2026-04-07
Hey friends,
Did you hear about the SQL query that walked into a bar, walked up to two tables, and asked…
"Can I join you?!"
There are so many different emerging approaches and workflows for leveraging AI — it's pretty crazy, and frankly overwhelming. Some approaches are incredibly complex (like building your own multi-agent ecosystem a la Gastown), and some techniques are a little more approachable...
One relatively simple approach I've had some success with recently is spec-driven development as Addy Osmani outlines it. The basic premise is:
CLAUDE.md for Claude Code) that sets some boundaries in things like 'Always Do', 'Ask Before', and 'Never Do'. One of the 'Always' enforces writing and running tests.SPEC.md) BEFORE you let it loose.The multi-step plan helps keep the agent focused on a single vision. I've found fewer instances of the agent wandering off into the woods chasing something irrelevant. It also helps when you work in chunks, as you don't have to work to preserve ephemeral context inside the agent — it's all written down as a plan!
The tests also help the agent check itself, allowing it to identify errors early and fix them before they become more difficult to resolve.
Web Stuff
Other Stuff
👽 Weird, 📺 Watch, and 🎮 Play
Don't miss what's next. Subscribe to Net Noodlings with Nathan: