June 5, 2026
Bug drama, but make it housework
My Agent Skill for Test-Driven Development
One coder says AI needs strict homework rules—commenters are calling it genius, outdated, and total snake oil
TLDR: A developer says AI writes much better bug-checking code when you force it to follow a strict step-by-step routine. Commenters were wildly split: some called that smart coaching, others said it’s overcomplicated hype—or that writing tests at all may be losing value in the AI era.
A programmer shared a playbook for getting AI coding helpers to stop churning out what he basically calls useless, fake-looking tests—the little automatic checks meant to catch bugs before they spread. His fix is a step-by-step coaching method: first decide what the software should do, then turn that into tests, then write just enough code to pass them. He even has a second AI review the first AI’s tests, plus a memorable house-cleaning metaphor for tidying messy code before adding more. In short: the bots need supervision.
But the real fireworks were in the comments, where the crowd split into camps almost instantly. One side rolled its eyes and yelled “snake oil”, arguing that fancy custom “skills” are overhyped and you can get similar results by simply telling the AI what you want. Another group said the advice may already be aging fast, because these tools improve so quickly that a blog post without a clear date can feel old on arrival. And then came the true chaos agents: people declaring that test-driven development—writing tests before code—might be one of the worst ideas in the age of super-fast AI coding, because why pay for all those extra checks if the machine already writes mostly bug-free code?
Meanwhile, the funniest mini-debate wasn’t even about code. It was about whether telling an AI “don’t make dinner in a dirty kitchen” is clever coaching or just adorable nonsense. Somehow, that line may have stolen the show.
Key Points
- •The article says AI agents are currently poor at writing useful tests without explicit guidance.
- •The author bases a personal agent TDD skill on Kent Beck’s Canon TDD and hosts the living version on GitHub.
- •The article introduces a three-step TDD loop called specify-encode-fulfill: define specifications, encode them as automated tests, and write code to satisfy them.
- •The described process emphasizes minimal code changes to satisfy the current failing test, with refactoring separated from behavior changes.
- •The author supplements the TDD skill with separate Test Design Review and Software Design Review skills, and says Claude often responds to a cleanup-first instruction when tests are hard to write.