I'm coding again and it's because I stopped coding

I run a data warehousing and GenAI team at AWS. As a manager, I haven't written production code day to day in a while. My job is strategy, prioritization, and high level architecture. I still think and talk technically, I just stopped writing code day to day.

Then I started using Claude Code and something amazing happened. I built a complete application over quick vibe coding sessions over the course of five days. It's a script that scans AWS accounts, generates architecture diagrams, and runs Well-Architected audits. 3,600 lines of application code, 108 tests, CDK deployment stack.

Using Claude I was able to operate at a high level of abstraction - giving Claude direction on which library to use, but not needing to dive in and learn it myself. I decided what the architecture should be and described it.

Abstraction

I think in terms of systems. What should this component do, what should it talk to, how should the data flow. That hasn't changed. What changed is I no longer have to also be the person who figures out the D2 syntax for a node with a nested shape, or why boto3 is returning a KeyError on a field that should exist, or how to mock a Bedrock API call in pytest.

I could just say "the diagram generator should use a planner/generator/judge loop, where the judge evaluates the rendered PNG." Claude built it. I didn't have to read the bedrock docs to figure out how to chain the LLM calls, I didn't have to tell Claude that we should stop optimizing after a certain number of iterations.

It's important to call out that I made this project as a side-project. If this were production code with real business impact and security implications, I'd be a bit more careful at each of these steps.

When the Graphviz diagrams came out flat and unreadable, I didn't troubleshoot Graphviz. I said "this layout approach isn't working, research alternatives to Graphviz that handle hierarchical AWS architecture diagrams better." Claude came back with D2, explained why, and I said go.

The part I didn't expect

I'm having fun. Like, it's actually fun to be able to build without worrying about lower level details. What I imagine it's like to go from writing assembly code to using an intuitive language like Python. Being able to build so quickly, go from idea to implementation is fun. And really addicting. I've never been one to stay up late working, but lately I've been doing exactly that.

I built more in these five days than I would have in a month doing it the old way. And I'm not exaggerating. I would have bailed on the diagram generation entirely. I would not have learned D2. I would have shipped the ugly Graphviz output and moved on. Instead I have three AI agents collaborating on layout quality because saying "build a judge agent that evaluates the rendered diagram" is easy when you don't have to be the one writing the Bedrock API integration.

It's not autopilot

I want to be clear about this. Claude didn't architect this tool. I did. Claude doesn't know the problem space, Claude didn't know what would be useful to an end-user and which features to prioritize. Claude also didn't always know when to give up on a given path and pivot, or when to introduce new ideas entirely. Maybe the new iteration of models coming soon will, but for now I'm still comfortable that my job is necessary.

I'm coding again because the bottleneck moved. It used to be implementation. Now it's ideas. And I have plenty of those.

The project is viewable here: github.com/blowenstein0/aws-cartographer