dhung.dev

Everyone fed up with Opus 5 Claudish. Here is how to shut it up.

A decision framework for what to test, how to structure it across a backend.

I built a pipeline to compare air conditioners across two retailers. The unit that costs least at the register costs 9.5 million dong more over five years.

RTK, Caveman, Context-Mode, Lean-ctx, Headroom, Repomix, SigMap, Graphify, Distill, Codebase-Memory-MCP, what they actually do.

No training, 12 hours of meetings a week, public sprint shaming, and a probation period that guarantees silence. This is not high performance culture.

xAI lost all 11 co-founders, utilize 11% its GPUs, and IPO next month. So Musk rented Colossus to Anthropic.

Most CLAUDE.md advice is wrong. I experimented with what agents actually follow and ignore. Here's the guideline I use.

Prisma's flow is fast until you need a rollback, a CHECK constraint, or to not lock whole db. I flipped the workflow: SQL first, Prisma schema follow.

I went full agentic AI on a feature and lost all sense of what my codebase was doing. I posted about it, and hundreds of devs said they feel the same thing.

Deployed Fastify, PostgreSQL, and Redis on GCP's free-tier 1GB VM for a pre-revenue product. Here's the full playbook.