📄 claude-cowork-oblivion.doc
_

← Back to Things I’ve written about

Claude Cowork is Elder Scrolls: Oblivion all over again (in a bad way)

In 2006, I bought my first PC for gaming. Two graphics cards (two!) using then-fancy AMD Crossfire technology. I did not, and still do not, know much about building a PC. I also bought a copy of Elder Scrolls: Oblivion, because I’d heard it required a high-spec machine to play.

What followed was a couple of weeks of tweaking settings, fiddling with sliders, installing mods, uninstalling mods, learning what things like ‘anti-aliasing’ and ‘draw distance’ meant, and so on. It was great. And of course, there was a ton of character customization to do as well.

I also spent about 5 hours in total playing the actual game, but eventually moved on because it just didn’t grab me. It looked incredible but was kind of… underwhelming? I just wasn’t having fun, which is pretty much the whole point of games. Seems it was critically acclaimed though, so I was probably just wrong here.

Flash forward to 2026 and I’m having a blast with Claude Cowork. I’ve got my set up just right, tasks firing and the best LLMs the planet has to offer working overnight to deliver incredible business insights to my inbox every morning. But as the models have apparently improved, I’ve noticed the list of what I find Cowork good for hasn’t really changed, and neither has the list of tasks I send its way.

Good

So basically - the dumb bits of middle management. Genuinely useful! Not exactly life-changing.

Not good

So basically - the fun/terrifying AGI bits.

What’s more, I’m spending an increasing percentage (25%?) of my time with Cowork feeding it context - about me, PostHog, the teams and people I work with. Because of our transparent handbook and LLM-friendly repos stuffed with markdown files, I suspect this is actually a significantly easier job here than at most companies. Every few weeks Anthropic does something weird and tasks suddenly don’t work as they used to, or the LLM in all its probabilistic glory starts behaving in an unexpected way despite apparently nothing changing.

Oh no - I’m back ‘playing’ Oblivion again! I’m in tweaking mode, getting everything set up nicely, tuned, humming along. But I’m not actually playing the game, i.e. getting real, usable output a real chunk of the time, I’m just playing with settings. It’s the illusion of fun, except this time it’s the illusion of actually getting anything done.

I’ve accidentally drifted into the bad part of context engineering - context theater. For the uninitiated, context engineering is where you, human, manage the information a model gets in order to complete a specific task better than the model could otherwise do alone. Context theater is a thing I just made up, where you spend most of your time context engineering with nothing to show for it at the end.

So what’s the lesson here? Something something opinionated? General purpose business harnesses will never work? If you’re spending more time perfecting your setup, maybe you just need to find a different game?

(Side note - I played Portal 2 on the Switch last year for the first time, and had a great time 15 years late. Didn’t spend any time in settings, because there was nothing to adjust and that’s not really what the Switch is all about. Maybe this is the way.)