I Tried to Catch Coding Agents Cheating. They Mostly Didn't.
I built cheatguard, a linter that blocks the shortcuts coding agents take to make checks pass. Then I ran 798 agent runs to prove it works. Not one cheated.
Technical posts about software, AI tooling, and building in public.
I built cheatguard, a linter that blocks the shortcuts coding agents take to make checks pass. Then I ran 798 agent runs to prove it works. Not one cheated.
I'm releasing docmux, an open source document converter written in Rust. It runs in the browser through WebAssembly: LaTeX papers to Typst that compiles, Word with native equations, and a playground to try it with your own documents.
We built in two months a product that would have taken a year. When we reviewed it, a small change touched 80 files. The AI didn't write bad code: it made design decisions nobody reviewed.
A code review of a SaaS app a non-technical colleague vibe coded in 5 days. God tables, missing auth, dual systems, and the meta-lesson about AI-compressed timelines.
A security review of millionco/expect — a Claude Code plugin that runs AI-driven browser tests via Playwright. The skill is safe. The CLI decrypts every cookie in your browser.
The skills ecosystem is growing fast. A recent study found malicious payloads in 13% of published skills. Here's what that looks like in practice and what you can do about it.
Your AI coding agent installs packages to solve problems. It doesn't check if those packages are GPL, unlicensed, or pulled from a sketchy mirror. Here's how to protect your project.
Claude Code's instinct to parallelize with worktrees is powerful — until your project has Docker Compose, env secrets, and 5 services fighting for the same port.
Vibe coding works great solo as is. But what happens when your team grows exponentially?