I Wrote Up My Whole Claude Code Workflow

People keep asking how I actually get real software out of a coding agent — a compiled language, a production SaaS, a strategy game, solo — instead of the plausible-looking mess an agent hands you when you point it at a repo and hope.

So I wrote the whole thing down. Not a hot take, not a thread — the actual method, start to finish. It's a long read, and it lives over on my personal site rather than here, because it's less "game dev log" and more "here's how I engineer."

Read the full guide → How I Drive Claude Code: A Practical Guide to Agentic Engineering

The short version, so you know if it's worth your click:

• Treat the agent as a capable but fallible engineer, not an oracle. Everything else follows from taking that one line seriously.
• Run everything through a named document pipeline. IDEA.md → DESIGN.md → PLAN.md → BUG.md, each captured in its own conversation and its own specified format — so both Claude Code and my own tooling can read the state of every project at a glance.
• Break the work into phases with real verification gates. A phase isn't done because the agent says so — it's done when something other than the agent confirms it.
• Review continuously. Fast agents don't produce bad code; they produce plausible code that quietly rots. Pay the debt as you go.
• When it gets stuck, change tactics, not prompt volume. Switch personas, stage a cross-model debate, feed in your own hunch.

There's also the part nobody talks about: how to stop. When a session runs out of useful context mid-task, I have the agent write a RESUME.md — a save point a fresh session reads to pick up cold, no re-explaining. It's how one task survives across three sessions without losing the plot.

Go read it. Then come tell me what your workflow does that mine doesn't.