← All writing
ON BUILDING · JUN 2026

Build, find the gaps, fix them, repeat

Risograph illustration: a person adding a block to a modular structure with gaps, a circular arrow showing the build-and-improve loop.

Anthropic ships new features almost every day. I run my consulting practice on an AI operating system I built in April that functions in Claude Code, and I don't have time to track release notes, follow every newsletter, and work out what applies to my setup. Nobody does.

Yesterday while jogging, I heard a suggestion for how to keep your system up to date, on The Growth Podcast. The host, Aakash Gupta, was talking with Hannah Stulberg from DoorDash about her team's Claude setup. He described a prompt he runs on his own system: research everything Anthropic released in the last 90 days, look up what the top practitioners are recommending, then compare it all to my system and tell me how to 10x it.

When I got home from my jog, I ran this prompt.

After gathering all of the background data and analyzing it against my system, it came back with 12 concrete ways to make it better, ranked by effort. Most of the suggestions were based on features that didn't exist when I built the system.

We spent part of the day working through the list. My system now does things it couldn't do in April:

  • Every scheduled routine reports when it runs. If one fails, I get an email, using a monitoring pattern Claude found in its research.
  • Every initiative I'm working on shows a green, yellow, or red freshness flag, built on the official Obsidian skills the Obsidian team released specifically for Claude.
  • When something breaks, Claude writes the lesson into a log, and recurring lessons become permanent rules.
  • On the first of every month, the system re-runs the same research prompt itself and emails me what to improve next, based on what has changed, tailored to my setup.

If you've built an AI-driven operating system you rely on, this is a way to make sure it keeps improving and adding more value to your work over time.

The prompt

Adapted from Aakash Gupta on The Growth Podcast. Run it in Claude with web search on:

1) Research everything Anthropic shipped in the last 90 days (search the official release notes and changelogs, don't answer from memory) and build a dated list. 2) Research what the most followed Claude/AI practitioners have recommended in the same period. 3) Read my setup files. 4) Compare and give me a ranked list of improvements with the effort each would take, plus what to deliberately ignore.

One caution: anything your AI reads from the open web is untrusted input. Nothing in my system gets merged, archived, or deleted without my sign-off.

The full write-up, with every prompt, is in my newsletter.

AI on Purpose is where I share the whole process, step by step, so you can run it on your own setup.

Subscribe to AI on Purpose