Cheap to generate a thousand game concepts, expensive to pick one

A concept generator will hand you a hundred game pitches in the time it takes to make coffee. A roguelike about tax auditing. A city builder where the citizens are ghosts. A rhythm game about repairing bridges. A survival game set entirely inside a moving train.

Some of those will make you sit up. That is a real effect and it is worth something. Combinatorial novelty is genuinely one of the things models do well: they hold a very large space of references and can join two of them that nobody had joined before.

But notice what just happened. You now have a hundred pitches and one life. The generator solved the part that was never expensive. Nobody who has failed to ship a game failed because they ran out of ideas.

The expensive question is which one is worth six months, or eighteen, and that question is not a text-generation problem at all. It is a judgement problem, and judgement runs on information the generator does not have.

It does not know what you are good at. A concept that leans on procedural systems is a gift to one person and a trap for another. The same pitch has completely different odds depending on who picks it up.

It does not know what you can stand. Development is long and most of it is unglamorous. A concept you find mildly amusing today will be intolerable in month nine. The people who finish games are usually the ones who picked something they were still curious about after the novelty wore off.

It does not know your scope. An open-world detective game and a single-room detective game read almost identically as a sentence. One is a weekend and one is a decade. Nothing in the pitch tells you which you just received.

It does not know what already exists in the shape you are describing, or how those attempts went. And it cannot tell you whether the pleasure it is describing actually appears when you play it, because nobody can know that from text.

And it does not know your circumstances. Whether you have three months or three years, whether you are doing this alongside a job, whether you have someone to make music, whether you are willing to spend a year on something nobody may buy. Every one of those changes which pitch is the right pitch, and none of them are in the pitch.

New combinations are easy to produce and easy to over-value. A pitch that nobody has attempted is often un-attempted for a reason, and the reason is usually that the two halves do not actually reinforce each other.

Compare that with combinations that worked. Crypt of the NecroDancer joins rhythm and roguelike, and the join is load-bearing: the beat forces you to commit to moves without deliberating, which fixes a real problem with turn-based dungeon crawling. Vampire Survivors strips a genre down to almost nothing and the reduction is the point.

In both cases the novelty is not the value. The value is that one half solves a problem in the other half. That is a much stricter test than surprise, and it is the test most generated pitches fail. A roguelike about tax auditing is surprising. Ask what the auditing does for the roguelike and you usually get silence.

So when a generated concept excites you, run one question at it immediately: which half fixes something in the other half. If neither does, you have a joke rather than a design, and jokes do not survive month three.

Here is a small set of questions that a concept has to survive before it earns any real time. None of them require building anything.

What does the player do, in a verb, second by second? Not what the game is about. What their hands do. If you cannot answer in one short phrase, you have a setting rather than a game. Setting is not worthless, but it will not carry production.

What is the smallest version that would still be interesting? If the answer requires an open world, a hundred hours of content, or an economy simulation before anything is fun, the concept has a bad shape for one person or a small team.

What is the risk, and can you test it in a week? Every concept has one thing that decides whether it works. Find it and name it. Being unable to name it is itself a warning sign.

And finally: would you still want to play this after a hundred hours of making it? Not the pitch. The specific repetitive activity at the centre. That question has ended more shortlists than any other.

Generated concepts are almost always hooks. A hook is the sentence that makes someone want to see more. A loop is the thing they do for eight hours.

Papers, Please has a fantastic hook: you are an immigration officer at a miserable border. But the game works because the loop is a document-checking puzzle that gets progressively harder as the rules multiply, with a family whose survival depends on your speed. The hook sold it. The loop is why anyone finished it.

Concept generators produce hooks because a hook is a sentence and a loop is a system. You can usually tell within a line or two. If the pitch describes a situation or a theme, it is a hook. If it describes an activity that repeats with variation, it is a loop.

The conversion from one to the other is design work, and it is the point at which most exciting pitches quietly die. That is fine. A pitch that dies on paper cost nothing. A pitch that dies in month seven cost everything.

It is worth doing this conversion out loud before you commit. Take the pitch and write the eight-hour version: what the player is doing in minute one, minute thirty, hour three, and hour eight. If the activity in hour eight is identical to minute one, you have found the shape of the problem, and you have found it for the price of a paragraph rather than a year.

None of this argues against using a concept generator. It argues for using it the way you would use a brainstorm: as raw volume that you then cut hard.

Volume has a real benefit. Reading fifty pitches quickly teaches you what you actually respond to, which is often not what you thought. If forty-eight leave you cold and two make you open a notebook, that is genuine information about your own taste, and taste is the thing you are actually trying to develop.

Generation is also good at variation on a concept you already have. Take a pitch you like and ask for twenty adjacent versions with one element changed each time. That is a cheap way to explore the neighbourhood of an idea, and the neighbours are sometimes better than the original.

And it is good at forcing constraints. Ask for concepts under specific limits: one room, two colours, no text, one input. Constraint is one of the more reliable idea generators in any medium, and a model will happily obey a constraint that a person would keep wriggling out of.

The discipline is that generated concepts are free to produce and free to delete, and you should exercise both freedoms. Keeping a pitch because it took effort to make is a trap that does not apply here, which is genuinely liberating if you let it be.

The through-line for all generated material applies here too, in its own form. A generated concept is a strong placeholder and a weak final decision.

It is a strong placeholder because it gets you out of the blank page instantly, gives you something concrete to react to, and is easy to discard. Reacting to a bad idea is much more productive than staring at nothing, and models are excellent at producing something to react to.

It is a weak final decision because the pitch cannot tell you the one thing that matters, which is whether the activity at the centre is fun. That is not knowable from a description. It is knowable from a rough playable version, and the gap between the two is where most projects that fail actually fail.

The gap is not large. For most concepts a week of scrappy building will tell you. What it will not tell you is anything about art, story, polish, or scope, and that is fine. You are answering one question: is the verb good.

Generate widely. Fifty pitches, one sitting, no filtering while you read.

Cut to five on gut reaction alone. Do not justify. The ones that made you want to open a project are the ones that survive; everything else goes, including clever ones you feel obliged to like.

For each of the five, write the verb, the smallest interesting version, and the one risk. Two of the five will collapse during this step because you cannot write the verb. That collapse is the process working.

Prototype the remaining ones roughly, cheaply, and in whatever medium is fastest. Cardboard counts. A spreadsheet counts. The point is to feel the activity, not to build a slice of the game.

Then choose, and accept that the choice is a bet made on incomplete information. It always is. The value of running the shortlist is not certainty. It is that you spent days rather than months finding out which bets were bad, and you are the one who played them and decided.