Prompting isn't the interesting skill with coding agents

The more I use coding agents, the less I think prompting is the interesting skill.

Prompting still matters, but I’ve found that context, decomposition, architecture, validation and judgement matter more.

One thing that has worked well for me is preserving useful context between sessions. At the end of each session, I ask Claude to capture the key decisions, assumptions and learnings in a short project note or README, and retain the relevant context for future work. Over time, I’ve found that I spend less time re-explaining the data, coding style and so on, and Claude becomes much more useful within that project.

Decomposition has become equally important for me. If I’m building a dbt model or doing a larger analysis, I rarely ask Claude to solve the whole thing in one go. I ask it to break the work into phases, complete one phase at a time, and validate the data before moving on. That has become a critical part of my workflow.