As workers at tech companies are sometimes wont to do — even in non-tech roles like mine — I had occasion last week to audit 1,500 Git repository logs and tally up recent code reviews and commits to help gauge activity levels. Back in the day, I would have had to check these manually one at a time or hunt down a developer to do it for me programmatically.
Now, though, it’s AI to the rescue.
Step one is figuring out the right tool for the job. There are several proprietary ones at our disposal, but since this is a pretty large dataset to comb through, the tech-ier tools (think: Claude Code or Codex) tend to have bigger context windows that can handle it.
So I fired up my desktop IDE (integrated developer environment), which is a friendlier wrapper for a CLI (command line interface) tool and has the right MCP (model context protocol) connections to access what I need via API (application programming interface). I’ve also given it access to a local file folder, where it could interact with my csv (comma-separated values) file of the 1,500 repo names.
Did I lose you at Git? Did you make it through the acronyms? Are we down the techie rabbit hole yet? Hold on, because we’re only warming up. (If you don’t want the full saga, skip down to The moral of the story.)
Hitting the wall. A lot.
Almost immediately, the tool — let’s call it CindyBot — hit a wall. The APIs weren’t working the way it expected, so it suggested doing a scrape in the repo browser instead. OK, but then it hit an authentication (login) wall. It advised opening a PowerShell terminal window (I have a Windows work laptop), reauthenticating, and all would be well. Nope. Same wall. CindyBot then suggested I use its full-on CLI version that runs in a virtual-machine terminal behind the authentication wall.
Hm. I’m not a fan of working in that terminal. I’m never really quite sure what it’s doing, and have to rely on the kindness of reference material to copy/paste commands. But if I must…
So I accessed the virtual terminal, logged in to CindyBot2, and repeated my prompt. It, too, started down the API route, hit the wall, and suggested the scrape. Now it just needed the list since it didn’t have access to my file. Alas, I couldn’t figure out how to upload it to the virtual machine. I asked CindyBot2 for the instructions, and it returned some sort of developer gibberish and implied strongly that this is a very basic task, and if you’re working in this environment, you should know WTF you’re doing. Awesome.
After a few rounds of terminal-shaming, CindyBot2 suggested just copy/pasting all 1,500 names into the chat, which I did. I gave it the go-ahead for the scrape and — d’oh! — it ran into the same authentication problem that wasn’t supposed to happen. It can’t physically authenticate for me, so it advised opening a separate terminal window to re-run those commands in case the session had expired, and then all would be well.
Small problem. CindyBot2 assumed I was using a typical dev setup to access the terminal. I was not. I was using the total-noob method that uses a browser and chewing gum. In this particular interface, there was no way to open a separate terminal window, which I pointed out. It said, in essence, I was lying. I said, in essence, oh no you did not!!! It begrudgingly admitted it was wrong and told me to log out of it, reauthenticate right there in the same window, and re-log in, because it would pick up where we left off. I dutifully did as instructed.
But CindyBot2 had punked me. It didn’t pick up where we left off because it doesn’t save sessions and didn’t remember the chat. Fine. I repeated the prompt and the copy/paste. It repeated the API attempt. It re-suggested the scrape. I gave it another go-ahead.
<sad trombone> Authentication wall.
CindyBot2’s next suggestion was asking a developer colleague to do the audit for me. Implied: Because clearly you’re an idiot.
It’s all in the context
Having gotten nowhere but furious, I tried CindyBot3, which is an environment meant for non-techies to use agents for the techie work. I didn’t go here first because it has a smaller context window. But if I must…
It tried once again to go down the unsuccessful API route, hit the wall, and suggested the scrape. It also couldn’t read my local folder, but it could access a separate Notion-like tool where I copy/pasted the list. Holding my breath, I gave CindyBot3 the go-ahead.
Success!!! It scanned a sample set of 5 logs and returned the data I needed!
Just to be on the safe side, I had it scan another 45 … which also worked! Now we’re cooking with gas!
It advised working in small batches because of the size, so I told it to do the next 50. That went well, too, and we started on 50 more. But while watching CindyBot3 think through this batch, I noticed a disturbing line flash past.
Based on the naming conventions of the remaining 30, I can do a pattern match and assume these will return zero code reviews.
<sad trombone> A full audit was too token-heavy, and so CindyBot3’s context management kicked in, and it decided to guess. The context window got me.
This wasn’t a deal-breaker. I asked it to create a blueprint of what it had done so I could start a new session without eating up context going down the full API -> wall -> scrape path again. It did, and then I ran into another blocker. I haven’t used this tool much (my work usually doesn’t need this much techie-ness) and didn’t realize it had no way to create a fresh session without exiting and re-creating a new agent environment. This is a several-minute process. But if I must…
And then the rote work began. New environment -> read blueprint -> run maybe 3 batches of 50 each -> “I shall pattern match” -> context window full. Each cycle took about 20 minutes apiece, and all told, I spent 6 hours that day wrestling with the CindyBots and going through all the batches.
The moral of the story
I did get the data I needed, I was able to do it without developer help (which wouldn’t have been possible a few months ago), and it did save me at least a day’s worth of work.
However, I expect the cumulative anger/cortisol surge from the tool churn and tech-shaming has shortened my life by perhaps an equal amount of time.
I’m fortunate to have access to a wide set of AI tools at work. Almost too many, in fact, and it’s easy to get lost in the ecosystem, especially if you’re new to the whole AI thing. I’m also fortunate that the team I’m on has embraced the tools and freely knowledge-shares what does and doesn’t work. For the most part, though, we’re expected to experiment and learn as we go, which, frankly, often feels like dumping someone in the NASA control room with a calculus textbook and a keyboard and asking them to calculate orbital velocity.
That can be addressed with more structured learning paths. A much bigger issue — and one many enterprise environments share — is that a lot of this technical complexity is because all the AI tools and all the other tools and platforms (Outlook, Slack, etc.) and all the various datasets don’t talk to each other very well. There are good reasons why, security being a big one. Another is that they were straight-up never built for this sort of thing, and need a lot of retooling to set up the right connections.
But if you’re told to automate your work and you’re not a dev and you keep running into an authentication wall or a lack of API access or some other barrier that requires in-depth knowledge of back-end systems, well… <sad trombone>.
AI fluency is one thing. Dev fluency is another. And having to learn both on the fly is daunting for many. Right now, though, the folks who are unlocking the most power with AI are the ones who are willing or already comfortable with diving into dev stuff. It’s difficult to bridge the technical complexities any other way.
It won’t be like this forever, though. Anyone remember Internet Relay Chat? MS-DOS? They’ve long since been replaced by systems that don’t require degrees in computer science to operate. Here in the present day, non-techie alternatives are catching up, too, and they’re catching up quickly.
So take that, CindyBot.
All opinions here are my own. All text is my own, too, including the em dashes. I welcome constructive comments and discussion on LinkedIn and Bluesky.