Quick version for the skimmers: I spent a month letting ChatGPT decide what I played, using it as a personal game sommelier for someone with a 90-game backlog and a five-hour week. The verdict is a genuine split decision. As a librarian for your existing backlog, it’s shockingly good, and I’m keeping it in the rotation. As a taste-maker discovering things for you, it’s a Reddit consensus machine wearing a trench coat. And once, with full confidence, it recommended a game mode that does not exist. Details, prompts, and the system I kept are all below.
The experiment started the way most of my experiments start: with me doing something dumb. It was 9:45 PM, I had maybe an hour, and I spent 25 minutes of it scrolling my Steam library like it was a menu in a language I don’t read. Ninety unplayed games, and somehow nothing to play. Sound familiar?
Decision fatigue is the silent killer of adult gaming time. So I outsourced the decision.
What Were the Rules?
Simple ones, so the experiment meant something.
For one month, ChatGPT picked what I played. I fed it my backlog list, my constraints (about five hours a week, evenings only, a preference for finishing things over sampling them), and my general taste. Before each session I’d tell it my mood and available time, and it chose. I committed to actually playing the pick, no vetoes, unless a pick was literally unplayable.
I wasn’t testing whether an AI knows games. I was testing whether removing the choosing from my evenings would give me more actual playing. Spoiler: that part worked almost embarrassingly well.
Where Did It Shine?
Three places, and they’re all versions of the same trick.
Mood-matching my own library. This is the killer feature. “I have 50 minutes, I’m fried, nothing that punishes mistakes” reliably produced the right answer from my own shelf, stuff like Balatro, a Hades run, a quiet hour of Stardew Valley. None of these picks were revelations. That’s the point. I owned the right game all along; I just couldn’t see it through the fatigue. The AI’s advantage isn’t taste. It’s that it doesn’t get decision fatigue at 9:45 PM.
The regret-ranking exercise. Early on I used a prompt I borrowed and adapted from a Tom’s Guide experiment: “If I only finish 10 more games from this backlog over the next year, which would I regret missing most? Rank them and say why.” The output was uncomfortably clarifying. It put Clair Obscur: Expedition 33 at the top (correctly, as it turns out, and I say that as someone who cried at my desk), and it quietly demoted a bunch of games I’d been keeping out of obligation rather than desire. Seeing your backlog sorted by projected regret instead of purchase date changes how you see the whole pile.
Session planning. Given “I have Saturday afternoon and two weeknights this week,” it built a genuinely sensible split: story progress on the long block, systems-y games on the short ones. It’s the kind of light structure I’d never bother making myself, and my finish rate for the month went up because of it.
Where Did It Fall on Its Face?
Now the fun part.
The hallucination. Midway through the month, it recommended a co-op mode for a game in my library, described how the mode worked, and suggested my wife and I try it. That mode does not exist. Never has. When I called it out, it apologized like a waiter who brought the wrong table’s food and then recommended a real feature instead. Nothing about this surprised anyone who uses these tools, but let it be said plainly: verify anything specific an AI tells you about a game’s features, especially before you buy. Confidence is not accuracy.
The taste flattening. When I asked for new games beyond my backlog, every list converged on the same 15 titles you’d get from skimming Reddit’s greatest hits. Solid recommendations, sure. But “what should a busy adult who liked Hades play” produced the exact list you already know if you read sites like this one. Not once in a month did it surface a genuinely obscure gem that felt found. AI recommendations are consensus with good manners, and consensus is where the interesting stuff isn’t.
The energy blindness. It can’t see you. It took my word every time I said I was “kind of tired,” and words undersell a Wednesday. A friend who games would’ve looked at my face and said “dude, just play the poker game.” The AI needed me to self-report accurately, which, at 10 PM, adults do not.
The Prompts That Earned a Permanent Spot
If you try this, skip the month and steal the system. Four prompts did all the real work:
The regret ranker: “Here’s my backlog: [list]. If I only finish 10 of these in the next year, rank which ones I’d most regret missing, and be honest about which ones I should delete without guilt.” Run this once. The delete permission is worth more than the ranking.
The 9:45 PM prompt: “I have [X] minutes, my energy is [honest number]/10, and I want to feel [relaxed/accomplished/absorbed]. From my library, pick ONE thing and tell me why. No lists, one pick.” The “no lists, one pick” clause matters. Lists reopen the decision you were trying to close.
The week planner: “My gaming windows this week are [blocks]. Assign my current games to them so long-form story stuff gets the big block and short-session games get weeknights.”
The fact-checker reflex, which isn’t a prompt so much as a policy: anything specific it claims about a game (a mode, a length, a DLC, a price), search before you act on it. Thirty seconds. Every time.
What Would I Do Differently Next Time?
Two upgrades, both about giving the machine better inputs.
First, I’d export real data instead of typing a messy list from memory. If you track your games in Backloggd or one of its rivals (we compared the big three in Backloggd vs GG vs Savepoint), your ratings and play history are sitting right there, and pasting that into the conversation gives the AI actual evidence of your taste instead of your flattering self-description. My self-reported “I love challenging games” and my actual history of dropping challenging games at the 30% mark told two different stories, and the data version gives better picks. If you’re not tracking yet, the Letterboxd-for-games options make it painless.
Second, I’d set the rules in one standing note instead of re-explaining nightly. Every chat tool now has some version of custom instructions or memory. One paragraph (“I game 5 hours a week, evenings, I value finishing, never recommend live-service games, always give one pick not a list”) upgrades every future conversation for free. Half my month’s friction was repeating my own constraints like a man introducing himself to a goldfish.
Does It Matter Which AI You Use?
Less than the internet wants it to. I ran the month on ChatGPT since that’s the one everyone asks about, and I spot-checked picks against a couple of rivals out of curiosity. The recommendations overlapped heavily, which tracks: they’re all drinking from the same well of reviews, forums, and lists.
The differences that did matter were features, not taste. Tools that search the web and cite sources are safer for factual stuff like “does this game have couch co-op” (the exact question mine faceplanted on). Tools with persistent memory make the standing-instructions trick smoother. Pick whichever one you already pay for or already use, apply the prompts above, and spend zero minutes agonizing over the brand. The agonizing is what we’re trying to eliminate, remember.
So Is AI Good for Game Recommendations or Not?
Here’s my honest take after a month: it’s a fantastic mirror and a mediocre compass.
Pointed at your own library, your own constraints, your own stated tastes, it reflects them back organized, and that organization is worth real money in reclaimed evenings. I finished three games during the experiment month. My usual rate is one, maybe.
Pointed outward at “what should I play next in all of gaming,” it’s a nicely-worded average of the internet. For actual discovery, humans still win: the friend with weird taste, the small blog that champions a game nobody’s heard of (hi), the Discord thread where someone won’t shut up about a puzzle game from Estonia. Taste requires someone who can be wrong in an interesting direction, and consensus machines structurally can’t be.
So the system I kept is a hybrid. Humans and places like our Discord fill the pool with candidates. The AI handles the nightly draw from that pool, where its tirelessness beats my tiredness. Discovery stays human. Dispatch goes to the machine.
Ninety games, five hours a week, zero minutes left for scrolling. That math finally works.
Would ChatGPT rank YOUR backlog correctly? Try the regret prompt and bring the results to the TAG Discord, the reactions are half the fun: https://discord.gg/bp2qn6mwC3
FAQ: Using AI to Pick Your Games
Can ChatGPT recommend video games accurately?
For organizing and picking from games you already own, yes, very. For facts about specific games (modes, features, lengths), verify everything. It stated a nonexistent co-op mode as fact during my month.
What’s the best ChatGPT prompt for choosing a game?
Constrain it hard: time available, honest energy level, desired feeling, and “one pick, no lists.” Open-ended prompts produce generic lists; tight prompts produce decisions.
Is AI better than Reddit for game recommendations?
It mostly IS Reddit for game recommendations, with better manners. Use it to sort and schedule. Use humans with strong opinions for discovery.
Does this actually save time?
That was the clearest result: outsourcing the nightly choice roughly tripled my finish rate for the month, because the 25 minutes I spent scrolling became 25 minutes playing.
