One rule that has held up: never hand the machine a task you could not do or check yourself. Why fundamentals are the error-correction layer no abstraction has replaced, why AI is not the next compiler, and what the industry's drift toward tool-fluency costs the people who learn and hire by it.
Read articleFor two years the scarce thing was silicon. Now it is electricity: data-center power demand is set to double by 2030, roughly 75 US projects were canceled for lack of it, and the squeeze reaches all the way down to your cloud bill and your architecture.
Read articleVibe designing lasts where unguided vibe coding did not, but not for the reason people give. The durable reason is a faster first loop that lets taste lead, plus the same structure that makes any generation trustworthy. And the checks the eye can never run on its own.
Read articleThree hours in, your context window is full. A subagent hands the side task to a separate instance with its own context, tools, and model, and returns only the result. How to build narrow specialists, four to steal, and why it ports to Codex.
Read articleForget one dimension versus two. The question that actually decides it is who sizes the items, the content or you, plus the five gotchas agents get wrong most.
Read articleAsk an agent for a dropdown and you usually get one the keyboard can't reach. The div that should be a button, the missing focus ring, tab order that fights the markup, all found in a thirty-second walk.
Read articleSpecificity is step four, not the whole story, and it never carries the way people think. How the cascade actually picks a winner, and why :where() and @layer end the !important arms race.
Read articleA model can't reliably separate your instructions from instructions hidden in the content it reads. Why prompt injection hits local coding agents hardest, and the architecture that contains it.
Read articleMost of what I believed in my first month with AI agents turned out to be wrong. Ten lessons - from why prompts were never the bottleneck to why judgment became the scarce skill.
Read articleAI code fails in predictable, well-documented ways. This is the review gauntlet: five failure classes and the fast tells that catch the most damage per minute of attention.
Read articleYou've gone from conductor to orchestrator. The bottleneck moved from the agent's capability to coordination. Here's how to run parallel agents without quietly corrupting your codebase.
Read articleThe people who build the most-used coding agents stopped hand-prompting. The unit of work is no longer a prompt. It's a loop that prompts, checks, remembers, and re-runs the agent for you.
Read articleEvery AI-generated site has the same tell. The fix is direction: a real brief, a design plan, and your boldness spent in exactly one place.
Read articleThe honeymoon with “vibe coding” ends the moment you point an agent at a real, messy codebase. What replaces it is a discipline you have to practice.
Read articleWhen an AI agent keeps getting your codebase wrong, the answer is to decide, deliberately, what it gets to see. A smarter model or a longer context window rarely helps.
Read articleAI agents write code fast, and they write insecure code with exactly the same confidence as secure code. Five practices worth building into your workflow.
Read article