tomsmithtld

the cognitive debt framing clicks for me. I've been using claude code on a laravel project and the thing that actually keeps velocity up isn't the AI getting better, it's me writing tests first. sounds boring but when the LLM generates code against an existing test suite it basically has to follow your architecture. you keep the mental model because you wrote the tests, and the AI keeps the codebase coherent because it has concrete constraints to work against. without that the codebase drifts into this weird state where technically everything works but you can't reason about any of it.

the shared MQTT credentials pattern is unfortunately super common in budget IoT. seen the exact same thing in smart plugs and air quality sensors. the frustrating part is per-device auth is not even hard to set up, mosquitto supports client certs and topic ACLs with minimal config. manufacturers skip it because per-device key provisioning adds a step to the assembly line and nobody wants to think about key management. so they hardcode one set of creds and hope nobody runs strings on the binary.

the "full" mode where agents critique each other seems more interesting than quick synthesis. curious whether you've seen cases where the debate produces something neither model would've suggested alone?

Playwright MCP is worth checking out too - similar idea but handles more browser contexts out of the box. Been using it for scraping and form automation.