Doomprompting
Is “doomprompting” worse than doomscrolling?
Doomprompting is when you keep prompting an LLM over and over, hoping you’ll strike gold. You tweak a word, hit send, get something close-ish, hate it, send it again. You’re doing a bad job at explaining what you want and hoping the model guesses for you.
The thing is, you usually do have something in mind, but you just have writer’s block, or can’t express it well, so you try to outsource your thinking. You turn the LLM into a third-party thinker and hope it returns something better than the mush in your head.
But you can’t make something out of nothing.
I’ll call it the law of conservation of tokens or something: you can’t pull more out than you put in. If you never specified the thing, no amount of regenerating is going to generate it.
This bit me recently while trying to name a new product. The domain I wanted was taken, so I asked Claude for alternatives. It gave me a list of similar, equally bad and weird names. I tried rolling again, this time with some more context. Obviously, I got more weird names. I don’t blame Claude because it really had no idea what I wanted, so it was no better than your phone’s autocorrect, just guessing.
I notice this happens way less when I’m coding though. Not because code is easier, but because of how I’ve set things up: my AGENTS.md tells the agent to grill me before it writes anything, to interview me until we actually share an understanding of the goal. The forced back-and-forth drags the spec out of my head.
It’s not unusual for it to ask me 30+ questions for a simple stupid game I hope to make.
With naming a product, writing a sentence, picking a direction: there’s no interviewer there. It’s just me and a blank box, and if I don’t know what I want, the loop has nothing to converge on.
Like sure, I could have it grill me, but it often makes less sense. I have no idea if I want my product name to be more “calm and informative” than “energetic and playful”.
Even when I do know what I want, I still get stuck if I can’t put it into words. Sometimes switching from Claude Code to Codex helps, but not always.
When the task is spec-able, asking the LLM to “interview me” before implementing helps a lot. It’s gonna ask you a ton of questions, but it’s better to spend the time planning and building the right thing than building the wrong thing and attempting to fix it. The catch is that this only works when the answers exist in your head somewhere. For the fuzzy stuff, no interviewer can pull out a preference you don’t have yet.
Back to the question: is doomprompting worse than doomscrolling?
I think it’s sneakier. Doomscrolling announces itself more. You know you’re wasting time, the guilt shows up eventually after you realize you’ve been sitting on the toilet for 45 minutes. Doomprompting hides inside real work. You’re typing, you’re getting output that looks like progress, so you likely don’t realize what’s up. You can burn an hour this way and feel busy the whole time.
But it’s not all bad, kind of like video games versus reels. Both can eat your night, but at least a game stimulates your brain a little, and maybe you’re playing with friends. Doomprompting is similar: even when you spiral, you might still walk away with something, a shipped feature, or a pattern you picked up. Doomscrolling leaves you with nothing.
When I catch myself in the loop, the fix is never another prompt. I really need to get my mind off things, and maybe go for a walk. This worked long before LLMs showed up, it’s just easy to forget now that there’s always a slot machine ready for another spin.
Because in the end, you can’t make something out of nothing. The LLM will happily try to fill that gap, and it’ll hand you something that looks good enough every time, that’s the trap. The work it’s pretending to do for you is the deciding, and that was always yours.