Daniel Kennedy

What skills make for a good vibe coder? I think it’s: 1. Relentless work ethic 2. Insatiable intellectual curiosity 3. Strong systems thinking (OS, to networking, to full app stack) 4. Very strong communication skills to both LLMs and humans

Fantasy football drafts are optimization problems. This app ingests your league history analyzes it for opponents tendencies, then gives you the optimized pick (based on final roster build). It’s really just a question of when to draft 1 position vs the next. footballquants.com

There’s no reason to use an ORM anymore. SQL is still high level, but it’s more descriptive and controllable than an ORM. With AI assistance, why use an ORM?

Writing better code than a frontier tool is becoming like beating an s&p index fund. You can try, but for 99% of people, just go with the market (tools suggestion). But for those that can beat the market, they will be paid handsomely.

I’ve cancelled my Claude max subscription. Codex is simply better, and cheaper per task.

You can still read every line of code, you will just be much slower and less productive than those who don’t. Figuring out how to safely ship code you don’t read is the challenge.

For fantasy football 2026: here’s a guide for standard scoring. Get your RBs early. WRs at value. And wait on your QB. footballquants.com/reports/s…

What’s the best OS to use with codex or Claude code? I can imagine a specialized OS fit for optimal tool use would make a difference. I mostly use windows or macOS. And find windows less efficient. What’s best?

Coding agents really don’t have strong opinions or engineering intuition. Depending on your prompt they will make decisions that sound correct, but if you push back even a little it will flip flop and assert something else is right, despite saying something entirely different before. This is why it’s still entirely on the engineer doing the prompting to know what is correct. Agents can’t hold real, grounded opinions, like a senior architect would. A real architect would never get pushed around on technical decisions by a junior product person suggesting slightly different things.

Anyone can prototype an idea with AI, but any large project will fail unless you apply strong engineering fundamentals. I pointed my codex cannon at this idea to build a tool to help build the engineering intuition for the projects that matter. engineeringintuition.com

My workflow: 0. High level behavior / architecture diagram of a feature. 1. Define the intent of the code with plain English assertions. 2. Write tests that validate those assertions at the e2e level, backed up by integration tests, and unit tests. 3. Have AI Write the application code. 4. Test.

Fantasy football has changed over the past 4 years dramatically. Managers have gotten significantly sharper, very likely because the apps we use publish more accurate projections. It’s incredibly hard to find an edge in 2026 over any of the main apps projections. footballquants.com/reports/w…

Another risk of using cloud based AI. They can literally stop your work midstream whenever they like. Regardless of your plan, or how much you’ve used. I have no recourse here with Claude. If I want to continue work, I switch to a different provider, continue by hand, or wait. Now imagine there’s a war room going on and customers can’t access your service. But you have no idea how it works. You’ve relied on AI to build and maintain it. And you can’t use your AI provider.

AI development is all fun and games until you try to work with 5 other people on an enterprise app. All slinging thousands of lines, dealing with vulnerability scanners, merge conflicts, etc. It’s workable with the right systems in place, but it’s not easy.

There’s so many great resources and approaches to early childhood education, but it’s not easy to see them in an organized way. I made this to help all the parents and educators out there. It’s open source and I’d love for anyone to contribute a new activity you do with your kids. If you’re into AI, point your agent at the MCP endpoint and ask how you can help! earlyatlas.com

I still review PRs for developers who write code by hand, and it’s a stark contrast looking at . AI is not perfect, but neither are humans, and handwritten code is much more prone to bugs and poor style. Pre AI projects with no automated tests/CI gates still exist and it’s like stepping into a different era when reviewing them.

MCP hate is misguided. MCP is simply a protocol that works well with agents. I can make a good MCP server that a new agent who has never used my site can interact with and learn how to use the features on my site. No need to read openAPI specs or external documentation. The concern about schema dumping 10k + tokens on an initial tools/list is real, but is avoidable with smart design.

If you’re building a basic marketing site for a small business with SEO and an inquiry form what approach would you recommend? A builder site like wix/squarespace? Or do a hosted Wordpress? Or something custom like Astro/aws/cdn/lambda?

Loops are great, but too expensive to run in pure frontier models. And not high enough quality to run in pure open source models. I’m building a hybrid approach where frontier gpt acts the supervisor, and qwen works as the worker implementing what gpt tells it to do. I’m finding this reduces token usage on the gpt model but maintains high quality. Anyone else doing this?