Articles

Gemini 4 Carbon is Google's strongest security model, but only selected teams can use it. What it does, why it launched closed, and what to use today.

GitHub's Octoverse 2026 shows 40% of pull requests now start with an AI agent. The key numbers, and what the teams where it works do differently.

Tests, legacy code, debugging, review, refactoring and SQL: what I ask the AI for in each, what I watch for, and how I check it in minutes. And the three decisions I never hand off, with a real case from this month that shows why.

Developer job postings are rising again, but not for everyone, and interviews are dropping algorithm puzzles. What October's data shows, what was debated this week, and six skills you can start practicing tomorrow.

A developer let Claude Opus 5.5 translate the TypeScript compiler to Rust for two weeks. It passes all 184,749 tests and he says he never read a line. What happened, what's overhyped, and how to use the same structure for a migration in your own project.

A Harvard study of 718 companies measured what happens when a team adopts coding agents: people write 30% more code, but no more work gets finished. The time is lost in code review. What the data says and what you can do before you send a change.

TypeSafe launched Jev, an AI model that doesn't generate text: it answers closed questions about a piece of text in a tenth of a second, for a fraction of the price. What it is, how to use it, where it fails, and how to try it.

What a language model's context is, why it fills up without you noticing, and four simple habits so the agent doesn't forget what you asked. Nothing to set up: it works from your next conversation.

Anthropic released Claude Opus 5.5: cheaper, faster, and with three new restrictions. What it does to the bill for a real session, when fast mode is worth it, who it breaks things for, and how to test it on your repo.

What should run on its own every time you push code, in what order, and what is better left out. A complete example and the rules that keep automatic deployment from being scary. It takes an afternoon to set up.

The anatomy of a specification, the five EARS templates for writing requirements with no gaps, and why plan mode is the cheapest way to start. The practical side of working with specifications.

The specification as the source of truth and code as a derived artifact. How the workflow actually looks, the real problem it attacks, the three traps teams hit with it in production, and the size of change where it stops paying off.

Claude Code, Codex, Copilot and Cursor now run the same frontier models and cost roughly the same. What separates them is the harness: how much you supervise, how much you delegate, and where it runs. How to choose by the work.

Why the agent hands you code it never verified, and the six things about the repository to fix so it can check its own work. None of them is a rewrite and none takes more than an afternoon.

What changes when the migration of a Web Forms app —or any other legacy— to a modern stack is done with a coding agent: the seven steps I would not skip, the most expensive mistakes, and how to avoid ending up with a new legacy written in a newer language.

Coding fast with an agent and coding well aren't the same thing, but they're not contradictory either. The habits and tools — CLAUDE.md, plan mode, subagents, hooks — that keep code scalable and clean, even when Claude wrote it.

A Skill is a folder with instructions that Claude reads on its own when it's needed, without taking up context until then. How to use the ones that already exist, how to build your own step by step, and what to check before installing someone else's.

Archify is a free skill for Claude Code, Codex, Cursor, and opencode that draws verified architecture diagrams instead of invented ones: it splits writing a typed description of the system, validating it, and only then drawing it.

A coding agent can burn half its tokens on a log or a JSON blob that never needed to be sent whole. Headroom is a free tool that compresses that before it reaches the model, without changing the answer, and without sending your code anywhere to do it.

Ten commands that solve the same four problems you already have: the context that fills up, the setup you never did, the change that has to be undone, and the work that does not fit in one turn.

Three free tools you need to know exist. One imposes a way of working before any code gets written. Another shows you where your tokens go. The third gives it design judgement. What each one does, how to install them, and what they cost you.

Almost nobody commits a password on purpose. They leak through four routes that do not look dangerous, and one of them leaves the key visible even after you delete it in the next commit.

Nobody has time to test everything. But there are four kinds of test that give you almost all of the benefit for very little effort, and several that get written constantly and are worth nothing.

The screen takes four seconds and the server is not under load. It is almost always the same problem, it has a name, and it is fixed by changing one line. How to spot it without special tools.

You deleted a commit, something broke and you do not know when, or you need to undo something you already pushed. Five commands almost nobody uses until they need them, explained with the exact situation that calls for each one.

It compiles, it passes the tests and it looks tidy. That does not mean it is right. Six concrete things worth checking, in the order the real problems show up.

A user reports an error from two hours ago and there is no way to find their requests among millions of lines. Three changes that turn your logs into something you can query, without switching tools.

Your assistant answers general questions well but knows nothing about your database. Giving it access means writing an ordinary program. What MCP is, how to write one, and the three security mistakes that keep repeating.

You touch the sensor and you are in, without typing anything. What exactly happens between the browser and your server, why your database stops storing secrets, and the three problems nobody mentions in the announcements.

Four changes to the Dockerfile that cut the size tenfold, speed up every build and remove almost all of the attack surface. With the numbers for each step and the complete Dockerfile.

Job queue, full-text search, cache, events and semantic search. Four services people tend to add to the stack that the database you already have can handle, with the exact SQL for each one.

In almost every language, incorrect code breaks and tells you. An incorrect SQL query runs, returns rows and nobody finds out. What to check before accepting one, and what changed in the databases themselves.

Both were dragging the same flaw around: the application did more work than it needed on every change. Angular 21 and React 19 fixed it, and the fixes are exactly inverse to each other.

What you actually pay for when using a language model to write code, why the bill almost never comes from the model you picked, and the practices that cut spending most without lowering the quality of the result.

From a platform that only ran on Windows to a cross-platform one with artificial intelligence support in its base library. The complete path, and the decision that defined each stage.