AI is one-shotting novel tasks

Coding agents have been helpful in the past. But, they always seemed to need guidance and control. They hallucinated and made mistakes that would get junior developers fired.

Recently the new models seem to be different beast. I had a complex feature that required using ancient technology so handed it over to Claude Code.

The problem

I had data in a database that needed to be enriched from another datasource.

That other datasource came in the form of a ZIP, which contained a single database export of a database system that was deprecated in 2003.

Deprecated in 2003. I was still in elementary school then.

No schema descriptions, no documentation, just a vague description what was in there valuable.

I gave it to Claude Code to analyze, told it the promise of the data and was curious to see if he could leverage it to enrich our data.

Claude absolutely one-shotted it

Claude spent little less than 12 minutes thinking and working. It prompted me for CLI tools needed, and whether he could install & use them.

The output was a Dockerfile that would take the URL, download the zip, export it, join it with our data and write it to our database in a single call. It had seen we use GCP in this project. This allowed me to set up a nightly Cloud Run job, since the ZIP file was at a static URL daily updated.

I reviewed everything; small touch ups were needed but nothing fundamental.

12 minutes... This would have taken me hours to set up myself.

Best of all; three other claude instances were doing similarly complex tasks in parallel with similar spectacular results.

We're accelerating

Less than a year ago I wrote about disappointing results using Lovable. Today I can give it complex novel problems and it one-shots practical and tailored solutions.

The next years are going to be crazy. We're in for a ride.