About three months ago, I started an experiment: build a new product while asking the team to lean hard into coding assistants; we picked Claude Code, but these lessons are generally applicable to Codex, Cursor, etc.
For me, it began as curiosity: if generative AI is here to stay, what does the software organization of the future actually look like?
Here’s what I’ve learned so far …
- START SMALL. Find your “lighthouse teams.” Give them access, clear goals, and support. Let them experiment, stumble, and share what works. Coaching matters more than policy early on.
- INTEGRATE, DON’T REPLACE. GenAI’s power is in how it fits into your workflow, not in rewriting it. Capture your standards, patterns, and review criteria as “instructional context” so the assistants stay grounded in how your org builds software. Revisit these regularly as part of your continuous improvement loop.
- ENFORCE TRACEABILITY. Yes, every compliance framework loves it, but it’s also just smart engineering. Tie code to requirements through prompts or PR metadata. It keeps assistants focused and makes the code instantly auditable (your CISO will thank you).
- DOUBLE DOWN ON CODE REVIEWS. Assistants can check syntax and structure, but only humans can confirm intent. Frequent peer reviews keep collective understanding alive and catch subtle misfires before they calcify into technical debt.
- KILL TECH DEBT EARLY. Agents learn from the repo’s context, and so bad code spreads fast. Use them for refactors, not hacks. If you must take on debt, document it like you’re writing an instruction manual for the next agent.
- DOCUMENT EVERYTHING (WELL). Requirements, architecture, test strategy; all of it. Create a
/contextor/docsdirectory, and keep the directory lean and accurate. Sloppy docs poison the agent’s knowledge just as fast as bad code. - COMMUNICATE LIKE IT’S DAY 1. When an engineer can ship a week’s worth of work in a day, alignment becomes a daily exercise. Make conversations about intent, goals, and impact the center of your stand-ups.
Leaning into AI-assisted development isn’t about speed, it’s about discipline at scale.
The teams that win won’t be the ones that code fastest, but the ones that learn fastest from how they code.
Originally shared on LinkedIn, November 1, 2025.