Most of my coding happens in a terminal chat. The agent does the typing and the file-digging; I hold the architecture, the constraints, and the bar for what ships. The day is explain, check, correct, explain again.
People picture this as delegation: write a good prompt, go make coffee, merge. That's not how it works. Agents drift-they'll pick a shortcut you'd never pick, or build the right feature on the wrong layer. I keep the loop tight on purpose: small prompts, constant check-ins, corrections made while the work is two minutes old instead of buried in a five-hundred-line diff. It's pairing with something very fast and a little too confident, and the steering is where the engineering lives.
The flow rewards the discipline I already work with. Design comes before the prompt-when the architecture is clear in my head, the agent builds it straight; when it isn't, no prompt saves it. Sharp briefs, because vague asks come back vague. And nothing merges on the agent's word alone-type checks, tests, a real run of the thing. Trust the verification, never the summary. The tools will keep changing. The standards are what make them work.
Keep reading
- Balancing AI Help With Keeping My Own Edge
Relying too much on AI can chip away at your ability to focus and reason deeply. The real trick is using AI without losing your edge.
- Sub Agents Changed Architecture
Monolithic agents are brittle. Hierarchical delegation might be how complex AI systems should work.
- Agent Guardrails in Production
Letting agents handle real money requires serious safety measures. The industry has learned some lessons the hard way.