I have had access to claude at work now for maybe two weeks now. Incidentally, I had already been using open code at home for a few weeks now at this point, so I’m wary of slot machining. I notice myself leaning into lift coding. So in data science speak, lift is the benefit above a baseline. So this can be using claude to rewrite a small piece of code based on a colleague’s PR feedback almost on the spot as opposed to perhaps waiting for some down time when I might get around to it. And I was iterating on a function, replacing it with a new one, and I found myself using claude to update the unit tests. That was really nice. Maybe I would have put off writing the unit test until later.

Lift tasks I am observing in myself, are often mini side quests I would have never otherwise had the spare time for. A week ago, my company’s python package proxy, which blocks downloads loosely based on new CVEs, blocked a deploy of mine as per usual with some fresh 403. But this time, instead of rummaging around in the effectively broken UI, which often shows lagging 🟢 clean packages that disagree with the 🛑 403 when downloading, I asked claude to build a table of a package’s version specs under CVE sorting by date. I did find this helpful for choosing a package version the first time I used it. However, because our package proxy really is trash, other CVE-free version choices I make using official CVE lists unfortunately don’t agree with the 403s I end up with. So then I ended up with trial and error and fighting the proxy web gui again.

Side note, in an ideal world, a package proxy UI and API should agree with each other and with opencve [1], nist nvd [2] and osv [3].

I also had a self review task due, where I ended up creating a script for pulling in my code pull requests I could use as the work history and a web ui I could use to help me more easily filter and annotate the more relevant work. The self review stuff can be a bit of a slog and a kind ui can be the spoonful of sugar to make the medicine go down.

Another kind of crazy one off I used claude on was related to being in a large company and recently joining a new large project with lots of people I don’t know. In the past I had made several attempts to learn enough about the Microsoft Graph API to just mess around with Microsoft Teams for various reasons, but the learning curve was too high for the pockets of time I might stumble upon here and there. But the other day, I used claude to create an org explorer from a list of emails I dump from an email meeting invite. I first used it to just validate whether my az login was good enough to even look myself up in the Graph API and when that worked, I asked claude to extend this to pull data build out a nice graph, so I could sort of understand who the people were in my meeting. I would have probably never found the time to do this otherwise. So that was definitely a nice lift.

References

  1. https://app.opencve.io/cve/
  2. https://nvd.nist.gov/
  3. https://osv.dev/list
  4. Staff Engineers Path, Tanya Reilly