The Thousand Tombs of Mystic Promise →
In Worm, a minority of the population — less than a percent — has superpowers. This requires a trigger event, the worst moment of your life. These have been spitting out capes since the 80s. The exact fraction that results depends on where you’re counting: it ranges from one in 8k to one in 26k. But let’s stop and multiply. That already implies millions of parahumans. We’re never given exact figures — wildbow isn’t a numbers guy — but this analysis examines the stats more closely and anchors it to canon details. It all lines up.
In the Worth the Candle series, the world is chock full of “entads” — which is to say, magic items: vorpal swords and rings of power. Each is produced in an random act of supernaturally-inspired creation, a fugue state granting no control over the magical abilities confered. How many entads are there? Alexander Wales, fortunately for us, is a numbers guy. We can consult the setting’s worldbuilding doc, where it’s explicitly noted the upper bound on entad quantity is about 2.1 billion.
Those are staggeringly large figures. Most of you can come up with a couple cool ideas for magic item, and if you sat down and got toiling, you could manage dozens, though the strain may soon become evident. I would consider RPG sourcebook abundant, if not excessive, if it had listed hundreds of magic items.
Once you’ve broken your back devising the most exhaustive power collection anyone’s written, you’d still need to 1000x that to describe these settings unabridged. So maybe you’d be tempted to just give up there. After all, those figures are the quantity of capes and magic item, not the number of unique effects.
The Binding of Isaac has hundreds of items. New ones get drawn, avoiding repeats — and one of the infamous tricks for “breaking the game” is by exhausting the item pool, which forces the game to reshuffle and draw from those already seen.
So maybe you come up with a few dozen core powers, but in practice, some people are just speedsters. When you’ve met one, you’ve met them all.
Except that’s not how these settings treat their powers. Etymologically, “entad” was coined to mean one-of-a-kind.
Weaver Dice consists of reams of documents on the topic of generating powers. In a cursory search, I couldn’t find it stated explicitly, but it’s not hard to read between the lines. Parahuman abilities are deeply personal — wildbow wouldn’t have written a booklet’s worth of exposition on power generation if just picking one off the rack was a viable option.
Burnscar is a pyrokinetic — except she can teleport through her flames and loses control the more fire burns around her. Lung is a pyrokinetic — except that’s a mere side-effect of how the longer the fight goes on, the more he transforms into a dragon, scaling to match any threat. Circus boasts minor pyrokinesis along with a grab-bag of other abilities like hammerspace and preternatural balance and coordination. Spitfire spews napalm-esque combustible chemicals that melt almost anything — including herself.
Essentially every classic power gets this sort of treatment: there’s always a twist.
So, again we have that dilemma. We can’t reuse ideas, and we probably can’t cheat and say this guy’s unique power is to resist fire damage by 20% while this girl’s “unique” power is 33% fire resistance. We won’t simply rescale and reskin powers.
Even still, another escape hatch beckons. Because how hard it is to hit a million power ideas depends on how good those ideas need to be, doesn’t it? In terms of scale, My Hero Academia outclasses Worm by a few orders of magnitude: it outright gives 80% of the population quirks. Except some of those quirks are things like “extra flexible fingers” or “blow bubbles (they’re just normal bubbles)” or “slowly attract small objects.”
But Worm and WtC do not take this route. The source origin of parahuman powers essentially requires that powers be, if not good, then useful. Some are C-tier, but I don’t think you get any F-tier powers.
For entads, it’s harder to say: the text gestures at a notion of “i-factor”, but never bounds it from below. It’s possible that shitty entads exist, but we never see them for for the same reason any story element is disproportionately interesting rather than boring — the narrative anthropic principle.
With that, I believe I have paced the perimeter bounding this cage we’ve found ourselves in. How do you write about a world with a million special abilities, each one unique and useful?
Perhaps this seems a niche concern — if it’s such a strain on disbelief, we can always not write settings that entail millions of bespoke effects.
But the true scope of this trope isn’t so easy to escape. How many times have you read a story with a magic library, if not a whole magic school? How many times have you seen a wizened magic veteran characterized by an appellations like “master of a thousand spells”?
I’ve always found myself faintly mystified, never fully satisfied by these portrayals. The art direction for magic libraries always renders them as having shelves upon shelves — at a glance, you can spot hundreds of books. Maybe not all of them are about magic, but some of them must be. What’s in them?
Even the most complicated magic system you’ve heard of is necessarily still described by a fraction of the book you read it in. In most cases, the rules of the system are either vague or clearly summarized, and both fail to truly convince me that you could write an entire textbook of exposition.
What amuses me — and helpfully serves to indicate why almost no one does this — is that the key to making a magic textbook feel necessary is, in some sense, writing the system poorly. If you’re designing a magic so rule-bound that you could write a textbook on it, then you probably care about making it clear, and limited, and all those Sandersonian virtues.
To motivate a textbook, you’d have to make your magic system confusing. The reader ought to be frustrated and dissatisfied with the simplified explanations offered in passing dialogue, wishing instead to just get a complete explanation.
Beyond spells and magic items, what about bestiaries? There are currently a thousand pokémon. That series relies on “regions” to explain why this new game has a new batch of monsters you’ve never seen before.
Kanto, famously, is an actual region of Japan, roughly 32 km2. Let’s round that figure up to 51 to account for the sourthern waters (some mons are aquatic, and we do visit islands) and compare that to the surface area of earth (510 million km2) and even supposing that each Kanto-sized patch of the planet has a paltry 1 unique pokemon will still net us millions — the very same order of magnitude that’s been haunting this section.
It goes without saying that no one can be expected to populate an entire roster on this scale. So why linger so long on this point? Two main difficulties arise.
The first is false abundance. We’re told a system is huge, but in practice it feels so small. This is what I was getting at when I digressed to magic textbooks: the empty promise of complexity when the rules of the system can be explained and understood in five minutes. The promise of endless diversity when, after millions of words and multiple volumes, we’ve seen maybe a few dozen spells total.
The second is blank page syndrome. I run into this often when drafting short stories set in my fantasy stories — especially Black Nerve and Vermin Cathexis. The magic systems I design are always strive for that sprawling abundance and complexity. So when my characters encounter a new magic-user, what should their gimmick be? It’s lame if it’s just a stock power like pyrokinesis or levitation. There should be more to it.
I can come up with some pretty damn creative powers when I need to But my process of is slow. I might spend all day working out the ramifications of an idea and crystallizing it into something coherent.
That’s unacceptable if the ultimate destiny of these powers is to get handed off to a team of jobbers that appear in a single random encounter. Not only is this inefficient, it’s wasteful.
One trope common to battle gimmick stories, played both for comedy and tragedy, is some character showing up to a fight — only to be killed before the audience gets any explanation of what their deal actually was. They had not yet realized the full potential of their gimmick. Now they never will.
A power that’s just pyrokinesis is lame, there should be a twist, but if we never see the twist in action, they effectively were just a pyrokinetic.
This finally brings us to the main subject of this post. Both of the problems outlined above could be solved by quickly sampling from a rich space of powers. If it’s quick enough, it’s useful for random encounters. And if the space is rich enough, it’s easier to believe “master of a thousand spells” is meaningful, once we’ve see enough unique spells to be convinced there really could be thousands of the things.
The solution we’re groping toward is clearly some variant of a random table. Design a couple of templates like {“control of”, “immunity to”, …} and {“fire”, “ice”, …}. Assign them to dice rolls or card draws, and you’re good to go.
← Combinatoric Extension Techniques →
Alexander Wales really likes combinatorics as a design tool. Rather having tables spit out ready-made powersets, he prefers a more interpolative approach, creatively combining two concepts.
One of my favorite worldbuilding things is combinatrics. Let’s say you have one physical affinity, like Power or Speed or whatever, and one elemental affinity, and you mix them together, and even with short initial lists you get a lot of interesting things.
Like what does it mean for a person to be Power-aligned and Leaf-aligned? Here you’re sitting down for a session of interpretation of your own elements and how they mesh together, and I find this easy to use as a starting point for getting creative with it. Maybe combining “Leaf” and “Power” means a single honking big leaf, the kind that you would wield like a buster sword. For me, that’s the kind of idea that immediately feels like it has something to it. And then you move on to the next one!
Leading to examples like:
Speed-Teeth is hard, sharp, cracking speed, not quite teleportation but the speed of something inflexible. It works better with a rigid subject, like someone in full plate armor, though practitioners without that kind of money sometimes just strap bits of metal to themselves.
Precision-Leaf is a striking power, best able to hit one thing of many, a knife thrown precisely through a crowd to hit its target, a jab to a weak spot. Metaphorically, it’s plucking the single correct leaf from a tree without even trying that hard. In theory, it’s taking a chaotic, varied system and isolating an element. In practice, it’s best at target selection.
The power of combinatorics is the numbers get very big very fast. If you can populate five tables with fifteen elements each then you’re already three quarters of the way to a million! Add five more elements and you’re good to go.
The weakness of combinatorics is the Ten Thousand Bowls of Oatmeal Problem. Classic elemental systems have 4-5 elements. If you need to push it to sixteen, then you need to split hairs: water isn’t the same as ice; earth isn’t the same as metal. But consider rolling a combination that describes something like charged + body-based + explosive + movement. That’s four of the tables locked down, how much difference does it make if the last roll said this was an “ice” power or a “water” power?
But let’s loop back to where this tangent began. Wales doesn’t just advocate for combinatorics, but a particular form: conjunctive combinatorics. In fact, he long ago whipped up a simple generator to that effect. We can spin the wheel a few times and get power-prompts like:
- Earth manipulation. Yes, but crowds. Could go either way: you can earthbend, but it gets weaker the more people are crowded over it — or the crowds are the source of the power, and you need an army to move mountains.
- Time travel. Yes, and absorption. This sounds like the plot of The Girl Who Spun Through Time, where the protagonist is drawn toward and consumes the magical power of herself in other timelines. A different possibility is time travel as armor — a weird variation of revision magic? — where traveling through time absorbs incoming damage.
- Air manipulation. Yes, and fear. Screamer, anyone? The obvious interpretation is being a horror movie sound effects team, so non-obvious option: you hyperventilate when scared, amplifying your power.
I wasn’t actually proving a point with that spate of brainstorming — it just happened. This generator is so compelling that you can’t help but try and make sense of its outputs.
To Wales, the strength of Worm’s system is the power of yes-and, though as you can see he tossed in some yes-buts. This is the simplest instance of conjunctive combinatorics. How far can we take this?
A few more rolls with new conjunctives swapped in:
- Fire manipulation. Yes, if absorption. You have to burn yourself first, and can only store so much fire.
- Animal control. Yes, for movement. You can only control animals that you ride.
- Probability manipulation. Yes, because feedback. Your ability to manipulate outcomes improves the more times you try (and fail) to affect them.
- Emotional manipulation Yes, or velocity. You can completely transform the moods of those around you as long as you and your target stand still.
- Fire manipulation. Yes, else age. You are immortal as long you burn the world around you, otherwise you turn to ash. Keep the fire alive.
- Air manipulation. Yes, then heat. You can fly and send objects flying, but they leave a kind of rocket trail of superheated air behind them.
- Shapeshifting. Yes, each pain. You transform uncontrollably whenever you’re hurt, more severe injuries yielding greater metamorphosis, along with incomplete healing.
- Telekinesis. Yes, while insanity. You can enter a flow state governed by random impulse and apply ghostly force to any objects — with no guarantee you’ll use for anything save self-preservation.
I’m going to call it there. I thought of a time-based conjunctives — before, after — but they gave me a profound streak of bad luck! Though in fairness, “before” is almost equivalent to “then” or “if not” while “after” is “because” or “each.”
I actually went as far as to write an entry for “once:” But I first had to reroll this a few times. But, still unsure, I tried one more roll, just to convince myself it was meaningful. “Plant manipulation once pain” is when I realize oh, this is literally just equivalent to “each.”
With that said, if you want to hack AW’s generator to spit out these conjunctives, open the dev console and run:
butand = ["but", "and", "or", "if", "as", "of", "like", "then",
"through", "else", "to get", "due to", "each", "while", "until", ]While you’re at it, you can add entries to powers to change the lefthand side and modifiers for the right. But if you’re going that far, you might as well start writing your own generator using all of the tricks this post will discuss.
← Reduced Instruction Set Worldbuilding →
See, I didn’t actually list every productive relationship more specific than “and” or “but.” There’s a special category of relation that Wales’s generator is rather unsuited to demonstrating — and I think it’s a key ingredient to escaping oatmeal combinatoric hell. By which I mean: context-aware conjunctives.
Two easy ones are exceptions and specialties. With effort, you can make arguably make sense of “Water manipulation except for happiness” or “Precognition, especially pain,” but the hit rate for this is pretty low. You’d instead want the generator to maintain some sort of conceptual hierarchy to draw on, so it knows what to lean toward. “Fire manipulation, except for smoke” or “Plant manipulation, especially grass.”
Another kind of context-aware conjunctive is generalization. I originally included this on the list above, but it was disproportately long and confusing. On revision, I realized the difficulties are because it works better with context. To illustrate, this is what I had:
Flight Yes, with theft. You lift things — steal away, liberate — transferring metaphysical ownership to yourself. Gravity is an expression of the planet’s ownership of all things with weight, so by stealing yourself, you can fly. It’s harder to use your power on larger or uncooperative things (living things don’t want to be enslaved), meaning the easiest targets are yourself and small, easily-snatched objects.
You , but if they’re randomly selected, the route is both tortured and annoyingly nonspecific.
Now, let’s consider a power like Grue’s — he has “darkness generation,” but I think the real heart of its structure rests on an analogy. He creates darkness that is like clouds. Once you look at this way, so many powers fall into place. What are kinetic powers but the ability to treat a medium like limbs? Scrying lets you treat a surface like eyes.
But try this with the toy generator, and you’ll get “Flight. Yes, like charging.” The intuitive read is that it’s flight you have to charge up. How is that any different from “and charging”? Key to generating interesting metaphors would be to know to pick something that doesn’t fit.
I tried generating random words until I rolled concepts that don’t fit together. The first promising set was “user glide thread snow illness.” Glide-thread suggests a neat power: spiderman web-shooting but hang-gliding instead of swinging, and perhaps a bit of kiting.
“Like snow” is easy — perhaps as you glide, the expended threads precipitate down. “Like illness” is a puzzle. “The power makes you sick” isn’t right — that would just be “and illness” or “but illness.” The power itself must be inherently akin to an illness. Are the threads symptoms, then? Discharged like pus? Perhaps the easy answer is the Haunting from Hollow Knight: Silksong — the threads possess you.
This in particular is hard enough I’m skeptical you can automate it without an intractable amount of data and precomputation. Conceptually similar things yield banal metaphors, but conceptual dissimiliar things yield random noise. The most effective approach may in fact me generate N random words and playing connections.
But if you return to our list, you’ll find that a few of these “conjunctives” behave differently. There’s a profound difference between “yes, but” and “yes, then.” Specifically, the connectors I want to highlight are “then,” “while,” and “each.”
If you have any experience programming, those neurons just lit up. The thing these all have in common? Control flow.
We’ve all heard of the power of “yes, and” — now let’s consider the power of “yes, and branch if”
For the laity, this means powers using these conjunctives track and alter some internal state. A simple example of how this makes for interesting power, consider an ability that allows you to bask in the sunlight, soaking up solar power. You can then expend solar power to grant you a subpower until the sun sets. These abilities are: (a) firing solar beams, (b) glowing eyes augmented to treat all darkness as if illuminated, and (c) equipping hard light armor whose durability depends on the strength of the light source currrently illuminating you. Neat power, but consider its structure.
Sunlight, then chargex Spend then no chargex and {{laser xor vision xor armor} while daytime}.
Why chargex and not “charging” (as in Wales’ generator)? The subscript variable is meant to indicate correlation. You spend exactly what you charged. It’s how we indicate this isn’t, say, a power that has two different chargeable states.
Introducing conjuctives are helpful — our original goal of reaching a million entries called for five 16 element lists, but we can get rid of two of them if we just insert conjunctives.
But introducing variables? Now taking N random words doesn’t just grant exponentially many combinations, because each of those combinations can be combined into some sort of random equation or graph structure. The equations themselves can define variables to be used in other equations! We haven’t defined our formalism yet, so we can’t precisely calculate, but I suspect this scales with the powerset 2N.
Before I started writing this blogpost, I had first tried to put together a computer program in the vein of AW’s generator. Unfortunately, recursively generating formulas with shared variables and derivatives is a bit more involved than rolling on two tables and sticking “yes, and” between them.
You could just abitrarily slap together variables and stateful conjunctives, but key to a good power is having a cogent structure. How do you make sure you don’t spit out nonsense? You could hard-code a structure, but then you’re restricting the results, closing off possibilities.
But the more I reflected on what I was doing, the more I realized trying to do this automatically is a somewhat doomed endeavor anyhow. An intuition I’ve developed, after studying Turing complete systems, is that any form of expression powerful enough to do interesting things is also hamstrung by chaos and incompleteness. This is the logic behind results like the Halting Problem or Rice’s Theorem.
Or put in a simpler way, the nature of the Oatmeal Problem is it occurs when procedural generation confines itself a well-behaved “safe” region of the possibility space.
But the only way to systematically allow for unexpected gems is accepting that you’re going to shovel a lot of dirt to get to them.
Which raises the question: what gems are we actually looking to find? Wales’s generator is pretty dang good, especially if you toss in our tricksy new conjunctives. I’m sure if you care enough about power systems to read this essay, some of the example abilities I dashed out above already struck you as kinda cool.
Am I just adding complexity for the sake of it? Is all of this recursive computer logic stuff pure wank?
← Praise Be Fiddly Bullshit →
Wales’s took Worm as his muse. For myself, I might pick Jujutsu Kaisen.
Consider Satoru Gojo, one of the most iconic characters. His ability ⸢Infinity⸥ lets him manipulate space. He can impose effectively infinite distance between himself and any danger, stoping its approach. Easy enough to call this, say, “Space manipulation and distance.” But he can also recreate blue orbs where points in space are drawn together. So: “…and distance or attraction”? He can also invert those orbs to get red projectiles that explosively repel. “…and distance or (forces and projectiles)”?? But finally, if red and blue collide, their overlap creates a purple orb unleashing tremendous destruction. So maybe:
Space manipulation for distance or projectilesz
projectilesz and {attractionx or repulsiony} or {explosion if attractionx and repulsiony}
Now consider the grand antagonist of the series, Ryomen Sukuna, and his ability ⸢Shrine.⸥ The core of this ability is easy enough to formulate. Dismantle delivers a ranged cutting attack, while Cleave penetrates magical defenses at close range. We have “slashes and {distance or pentration}”. A few arcs into the story, he demonstrates an inexplicable ability to summon a “fire arrow.” This ultimate move makes for a formidable enough of boss attack, but it’s hard to square with an ability that is otherwise admirably straightforward.
At the climax of the series, it’s revealed , which represents less a special attack than a state change in the dynamic of the fight.
More than anything else, Sukuna is a boss fight, and a good boss fight has multiple phases.
But I saved the throne of this tricolon for the most egregiously multifaceted ability in JJK: the Ze’nin clan’s inheritance: Fushiguro Megumi’s ⸢Ten Shadows.⸥ The ability centers around taming monster summons, and their abilities range widely, from flight to water blasts to adaptation to any attack. Briefly:
summoning if challenge each monster
monster and fragile or fusion and cost
dogs: multiple and free
nue: flight and eletricity
toad: strength and capture but tongue
serpent: size and capture
elephant: water and blasting but cost
rabbit: quantity and distraction
deer: healing and negation
ox: strength but charging but momentum
tiger: aura but mystery
mahoraga: negation and evolution but time
Ten shadows is not an ability. This is dozen abilities in a trench coat.
But what’s wrong with that? Every power entails Required Secondary Powers. All powers are a synergy package. Pyrokinesis tends to grant a form of fire-sensing or fire-resistance; each could be (weaker) abilities in their own right. So here we likewise have multiple abilities joined at the hip.
Many sentence can be broken up into indepedent clauses — it’s this recursion that grants language its means of expression.
Multifaceted or evolving powers proliferate in genres like shōnen and progression fantasy. Simple combinatorics can be a good starting point for creating them, but why not add a combinator for overarching structure?
We’re almost ready to prepare our own tables. The last remaining question is also the first question: what do we put on it?
After copying over every single generation, I’ve gotten quite tired of typing out “manipulation”. It feels like half the entries in the provided tables are some sort of kinesis, but it’s missing some pretty iconic powersets. No lightning or fleshcrafting or necromancy, or sleeper hits like evolution or astral projection or pocket dimensions.
Even “manipulation” is unhelpfully broad. One automatically assumes telekinetic control, but that’s no more than the lowest entropy prior for what “manipulation” means. Consider: Air manipulation, except you can’t move or stop the air, or vibrate it for sound. You can control other properties — increasing and decreasing temperature, or separate components to alter humidity, oxygen content, filtering contaminants. Can you ionize the air and generate lightning bolts?
More egregious is “meta powers,” which I take to encompass a whole industrial complex — power copying, power negating, power stealing, power jailbreaking, power scaling, etc.
We can improve this list just by patching every hole we see, but what about the ones I don’t see? How do we know I’m not missing something obvious?
A different approach can be more rigorous: if we start from the top, the base principles of what a power must do, narrowing it step by step, at each layer mapping out the boundaries of that subspace can contain, then it’s far less likely we’ll miss anything.
But that old tradeoff snags us once again: a space big enough to contain every interesting possibility has to include a bunch of boring ones.
Let’s consider elemental systems. If you want to get materialistic, one of the easiest ways to both harden and extend bending powers is by making it the control of an element. That is, atomic elements. At once, hundreds of possibilities open up! And most of them are meaningless. What use is “yttrium-kinesis”? I have the power to create Copernicium from nothing! I can’t make it last longer than 30 seconds, though.
You’ll find yourself drawn to the most abundant elements in the universe, and probably wind up with something like
“Niter” lets you control the gas that composes most of the atmosphere — nitrogen. “Sand” less you control most forms of stone, but especially sand and glass — it’s silicon. “Ash” is usually thought of as control over fire, but you can devestate flesh and direct a ghostly gas — all forms of carbon. “Water” is the cutest, because we know “water” is a molecule composed of two different atom types — but to a first approximation, all its mass consists of that one oxygen atom, fifteen times heavier than hydrogen: control that, and you control water. This has nifty knock-on effects like extinguishing flames by starving them of oxygen.
This is a fine enough scheme for the one setting it comes up in. It’s neat how close you can imitate conventional elemental system with materialist conceits — but that’s just the thing: this isn’t what people conventionally mean by elemental powers. “Fire” is not carbon — considering how often lightning is an subelement of it, it may well be about energy and intensity more than composition. This scheme would give each type of metal — iron and copper and tin and aluminum — its own element, when half of us consider all those a subtype of earth.
Who is right? Which assumptions should our generator privilege?
← Act I: From Firstest Principles →
Originally, I had spliced in a lengthy discussion on the ontology of powers, but this piece is long enough; so I cut it and relegated that to its own post: Escaping Capepunk Ontology.
With that, I think I’ve rotated this problem through every conceivable angle short of actually doing the work. So let’s get cooking.
For a start, we could define an ability as a trio of three elements: a trigger, a process, and a medium. (Equivalently: a cause, an effect, and a target.)
Hydrokinesis become: {at will : manipulation : water}. Water bending — as portrayed in Avatar: The Last Airbender — would swap “at will” for something like “somatic” or “gestural”. If some character copied Jesus’s famous feat of transfiguration, we might change “manipulation” to “transformation,” at least as a starting point. And of course, pyrokinesis subs out water for fire.
The trigger is what circumstances cause the ability; the process is the mechanic by which it operates; the medium is what it can operate on.
I came up with two class of process with four subclasses each.
- Manipulation: changing of the properties of a medium
- Movement: changing the location or velocity
- Rescaling: adding or removing energy; making it greater or lesser
- Reconfiguration: shifting or revising subtle internal state
- Tranformation: creation and destruction and transcendence
- Perception: acquiring knowledge of the properties of a medium
- Detection: sensing the distance, orientation, or address
- Measuring: sensing internal properties and processes
- Prediction: understanding future or hypothetical behavior
- Exegesis: discerning relations and connections
You might think this list would be more fractally symmetric if there were four types. Originally it was — “Alternation” would encompass transformations of medium, including “Appearance,” “Structure,” and “Composition.” But what is a transformation of appearence? Are you manipulating the light that reflects off an object? Are you manipulating the mental experience of sight? Are you manipulating how the medium reflects light?
After cutting that, I considered replacing it with a suite of “meta” processes, but there’s nothing in the definition of medium that suggests it couldn’t be another ability.
Even after cutting Alternation, I still had the category of “Modulation”, which included “Potency” (how the medium impacts others), “Resistance” (how the medium gets impacted), “Flexibility” (range of causes and effects), and “Evolution” (accelerating internal processes). But most of those are special cases of manipulation.
Exegesis merits some explanation — originally it was retrodiction (“History”), an obvious symmetric parallel. But besides that being a specific and uncommon power, what would sensing the history of a medium mean? Specifically, what knowledge would it grant you that’s meaningfully different from careful Detection and Measuring? Even Prediction steals from it — the past is a hypothetical.
Exegesis, then, is something else entirely. A character with the power of Water Exegesis can’t sense water. Rather, they sense through the lens of water. Perhaps they intuit when something is impacted or associated with water. Does this make sense? Maybe. It feels like there’s something interesting here.
(In the rest of this essay, I go on to make a few jokes about Exegesis being hard to understand, but it’s been a few days. After going over this post several times — and crafting my own generator — I think I have a better handle on it. Two iconic examples of Exegesis powers would be Tinkering and Communication. The former is understanding how best to use the medium (e.g. in devices) without having control over it, and the latter treating the medium as if it had a mind and will, again without control over it.)
All told, I’m not quite satisfied with this. To illustrate, the defining example of “Potency” is making a sword supernaturally good at cutting. How, though? There are a few ways this could work. The obvious, reductionist answer is by optimizing the material structure to be sharper — but this could only make a sword naturally good at cutting.
To go beyond that, you’d probably want to telekinetically maintain the edge and integrity, letting you do classic stunts like cutting lesser blades in half. Another option is telekinetically splitting apart the target around the blade, sundering with magic itself.
I suppose the synthesis here is to recognize that if your power is something like augmenting swords for cutting, then this can be expressed as sharpness manipulation.
But enough about media — what about triggers?
- Psychic: activated by mental states
- Willful: focusing attention and willpower
- Emotional: feeling certain emotions and intentions
- Principled: abiding by sworn rules
- Alienated: varying through separate, inscrutable whims
- Somantic: triggered by bodily activity
- Verbal: speaking words of power
- Gestural: performing, dancing, or signing
- Reflexive: automatic response to stimulus
- Metabolic: passive result of survival or grwoth
- Conditional: depending on properties of the medium
- Prepared: investing time or energy in the target
- Inherent: possessing certain natural qualities
- Contingent: attaining a certain state
- Personal: uniquely linked to the user
- Circumstantial: factors outside the user or the medium
- Expensive: depleting a resource
- Random: occuring unpredictably
- Consensual: considering the target’s preference
- Delegated: giving control to someone else
This is a pretty broad spread. And some of these distinctions are very fine. Is the difference between “Prepared” and “Personal” clear? The former is a telekinetic that can only control objects they’ve “charged up” beforehand; the latter is a telekinetic that can only control one object, e.g. their spirit weapon.
What about “Alienated”, “Consensual” and “Delegated”? An user with Alienated Portal Generation can’t control when or where they create portals, their power does what it thinks is best for them. Delegated Portal Generation is easy: rather than (just) creating them yourself, you empower others to do so.
Consensual Portal Generation is trickier to assign meaning to. See, my original idea was, e.g. “Consenual Animal Control” means that the animal can shrug off your commands. But in the general case, this leads to funny animism. “Consensual Hydrokinesis” suggests that each body of water has its own mind, and you have to appease water spirits.
Consensual Portal Generation, though? If the portals themselves have minds of their own, how is it practically different from Alienated Portal Generation? Or maybe it’s not the portals, but the space they tunnel through, so that you can only portal into someone’s house if that house allows you come through. Cute, but what determines what the house wants? Animism is one option, but a more natural option is that if it’s my house, then whether I want you inside is the operative question. But at that point, is it any different from Delegated Portal Generation?
Words and concepts are defined by deferal. If you have “Red” and “Blue” buckets, then you must categorize colors by how warm or cool they are. But if you add a “Cyan” bucket, suddenly a whole bunch of colors stop being “Blue.” The color hasn’t changed, but the concept gets redefined: even though they’re still blue, they’re more cyan than they are blue.
If I deleted the “Personal” category, then that niche of powers — spirit swords, astral projection, etc. — can be more properly understood as being Prepared or Inherent in some specific way.
Another implication here is mutual exclusion. If I roll a power as Reflexive, the fact that there’s a Willful category might lead me to conclude it’s therefore not willful — i.e., it’s involuntary. Yet by that logic, half the possibilities imply completely involuntary powers. But they just aren’t that common, in my experience.
While writing the list, I originally included words like “require” or “demand,” but I deleted them. Willful doesn’t mean voluntary as opposed to involuntary, it means your will is in some sense tested or depleted. Verbal or Gestural don’t mean the theater is necessarily, only that it notably empowers or affects the ability in some way.
All of these can stack — in fact, most written magic systems demand several of these factors.
But all of this fiddling leaves me with a question: what are we even doing?
← Act II: Dark Night of the Goal →
More pointedly, are we building a power generator, or a power prompt generator?
When I, bright-eyed, started putting together my own code, I had the idea that I would totally one-up AW’s generator by spitting out detailed, well-formed sentences.
One of the premature optimizations that kneecapped my progress was I payed a lot attention to how grammatical my outputs were. Achieving context-aware results required that my dataset encode relations — so the entry for “Manipulation” pointed to the entries for “Movement”, “Scaling”, etc. — as well as including fields for name, description, and the exact phrase to include in the output sentence.
Once I started making things properly recursive, I realized I ought to include a field for the gerund form of process verbs, so that they could be qualified in dependent clauses, which even at that point meant adding additional lines to two dozen definitions and…
My “am I just wasting my time” sense started tingling.
I try not to spend too much time on this site griping out current events, but at this point bear a mention. After all, if my goal is generating well-formed sentences, no script I cook up in a day is going to outperform the slop machines.
LLMs, of course, suck at brainstorming. Now, you can overstate the “glorified autocomplete” talking point, but the inescapable nature of the prevailing machine learning regime is statistical approximation — a form of averaging. The most likely token is the least surprising.
So, what use have I for programmatically generated text? Why would it be more impressive for my script to output more articulate sentences than some toy javascript applet from ten years ago?
What my desires amount to, ultimately, is producing more complex and coherent output. “Precognition. Yes, but happiness.” is obviously incomplete. I can design a generator that, after rolling “Fire manipulation”, knows that it can pull from a table of things related to fire — light, smoke, heat — and might preferentially steer toward those when generating further concepts.
Why?
What generators do, ultimately, is a form of automation. I could pull up a word list and scan it until one of them catch my eyes, or simply close my eyes and wander my mind palace until inspiration arises.
When take a prompt and flesh it out into a power, my thinking follows patterns: there’s a set of mental motions I do every time to make sense of things.
And that’s the key: I want outputs that make more sense.
I am turning input A into output Z, and wouldn’t it more efficient if I could start at step M instead of A?
Lately, there’s been a recurring phenomena where game developers have left diffusion-generated assets in published titles, ostensibly by accident. Why does this happen? State of the art generated output makes for poor placeholders, for the dual reason that they do not suffice as a final product, yet they still appear fine enough that, after rounds of testing and polish and putting out fires, you might forget about them. Good placeholders look obviously, unmistakably unfinished, precisely so that this doesn’t happen.
A good prompt almost works. An important part of “almost” is that it means “doesn’t.” Prompts do most of their useful work in those first moments of sense-making.
Consider the “bad rolls” AW’s generator sometimes spit out — “fire manipulation but heat” or “time travel and time” — they are uninspiring because they make already sense. There’s nothing to do with them but extend a rather rote exercise.
Part of why I conjecture LLMs are so deleterious is the anchoring effect. Once you have an answer, you’re inclined to stick to it. If you need an answer and don’t have any, then you know to start looking. But once you have an answer, instead you need only ask: is there a better one? Equivalently: is this one close enough? (An LLM’s answer is optimized to be one fits best on first blush — it makes the most statistical sense.)
How many species of fish are there? If I say “two or three,” you know it’s nonsense. If I said “around five hundred thousand,” contextually you already expect that I made that up whole-cloth, but doesn’t it sound about right? It makes sense, and cognitive psychology suggests that starting point will influence your guess, even when you try to correct it.
This creates a paradoxical scenario where a worse answer is better, when glaring insufficiency provides more motivation than vague imprecision.
To wrap up a confused tangent, I don’t want to press a button and get a complete power: I want to press a button and get an inspiring puzzle, a test of my creative muscles. But I do want that sweat to produce something worthwhile.
If the space of possibilities were a great big rolling plane, I don’t want to get stuck in one corner of it, I want to explore the whole thing. If I’m told exactly where to go, I’ll probably wind up somewhere nearby. But if I’m told what direction to head in, I’ll wander until I see a cool point of interest.
The former can only produce the best results if someone already knows the approximate location of the best result. The latter takes longer, but walk long enough and you’ll find the most interesting landmark in that direction.
This leads to a better objection to AW’s generator. If I roll dice to determine which direction to check, I can only ever go in six directions. As noted earlier, there are some powers that never come up on the lefthand side.
Is that all we need, then? Repopulate the list with an improved power pool? Well, no. Remember what my efforts to categorize powers yielded? We broke “Manipulation” powers down into four subtypes, only one of which is what everyone thinks of — Kinesis. Worse is “Perception”, which has three reasonable enough subtypes, then “Exegesis,” which I still haven’t quite wrapped my head around.
So perhaps one of the fixes is to introduce a weighting system? We can give Movement more weight than Reconfiguration.
But now we must ask: what do we want our distribution to reflect? The point of consulting a random table is for unexpected output. If we want obvious answers, we can think of them on our own, or ask ChatGPT. We want our generator to spit out unusual answers.
Don’t we? Let’s go all the way back — remember how this post started off? We analyzed settings that implied or stated there were millions of unique magical effects somewhere offscreen. We then trotted out procedural generation in the interest of more tractably reaching that target.
For that application, we genuinely do care about accurately modelling distributions. It would be weird if everyone in the world had bizarre off-beat powers. It’d have setting-level ramifications if only 25% of sensors can actually detect the location of their element, and another 25% are all doing some mystical bullshit you can’t clearly explain.
Is Fire Manipulation more common than Delegated Phantom Essence Transformation? The latter is a better answer to “what’s an interesting power?” and former is a better answer to “our characters just encountered a random guy, what’s his deal?”
The correct, boring solution to this problem is that you should math out the power statistics for your setting — e.g. 60% Manipulators, 10% Perceivers, 30% hybrid — then roll on your own custom tables.
But in order to construct those tables, you need to appraise the full space of possibilities. So let’s — finally — get into what I teased earlier, and explain how to go full recursive.
← Act III: Hard Takeoff →
This character has the power of Prepared Plant Movement. By touching flora, they can imbue energy that accelerates and guides growth. This means, specifically, that they most likely have: (a) the power to discern what is and isn’t a “plant”; (b) the power to exude energy; (c) the power to turn this energy into “growth;” (e) the power to sense where this growth is and could go; (d) the power to remotely control where this “growth” ends up.
Can they, with power-sense alone, tell the difference between a brick and a bit of driftwood? What happens when when they exude this energy into something that isn’t a plant? Can they grow food or poison — can they revive dead plants? Can touch a tree and extend their sense through its branches and roots — throughout the wood wide web? Can they control the growth precisely enough to remotely control puppets?
The point here isn’t base munchkinry or brainstorming. For one, each of these subpowers change what the “real” power is. Is this a mere healer or crafter with an affinity for plants? Or are they actually a looming threat — boasting an army and continental omniscience — for which the botanical nature is at worst a detail or at best a hidden weakness revealed near the climax?
But more importantly, by analyzing the substructure of this conceptually coherent power, we can reveal the patterns and discover how to tie this into a coherent superstructure. “Plant Detection” is a weak power, and still weak if paired with “Plant Measuring,” but when extended with Manipulation powers that rely on that data?
The internal structure becomes a long sequence of “then”s. Detect a plant, then energize, then initiate growth, then channel it to the destination, then manifest it.
So we can make this power more complex by adding more steps. Where does the energy come from? Easy answer: sunlight courses with magical potential. But that feels like the same level of abstraction as the other steps.
More interesting things happen if we canonize that “plant puppets” idea, change it from a clever use-case to a supported feature. If this character grows the plants into special bodies remiscient of animal anatomy, then they can animate plants with a spark of higher intelligence, creating triffid minions.
It’s tempting to carry this further — can the triffids be vessels for storing and widely distributing the growth energy? Can the triffids evolve into higher forms, like sapient dryads? — but again, my goal here isn’t pure brainstorming.
Instead, notice the symmetry here. “Plant Minion Creation” could have been its own power!
Structurally, you could generate this using conjunctive combinations, but omitting the medium (equivalently, letting the medium be a sub-power). So {Slime Reconfiguration} would be manipulating slime to impart useful properties, and {{Slime Reconfiguration} for Prediction} makes me imagine the slime is a distributed mind, learning and assimilating intelligence.
Another impromptu roll gave me “each.” How do we interpret {{Sharpness Transformation} each Movement}? ‘Transformation’ also encompasses creation, so one interpretation is you can psychically throw objects, and higher velocities radically enhance their cutting power.
But just chaining like this limits the recursion. Return to the Plant Uplift power. One of natural designs is to have the two subpowers share states. The growth energy can be used for kinesis, but also for empowering the minions. Similarly, the power’s state must be tracking whether a given plant has been grown into a proper shape for a minion. This notion of internal state is crucial to categorizing higher-order powers.
The simplest might a simple toggle power. You can do only X until a condition is met, where you can only do Y. Perhaps Y is a powerup state, or it’s a modal power. Wind powers, except you can only use your long-range air motion sense when you aren’t using your powerful air blasts. This generalizes to powers with longer lists of modes. Maybe you can fly, too, but that means no draft sense or air blasts. In short, these powers entail choice and tradeoff.
If you limit the transitions between the modes, you get a more specific archetype: the multi-phase “clock” power. The transitions might be based on time or injuries sustained, but it’s still interesting if the character can deliberately advance the clock.
Of course, there’s more interlinked shapes than simple cycles and fully connected graphs. With enough modes, it’d be meaningful to speak of branching tree shapes, or latticed grids — but at that point, you’re talking at least half a dozen states. While each of those could its own unique power, to maintain a coherent identity, we’ll likely go for modifiers and intermediate variants of general power.
If we ditch the modal constraint, then what results is instead a “menu” power, where various conditions unlock subpowers. We’ve already touched on an illustrative menu power — Ten Shadows.
But the Plant Mastery power we cooked up just above, I would argue, is more Menu than it is Modal. Minion creation depends on growing plants, but it’s not a separate phase, just a menu option that’s “grayed out” pending enabling circumstances.
The nature of Menu powers is aggregating different subpowers that synergize, whether it it’s a big “switch statement” like TS or a step-by-step procedural flow. But there’s a third possibility. There’s a meaningful sense in which Minion Creation is a much “bigger” power than Plant Manipulation. If these minions get large or numerous enough, they outscale the base power in plot significance.
Consider the narrative arc of, say, an incremental game. You start by clicking, then you gain access to auto-clickers, then you gain access to abilities that auto-buy auto-clickers, and so forth.
Some menu powers are steps toward an end goal, but some loop back into themselves and start to fractally enrich themselves. We don’t have to go full every time — maybe there’s only two or three levels of scaling supported. But it’s worth recognizing that even a limited feedback cycles is its own unique structure.
Next, it’s tempting to notice that both modes and menus are in some sense binary, so perhaps the logical next step to generalize is to consider powers that are more granular the that. Spectrums with integers or fractions instead of booleans.
But I don’t think this actually results in any novel possibilities. What is a “spectrum” power? Perhaps twin Light and Dark powers, which start with angelic abilities that gradually descend into corrupted eldritch powers? But this is no spectrum. It’s still binary: Two powersets (Light and Dark) and rules for when either subsets is applicable.
Ultimately, you’re always drawing from a number of conceptual wells when you design a power, and when decomposing an ability into subpowers, binary vs spectral doesn’t matter as much as what the concepts are and how complex and overlapping they are.
There is one variation that does merit special representation. Consider classic power copying. Here, certain states confer subpowers, that much is the same, but there seem to be abitrarily many such states. Equivalently, the state is the subpower.
This isn’t limited to such overt cases. A common enough trope sees clever characters “hack” their power, reprogramming and extending it to new heights. The basic rule of universal computability — wizard completeness, if you will — is that if the specification of your ability has enough “whiles” and “eaches,” enough internal states, then you can theorectically reconfigure its behavior to serve whatever end you desire — it just comes down to patience and imagination. (And, one presumes, mana reserves.)
Now, if you aren’t technically inclined, your patience may be wearing thin. These complicated variables and functionalities are all well and good if you want to make arcane video gamey abilities — but some of us care more about themes and such.
But I find there’s a more natural framework for thinking about higher-order powers. Rather than programs, abilities are analogies. Waterwalking is treating liquids like solids. The pyrokinetic treats fire like limbs, as an extension of their body.
Higher order abilities, then, are extended metaphors. This character may be crystal manipulator, but thematically, their power is like a snake. When she crafts crystal swords, it’s like fangs. When she crafts crystal armor, it’s like scales. What does it mean for her to be venomous? Constricting? To slither, to scent? Can she lay eggs? Does her power force her to be solitary, or to bask, or to hunt?
The general form, then, is X is like Y, where each x is like a y. Ten Shadows, for example, feels a lot less random when you learn that the thematic root is the Ten Sacred Treasures. Consider how many culturally canonized archetypes can you pull from to define your powersets.
Before we move on, let’s recap:
- Modal: powers are separate or exclusive
- Options: powers provide choices and tradeoffs
- Phases: powers rotate or alternate by a set pattern
- Graphs: powers transition by a complex logic
- Menu: powers are interdependent and overlapping
- Flow: powers advances or enables a procedure
- Branch: powers work indepedently yet synergize
- Fractal: powers builds a positive feedback cycle
- Metaphor: powers relates to an archetypal structure
- Mu-recursive: powers can modify their own logic
This list is my most obviously unfinished. Breaking into a rule of three rather than four feels like a failure of symmetry, and two I have failed to meaningful dissect into specifics at all. As much as I made fun of AW’s “meta powers,” I don’t think the all-encompassing broadness of the last two are an improvement at all.
We’ve climbed high enough on the tower of abstraction that I’m running out of air. Let’s user-glide-snow-thread-illness back down to earth.
← Epilogue: Ad Myriadem
You may assume, in this post titled “Recursive Magic Systems” which has spent nine thousand words examining magic system, that you have read a post about magic systems. But that’s not quite it.
What stood out to me was that, when writing about the procedural flow of powers, I recalled about the scheme for modelling nonhuman emotions I wrote about all the way back in . I thought about the model of character traits I’ve groped toward in some abandoned essays.. I already joked about how those nifty combining conjectives are exactly the sort of thing I’ve used as glue in every storytelling tool I’ve ever devised.
The only thing that made this essay seem to be about superpowers is merely that the word lists we pulled from had names like “flight” and “portals” rather than, say “betrayal” or “road of trials.”
The real magic of this recursive system is that it generalizes far beyond powers.
Ultimately, magic is a plot device. Put enough such devices together and there’s no better word for what results than “plot.” This means that, equivalently, a good magic power is no more and no less than a good story. Worlds are stories. Characters are stories. They have internal logic, themes that develop.
In a , I noted that I have fallen out of love with those rigid three-part abstractions that defined my early essays; I’ve since found flowing pairs more flexible and productive. This new system, I think, could be the fruition of that line of thinking.
Two ideas joined by a conjunctive, each recursively defined thusly… I find this a powerful formalism for expressing narrative ideas. But maybe you’re less interested in me generalizing the Curry-Howard correspondence to fiction than you are in just generating some cool superpowers.
I’ve programmed my own power generator based on these ideas. You can play with it here: Recursive Idea Generator. It’s still in a pretty unfinished state, but do find it fun to roll on.
Let me know if you actually end up with a library of a million spells.