS] Obsessed with using AI to learn. They connect dots faster, synthesize information, compound their knowledge like crazy. They almost treat it like a video game. How much leverage can I build? How much more can I do? They get a high from it. ——— [SKIPPERS] Use AI so they don't have to think. Not to think better. They're not reading. They're not critically evaluating anything. They're not connecting dots and bring legit judgement to the equation. BTW you can usually find this by their ability to write clearly. Bad writing = lazy thinking. ——— The gap between these two groups is getting insane. I'm talking 10x... and it's compounding! I've spent 200+ hours in the last two months building what I call a context operating system. For myself and for the AI agents working on my behalf. [TIPS TO COMPOUND] → Curate your context. You're building a knowledge graph for yourself and every agent working for you. Your context is your moat. → Use AI for what it's actually good at. Fetching information. Connecting patterns. Processing at scale. Not replacing your thinking. → Make every process compound. Every output feeds the next input. Nothing is throwaway. Nothing will replace the time you need to spend with it. PLEASE please stop trying to download some damn skill and think you're job will be done! Be the person that has the best know-how and knowledge and knows how to distill it all to build leverage. Not the one that comments on a post and produces 100x more slop daily. ——— [WORKSHOP] I built this for myself first. Now I'm training people inside our company on it. Started coaching a few people externally too. I'm putting together a small virtual workshop. Up to 50 people. 4 hours. The entire system start to finish. Drop a comment or text me if you're interested. I'm just finalizing details and date but will be in the next 6 weeks.
-
Matthew Panzarino liked thisMatthew Panzarino liked thisWe're hiring an M&A reporter at Axios Pro Deals. If you like breaking deal news and have sources across banks, law firms and the rest of the M&A world, come work with us. https://lnkd.in/g2itzxGq
-
Matthew Panzarino liked thisMatthew Panzarino liked thisI’m officially a boomerang. I’m excited to be back at Gazetteer, this time as a managing editor helping with editorial strategy and processes for both digital and print. A good amount of my work will be behind the scenes but you shouldn’t be surprised if and when you see my byline. I won’t have a beat, per se, but it’s safe to say you’ll see me writing about the Valkyries and professional women’s sports in the Bay (SF Firebells, anyone?). Very excited to be back. Let’s get it. https://lnkd.in/gP4YVHVr
-
Matthew Panzarino liked thisMatthew Panzarino liked thisHands. Tried something new: a visual exploration pulled together from a century of archival shots, co-edited in Palmier plugged into Claude Opus + ChatGPT Astra. Used The Moving Image Archive + Destockd to grab the right visuals & its set to an instrumental of "Us" by Regina Spektor, one of my all-time favorites ✌🏽
-
Matthew Panzarino reacted on thisMatthew Panzarino reacted on thisToday is my birthday, and I keep thinking about a school cafeteria table. I was 11 when we moved from Brazil. I didn’t speak English. I was a stranger in what felt like a completely different universe. School became a daily exercise in trying not to be laughed at. Trying to blend in. Trying to get through lunch. I can still see the cafeteria. It seemed like everyone had a table. Everyone knew where to go. I didn’t. I think some part of me has been standing in that room ever since, looking for a place to belong. For a long time, my answer was to work harder. Outsmart everyone who wouldn’t let me in. I watched people closely. I learned to spot patterns, think ahead, figure out how to get their attention. Those skills have helped me build a life. They also came from a kid who was scared. Then I learned graphic and web design. I could sit for hours making things, like I used to with Lego. I built a website and figured out how to get customers online. The first time a real business filled out my form and paid me for work, I was blown away. I was a teenager. They had no idea. I’d made something and someone found it. For once, I didn’t have to wait for someone to pick me. I can see that thread through so much of my life now. Building things people might care about. Giving away value. Hoping the right people would find their way to me. MY PEOPLE. I wanted more than customers. I wanted a team I loved being around. I wanted to make a place where we could do work we believed in and where other people felt they belonged too. My team and I are building that table at GrowthX AI. We help companies figure out what they stand for, make something valuable, and find the people they’re here to serve. I think of a website as their TABLE: a place to show up as themselves and invite those people in. And honestly, the fear hasn’t gone away. I still wonder if I’m faking it. If I’ll fail and everyone will see. I know where that voice came from. I’m trying to hear it without letting it run my life. … Today, I just want to express my gratitude to the people around me. Those that have joined my table in some way. At the team building this with me. At everyone who made room for me along the way. Thank you. It means the world to me. Much love ❤️
-
Matthew Panzarino liked thisMatthew Panzarino liked thisAI visibility can tell you plenty. Just not from a single score. Your AI visibility dropped 30% this week. That number can't tell you why, or which page caused it. It's one number covering every question a buyer might ask you. So ask one buyer stage at a time. Each one needs something different. - When they're still working out the problem: are you cited at all, and which sources are showing up instead of you. - When they're building a shortlist: how often you show up, where you sit in the answer, and what it says about your support and how easy you are to use. - When it's down to you and one competitor: which one the answer picks, and the reason it gives. - When they're ready to buy, or already bought: whether what the answer says about your product is right, and whether it came from your site. Each of those points at different work. Being included alongside a warning that your setup is difficult is a different problem from being absent altogether. "Improve our AI visibility" is not a task. Any one of those questions is. The full version, including how much weight to put on a single reading: https://lnkd.in/gdcWernmAI Visibility: What AI Visibility Can Actually Tell You | GrowthXAI Visibility: What AI Visibility Can Actually Tell You | GrowthX
Experience & Education
Other similar profiles
Explore more posts
-
Conduit Digital
Landing a "white whale" client is exciting. 🐋 Delivering on it without breaking your team, margins, or systems is where most agencies struggle. The agencies winning bigger accounts in 2026 aren't scaling headcount — they're scaling infrastructure. Clear goal frameworks, unified execution, and performance systems that hold up under pressure. That's how you say yes to larger clients and stay operationally sound. Learn how to land bigger clients without blowing up your ops: -https://lnkd.in/e-V5HaMq
-
Kristyn Wellesley-Weisz, M.A., AIGP
If you run a newsroom, this should get your attention. A New York State bill introduced this week would require disclosures on AI-generated news content, mandate human review before publication and reinforce protections for journalistic sources and data. What stands out to me: -- A clear expectation of humans in the loop for AI-assisted journalism -- Disclosure requirements for audiences when content is substantially AI-generated -- Explicit recognition that source protection and data privacy apply in automated systems This isn’t about banning AI in journalism but about who carries responsibility when it’s deployed. Whether this bill advances or not, it reflects a shift I’m seeing more broadly: policymakers are starting to encode long-standing journalistic norms into AI governance frameworks. For newsroom leaders, that’s the signal. If you don’t have clarity on: -- where AI is used in your newsroom -- who reviews and approves AI-assisted content -- what data (including source material) AI systems can access or retain, you’re behind the risk curve. Governance isn’t a brake on innovation. It’s how audience trust survives automation. https://lnkd.in/e_sBRUpU
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top content