Relearning one of the oldest cliches

I re-learned one of the oldest lessons. That when you use a hammer, everything looks like a nail. But for the AI era.

I became one of those people. I run a set of overnight jobs on a Mac mini. They watch my code bases, fix vulns, and help me manage my life. Then email me a digest of the day's tech news (and a bunch of other stuff) before I wake up. Until May, an AI model made most of the decisions and did most of the work in those jobs.

Then Anthropic announced that automated usage like mine would move onto a metered credit. So I measured what I'd been using. Claude estimated it would have cost about $300 a month to do all the things I was using agents for. Wild.

There was a second problem too. Models make stuff up and fail in super weird ways. Programs and deterministic outputs don't. A model once invented some Hacker News posts, complete with a convincing URL that returned a 404. It rewrote a repo with a totally different name. It skipped some of the top stories of the day because it didn't think they were important enough, despite the prompt saying otherwise.

I needed to do something. Just generally. So I started out by pinning everything to Haiku and Sonnet. That got the bill down, but it was still spending (theoretical) money I didn't want to.

I pivoted to a tiny local model. It did a decent job. But it was too weak to trust. So I started asking Claude how much could be programmatic. Picking the day's big stories turned into a ranking function. Deduplication became a hash. When I finished, code was doing 90% of each job and the model was just doing the words. Then I moved the words to a template.

One of my scans now runs in 9 seconds. It used to take 7 minutes and sometimes timed out. The digest reads the same as when the expensive model wrote it, and the running cost is zero. And, it can't invent random stuff.

I still use Claude every day. It planned and wrote every one of these pipelines. Its job moved from running my automation to building it, and freeing up usage for helping me in my other projects.

It wasn't until I realised that when your tool is a hammer every problem looks like a nail that I saw I could just move away from this setup. And since then I've pushed so many of my jobs away from AI and towards programmatic approaches, leaving agents to do the things that just can't be done programmatically.