There are moments when working with Claude Code or Openai Codex feels absolutely magical. Features get implemented faster than I could ever hope to accomplish. I see my project take shape with just the occasional prodding of my AI worker bees.

I envision, we plan, they implement. I dream, we strategize, they toil.

2FA? BAM! Database migration? BAM! Scraping of a relevant API? BAM! Building a user interface? Custom Django admin interface? A change to the data model? BAM BAM BAM. I reload the page and bask in the feeling of superhuman speed. It's like seeing a metropolis being built before my eyes in hyperlapse, and it feels incredible!

... and then I notice something in Claude's endless output. An innocuous function named update_constituencies. But wait, I think - didn't we already have another part of the code updating constituencies? I ask Claude about it.

"Hey Claude, are we populating Constituencies here from the API?"

"Yup, this fetches the Constituencies we need to create our Representative instances."

"Okay, but aren't we already creating these in update_constituencies over in TheOtherModule?"

"Let me look into that for you.... You're absolutely right update_constituencies does the same thing as update_constituencies, they are redundant!"

Cool, I think, let's refactor. We brainstorm, we plan.

Claude is off to the races and churns out commit after beautiful commit. With commit messages that would make any reviewer cry joys of tears - succinct and clear - it forges ahead with unerring confidence until the PR is ready.

I take a look. Nothing is fixed.

"You sure this is fixed? I still see two functions that do basically the same thing"

"You're absolutely right! Let me fix that real quick."

It's infuriating. My personal AI hype train comes to a crashing halt. I start sorting through the soup my enthusiastic AI chefs have cooked. It's not pretty, and there is oh so much of it. I'm mad at Claude, I am mad at myself for believing in a beautiful future where I tell my computer what I want and it understands.

I go to the kitchen, make myself a coffee and look out the window.

The future is not here. Rome was not built in a day after all.

I sigh, walk back to my computer and CTRL+C twice. Goodnight Claude. I put on my favorite playlist, open up my editor and start from scratch. This is way more fun anyway.

For what it's worth, I use Claude with superpowers and rigorously follow the plan, refine, implement loop. Still, it's not enough. Where I find Claude extremely useful is as a tutor when learning something new, or a sounding board.