LUMP Depot
It was the end of the day on a holiday weekend. I still had a lot of context left in my plan, and I thought: why not use some of it up before I go away. That is not a financial strategy. I just wanted to get my money's worth, ya know?
I asked Claude what it would take to make a coin pusher game in JavaScript using three.js. Typical vibe coding. Fun, low stakes, nothing riding on it. You know how this goes.
What would it take to make a coin pusher game in JavaScript using three.js?
One request. It used my entire five-hour context window on that single thread. The model went deep. I watched the meter run down and I did not stop it, because it all seemed well and good. It's Friday. I'm just poking at a toy.
Then the ticket landed
A critical issue came in. Something in the codebase needed investigating and fixing as soon as possible because customers were affected. Not theoretical. Not "when you get a minute." Now.
And I was out of context. Flat empty. The assistant that usually knows my repo, my last ten decisions, and the name of that weird helper function was suddenly unavailable unless I paid extra or waited.
So here's the new footgun, if we're keeping a list. You can burn your whole session on a single recreational prompt and leave yourself with two bad options when real work shows up: turn on more billing and spend something like twenty bucks on a problem you did not plan for, or sit there until the window resets in about five hours and hope the outage tolerates that.
That is a maintenance risk, not a vibe risk. It's the same reason I've been banging the drum about not getting too dependent on AI-assisted code and pure vibe coding for the stuff that actually pays the bills. The tool is incredible until the meter is empty at the worst possible moment.
The old way still worked
Luckily I wrote most of this code myself. I could go in, read the stack traces, grep the weird paths, and debug it the old-fashioned way. No context window required. It was slower and less glamorous, but it shipped.
The uncomfortable question is how long that stays true. If more of the system only ever lived in long chat threads and generated patches I skimmed, would I still have been able to do that under pressure? I'm not sure. That's the dependency curve nobody graphs on the same chart as "lines shipped per week."
tl;dr: treat your context budget like part of operational readiness. Fun experiments are fine. Draining the tank on a Friday lark before you might need the tank is shooting yourself in the Claude. I did it so you can picture it clearly. Don't save your leftover context for a rainy day in the cute sense. Save it for the rainy day in the pager sense.