Here are the links and tools that caught my attention in February.
The AI coding workflow wars
This was the dominant theme of my inbox this month. Everyone is figuring out how to work with AI coding agents, and the approaches are diverging fast.
Mitchell Hashimoto wrote about his AI adoption journey. He describes a 6-step progression from skeptic to productive user. The key insight: force yourself to reproduce your manual work with agents and invest heavily in “harness engineering” (AGENTS.md files, scripts, feedback loops). The efficiency gains come not from AI doing everything, but from learning what to delegate vs. what to do yourself.
Boris Tane wrote about his Claude Code workflow and his core discipline: never let Claude write code until a detailed plan has been researched, written, and iteratively reviewed. Implementation only begins after the plan is fully approved, at which point it becomes “boring” mechanical execution. I’ve been somewhere in between these two approaches.
The Claude Code ecosystem is exploding with session management tools. I saved at least five this month: agent-view, agent-conductor, Vibe Kanban, get-shit-done, and Plannotator (visual plan review for agent plans). I also bookmarked OpenSpec, a spec-driven framework where every code change produces a “spec delta” showing how requirements changed.
People are building real engineering processes around agents. However, despite saving all of those tools with good intentions to come back to them and give them a try, I’ve just been iterating on my personal flow which can differ from project to project. One thing that will be interesting to see is which of these tools are relevant in six months or twelve months? Either way, I love seeing people building and moving the ball forward.
Simon Willison wrote that AI doesn’t reduce work, it intensifies it. A Berkeley study of 200 tech employees found AI creates constant multitasking: managing parallel agent threads, checking outputs, reviving deferred tasks. My experience maps to this! Simon mentions his mental energy depletes after an hour or two but I’m not feeling the same thing. I feel energized by how quickly I can go from idea to prototype.
The taste debate
Two pieces wrestled with skill, taste, and AI.
“No Skill, No Taste” argues that LLMs have created an illusion of a lower barrier to entry, but the real barrier was always taste. Knowing what’s worth building and for whom. A flood of derivative vibe-coded apps now clutters Show HN because people overestimate both their skill and their taste.
A Hacker News comment offered a counterpoint worth saving. The real vibe coding pattern isn’t “make me Stripe.” It’s directing AI to study open-source alternatives and adapt them to specific needs. Google used AI agents to translate their entire x86 codebase to ARM, a project that was economically impossible with human labor alone. The tool isn’t the problem. The wielder is.
In a river of slop, good product will win.
Data and visualization tools
Simon Willison (again) announced Chartroom and Datasette Showboat, tools for live-updating documentation from coding agents. The idea: agents push document updates in real-time so you can watch progress live.
OpenAI published a piece about their in-house data agent built on GPT-5.2. It handles end-to-end analysis: discovering tables, writing SQL, running queries, generating visualizations, all with a self-correcting loop. Available across Slack, web, and CLI via MCP. The scale is wild (600+ petabytes, 70k datasets), but the pattern is the same thing we’re all trying to do at 1/1000th the scale.
The Pudding is a digital publication I somehow hadn’t encountered before. Visual essays that explain ideas through data journalism and interactive storytelling. 215+ stories on everything from mapping happiness to onion-dicing optimization. Bookmarked heavily.
Stuff I bookmarked
- Happy: Open-source mobile/web client for Codex and Claude Code. Claude Code also released their remote feature and I’ve been doing a lot of remote coding on exe.dev. Lots of people working on coding-on-the-go.
- Zellij: Terminal multiplexer.
- MonoSketch: ASCII diagramming tool. For when you want to draw architecture in a README.
- Valetudo: Cloud replacement for robot vacuums so they run locally only.
- comma.ai’s $5M data center: They run 600 GPUs on-premises instead of using the cloud. 5x cheaper than cloud equivalents. Managed by two engineers in San Diego with open-air cooling. I saved comma.ai last month too. These people are interesting.
- H3: Uber’s hexagonal hierarchical geospatial indexing system. Was recommended by a friend and I’ve already implemented at Hello Gravel for some projects.
- Brickanta: AI-native platform for construction planning.
- Build with Star: Create playable games instantly.
- Lethe: Autonomous executive assistant. OpenClaw with a fancier landing page?
- Strix: AI penetration testing agent.
- Anthropic Courses: Official Anthropic training courses.
- QBO CLI: QuickBooks CLI tool. Recently started using Claude Code to update our forecast each month and may give this a try.