A collection of 26 posts about ai.
I built Daytripper and RepoFinder entirely with OpenAI Codex, then turned my Claude Code tooling loose on the results. The QA and design findings are numbered in the git history, one regression test cites the report that found the bug, and both products came out better than either agent would have produced alone.
RepoFinder turns a project and a goal into a short list of open source tools that genuinely fit, then lets you interrogate every pick. Built entirely with OpenAI Codex on the Responses API: gpt-5.6-luna for extraction, gpt-5.6-terra for ranking and chat, strict JSON Schema outputs, a labeled fallback, a remote MCP tool, and 15 checked-in lessons on how Codex actually works.
I learn AI coding agents by shipping with them. Daytripper is the first app I built to learn OpenAI Codex: a travel planner built in one overnight session, deliberately zero LLM calls at runtime, with tests covering all 1,296 possible preference combinations, deployed to Cloudflare Workers. Here is what the build taught me about how Codex actually works in August 2026.
Anthropic ships cookbooks for financial workflows. I did not see one for the problem I actually have: taking over a codebase with no docs, no tests, and no author. So I wrote it. This is the whole thing, top to bottom - the recipe, the building blocks (commands vs skills), the guardrails (ask vs enforce, hooks and all), and why Claude Code is the tool for a regulated fintech with three teams to serve.
Rocket Mortgage put Claude Code in front of the engineers who own their mortgage origination platform. A migration that had dragged for nearly 3 years shipped in 4 months, senior engineer throughput went up 145%, and client-impacting incidents after the cutover: zero.
I caught myself confidently misremembering my own architecture, so I asked Claude Code to read the code and tell me the straight truth. It corrected three things I believed, found an MCP server I forgot I shipped, and left me with a mental model worth sharing: southbound REST, northbound MCP.
RepoRadar was vibe-coded in four hours. Repo Recommender is the follow-up, built slowly and in the open with Claude Code: plan mode, CLAUDE.md, custom skills, a commit-blocking hook, an MCP server, and a 23-lesson book documenting every step.
$150M. 1,000 fellows. $85K salary. Full time, in person, embedded in nonprofits across America. This is what AI "spilling into the real world" actually looks like.
On June 12, the US government ordered Anthropic to suspend Fable 5 and Mythos 5. No statutory process, no evidence standard, no appeal. If this can happen to a frontier model, it can happen to any model.
When Claude's Builders Day turned us away, Hudson and I ended up at Corgi Cafe reading Dario's essay about whether humanity can survive its own technology.
I'm tired of my Mac Mini powering down or dropping my agent while I'm away from home. Here's the exact pmset and LaunchDaemon setup that made it self-healing, including the bootstrap error that almost made me give up.
My BSides SF 2026 talk with Dan Hollinger is up. We showed how Claude Code can bypass bot detection and solve CAPTCHAs in minutes using Playwright Stealth. The full talk, live demo, and Q&A.
We shipped RepoRadar in four hours at the AI Tinkerers SF Generative UI Global Hackathon: a radar that ranks trending GitHub repos, then an agent deploys a working micro-app for any of them at its own subdomain. Here's what we built, how it works, and how to grab it.
Cloudflare and Stripe shipped an agent provisioning protocol last week that lets autonomous agents create accounts, register domains, and deploy production code. I work at Cloudflare. Here's why this is the announcement I keep coming back to.
There's a new agent product announcement every other day. Nate B Jones gave me a way to triage them in 30 seconds. I'm using it.
Cloudflare ran Agents Week last week. I work there, and even I needed a second pass to absorb the scope. Four announcements actually change how I'd architect a personal agent stack today.
Matthew Berman has spent 2.54 billion tokens learning what OpenClaw is actually for. The list of 21 use cases is interesting. The decision underneath is more interesting.
My friend Vic is a C++ programmer who teaches people how to maximize credit card points. He had an idea. We sketched it on paper for ten minutes, then built the prototype in thirty-five.
Matthew Berman's framing in his March OpenClaw tutorial reframed how I think about my agent. Hope isn't a script I run. She's someone I onboarded.
Alex Finn's case for running OpenClaw on local hardware is right. His specific model recommendation isn't, if you're on the base Mac Mini I'm running. Here's what fits and what to use instead.
Everyone's arguing about which model is smartest. The real game is being played across three layers - and the most interesting one is the one nobody's talking about.
The AI builder community is doing something rare - actually helping each other. From open-source tools to strangers encouraging you to start a channel. Here's why that matters.
Most people use AI as a chatbot. I run autonomous agent fleets that build, trade, research, and create while I sleep. Here's why the multi-agent approach changes everything.
25 years in solutions engineering taught me how to solve problems at scale. Now I'm applying that same thinking to building AI-powered products.
Everyone debates which model is smarter. The real constraint that determines what AI agents can actually do? The context window.
Hudson and I built Moder.fun together. It's not about the product - it's about showing kids they can build anything.