Agentic AI can give you superpowers

So I have been able to spend my usual few weeks of summer holiday both enjoying the sunshine and swimming pool as well as using my ‘build something while I can’ time. This is where I have a few hours of peace and quiet before the rest of the family wakes up and we get started for the day, and I dont have real work to do.

Last year I scratched an itch and built a team management app - to replace the many excel ‘people planning model’ spreadsheets I think all CTOs have - which was more about experimenting with Cloudflare workers / Hono / HTMX / serverless databases and seeing what the art of the possible was. I got quite far, but not far enough for it to replace Excel.

This year of course the world has really changed, so while we have been pushing hard on using AI both in our products and in our engineering teams at my day job as CTO @ Infinitas, I haven’t personally been able to get very hands on. So this year I put myself fully into the hands of Claude Code (https://claude.ai/) with the Max Plan and decided I would push my old app as far as I could in as short a time possible.

I learned a lot:

  • You must invest time in describing and documenting your code and app - adding it to claude.md files scattered into the different areas of the app. Tell it how it works, what the intent is, what good patterns are, what not to do. And of course you can use Claude to write the first version of this for you. This even worked on what was quite a ‘legacy’ code base . Where I first used claude to migrate from Hono / HTMX to Hono / Vite / Typescript / React (while maintaining the same Neon postgres database) - and it almost ‘just worked’.
  • Separate the ‘Investigate / Analyse’ step from the execution step. Use Opus for the more complex analysis, investigating and writing its output into a markdown file that you can review and edit. Then switch to Sonnet, and ask it to execute on the plan exactly.
  • You need to pay attention if it is doing anything complex, press esc if it is getting confused, add a nudge in the right direction and it will help it get unstuck.
  • Managing your context is critical, clear your session via /clear after each activity, rely on the documentation you created earlier and of course craft a smart prompt to get going.
  • Ask it to use agents to parallelise repetitive well documented work.
  • Finally, keep your prompts clear and crisp - don’t ask nicely, be concise, directive.

So did this make me more productive? I think immeasurably - for an experienced engineer who knows what they want it gives you super powers. I haven’t been able to write this much code to this level of complexity of product in this short a time (~10-15 hrs total) in my entire career, even when I was 21 years old coding every hour I could in my first startup!

You can try it here: https://www.fintune.app. It’s live but not a real product (yet?). Hopefully you will excuse the AI written copy :D

Fintune Team Visualization Dashboard

The team visualization dashboard I built using Claude Code - showing how teams are organized across different departments and skill areas.

How does this compare to your experience with AI coding tools?