cros.land

Two months ago I wrote that the statblock generator no longer worked from templates. It does again. This post is the correction, and a short list of what changed for you.

Abilities belong to the same creature now. A charge that knocks the target prone and a trample that only works on prone targets. A grapple that leads into a smother. A funeral bell that weakens the dying and then finishes them. The parts of a generated monster set each other up, which is the thing the last version could not do.

Fewer combat roles. Brute, defender, skirmisher, artillery, controller, lurker, or Auto. Balanced and Support are gone. Swarm is no longer something you pick: say “a swarm of” in your concept and the generator builds a swarm. The same goes for oozes and dragons.

The name field takes a description. Type “a cliff troll that hurls slabs at climbers” and the generator names it. Type “Brazen Bull” and it keeps the name.

2014 monsters flip to 2024 instantly. A monster generated under 2014 rules is saved with both printings, so the edition tabs switch with no conversion step. Monsters you made before this still convert the old way.

What I said, and what happened

In July I said the generator computed its templates instead of drawing them from a shelf. The math was computed from the rulebook’s targets, and the AI invented abilities into empty slots. I also wrote, in a section called “The cost,” that a computed template has the numbers but none of a designer’s judgment about which abilities belong on the same creature, and that correct parts can still combine into something odd.

That is exactly what went wrong. The numbers were always right. The creatures were often not the ones you asked for. A brass bull came out as a thing that swallowed people whole and never charged. A mountain of troll-flesh came out with no regeneration. A stone shrine guardian came out with a genie’s wish-granting clause in its traits. Every ability on those monsters was fine on its own. They just did not belong to that monster.

What the new templates are

Not the old ones. The library I built in 2024 had three labels, and a template came with whatever abilities its source monster happened to have.

The new library has a hand-written template for each combat role and each way of playing it: a lurker that ambushes, a lurker that is a ghost, a lurker that is a trap; a brute that charges, one that grapples, one that regenerates; and so on. Each is written against the official monsters of that shape, and the same math as before scales it to any challenge rating.

The AI’s job got smaller. It names the creature, renames the abilities to fit your concept, and writes one signature ability of its own. When that one ability fails the checks, the generator substitutes a proven shape from the rulebook rather than retrying until something slips through. That happens about half the time, and I decided it was a feature. A dull ability is a better failure than the wrong monster.

What I gave up

Variety is bounded by the library again. Generate twenty lurkers and you will see the ambush template more than once, dressed differently each time. What you will not see is a lurker with a breath weapon, because no lurker in the book has one.

The AI has less room to surprise me, and it still does. The monster that harvests a skull from a fallen victim and throws it. The duelist whose riposte fires the instant a blade misses her guard. Those came out of the one ability the AI still owns, and they are the best things the generator has made. Less room turned out to be the point.

The test from July still holds: would this break a table, or would a Dungeon Master fix it in half a second without noticing? What remains is in the second category. What I got in exchange is a generator that gives you the creature you described. I did not have that in July, and I said I did. Hopefully this has been fixed. Please let me know if you notice anything.

The statblock generator now works in both editions of 5e, and there is a new tool for converting the monsters you already have.

Generate in 2014 or 2024 format. One switch, any challenge rating, and the output uses the right layout and phrasing for its edition. The 2024 cards have the official structure: merged immunities, the MOD and SAVE grid, saving throws written as Failure and Success.

Pick a combat role, or let it read your concept. Brute, defender, skirmisher, artillery, controller, lurker, support, and swarm. The role changes the actual math, so a lurker and a brute at the same CR are meaningfully different creatures.

Flip any saved monster between editions. Tabs on the statblock card switch between printings once both exist. Converting rewrites the wording only. Every number stays exactly where it was.

Convert statblocks you already have. The statblock converter is a new tool: paste a monster in, and it works out which edition it came from and converts it. Conversions count as generations against the daily limit rather than sitting behind a separate gate.

Export in both editions. Homebrewery markdown in the official 2024 frame style, Foundry VTT, and Improved Initiative. The dungeon, NPC, and encounter tools also generate statblocks in whichever edition you prefer.

The library I built by hand

For most of this tool’s life, the monster generator worked from a library I built by hand.

I went through the SRD one monster at a time and had ChatGPT strip each one down to its mechanics. Take a creature, remove everything that makes it that creature, and keep what is left: the armor class, the hit points, the attack routine, the number and shape of its abilities. Do that across the whole book and you end up with a library of generic templates sorted by challenge rating, which I labeled with the only three distinctions I bothered to draw at the time: stronger defense, stronger offense, and balanced.

Generating a monster meant picking a template that matched what you asked for and having the model reskin it. The AI never decided the numbers then either. They came from a template, which came from a creature a professional had already balanced, and that is why the output was mechanically sound.

The limitation was never the numbers. It was that every monster you got was shaped like whatever template came up. Ability slots were inherited along with the math, so if the template underneath happened to come from something with a breath weapon and an aura, you got a breath weapon and an aura, however they were dressed. The model was adapting abilities rather than inventing them. With a finite library and three labels, you eventually start seeing the same creatures wearing different coats.

Adding the 2024 rules is what made this a decision I could not put off. Every template in that library came out of the 2014 SRD. Supporting a second edition properly meant either doing the whole stripping exercise again against a different book, or building something that could produce a template on demand.

So I took the leap I had been avoiding. Instead of drawing a template from a shelf, compute one.

What a chassis is

The core of the new generator is what I ended up calling a chassis. Give it a challenge rating and a combat role, and it computes every number in the statblock before the AI is involved at all, along with a plan for what kinds of abilities the creature should have. Not which abilities. What kinds: a signature attack here, a defensive trait there, one nova effect at this damage ceiling, a reaction if the role wants one. Slots to be filled rather than abilities to be reskinned.

The math comes from the same place a designer’s does. The Dungeon Master’s Guide gives target values by CR for defense and offense, and the chassis derives from those: armor class, hit points, proficiency bonus, attack bonus, save DC, and a total damage budget for the creature’s turn. That budget then gets divided up. If the monster makes three attacks, each one is a third. Area effects get roughly half the single-target number, because an area effect that hits four people is doing four times the work. Creatures with legendary actions get a discount on their basic turn, because they act more often than once a round.

The combat role is where it gets fun. My old three labels became eight roles, and each one shifts the math the way official monsters actually differ from each other:

Brute: more hit points, lower armor class. Easy to hit, hard to drop.
Defender: high armor class, slightly reduced damage.
Lurker: fragile, low sustained damage, and a much larger once-per-fight nova.
Controller: less damage, higher save DC, because its job is conditions.
Swarm, artillery, skirmisher, support: each with their own shape.

This is the part that expanded what the generator can produce. A lurker and a brute at the same CR are genuinely different creatures now, with different survivability and different damage patterns, because the numbers say so before a single word gets written. Nothing is drawn from a shelf, so nothing is constrained to the shape of the creature a template was stripped from. Old monsters saved under the previous labels still map onto the new roles, so regenerating one lands where you would expect.

The AI’s job changed in kind. It used to adapt: here is a creature’s mechanics, make it something else. Now it invents into a container. It receives a filled-in skeleton that says, in effect, “your melee attack is +7 for 14 damage in two dice plus five, your area effect is a 21-point save for half, you have one trait slot and three action slots, one of which should be a once-per-fight nova,” and its entire job is to decide what those things *are*. What the creature is made of, what its abilities do, what they are called. Then assembly stamps the engine’s numbers back over whatever the model wrote, so even a model having a bad day cannot ship a CR 3 monster with CR 12 damage.

A side effect I did not expect: because every number exists before the first API call, the loading state can show you the real armor class and hit points immediately while the rest fills in. The architecture leaked into the interface in a nice way.

The cost

Here is the honest part. The template library gave up variety in exchange for guaranteed sanity. Generating the template gives up some of that sanity in exchange for variety.

A template stripped from a published monster carried more than numbers. It carried a professional’s judgment about which abilities belong on the same creature, even after the flavor was gone. A computed chassis has the numbers and none of that judgment. It is assembled from correct parts, and correct parts can still combine into something odd. The failure mode is not usually broken math, since the engine owns the math. It is coherence. A creature whose abilities do not quite belong to the same animal. A name that describes an attack routine instead of a creature. An elemental that reads like an undead.

Most of the month I’ve been working to close that gap.

Roughly a hundred and fifty checks

Every generated block runs through a linter before it ships. The rules are boring individually and powerful together: an area effect that deals damage cannot also stun without a recharge gate, a grapple must carry an escape DC, a restrained condition needs a way out, Prone never takes a duration because standing up is the exit, a damage rider is extra dice rather than a flat bonus, an entry cannot reference an ability that does not exist in the block.

Where I could, the fix is deterministic. The linter does not just flag a missing escape DC, it stamps the right one in. About half of everything the linter catches gets repaired without another API call. What remains goes back to the model as a short list of specific problems with the exact text to fix.

The discipline that made this work is a rule I set early: no check ships until it has been swept against all 329 monsters in the 2024 SRD without a single false positive. Any rule that fires on an official monster is wrong about D&D, no matter how sensible it sounded when I wrote it.

That killed a lot of my ideas, and the kills were more useful than the survivors. I wanted a rule that no monster should have four different energy damage types. The sweep found exactly one official monster that does, so the threshold moved. I wanted to flag raise-the-dead necromancy on creatures that are not undead. An official dragon casts Create Undead. I wanted to catch monsters that give themselves conditions they are immune to, which is a real defect, and the first version of that rule flagged the Air Elemental, because “pushed away from the elemental and knocked prone” puts the elemental’s name in the wrong grammatical position. Each of those taught me something about how the actual game is written that I would have gotten wrong from memory.

One judgment call

Deterministic rules only catch patterns you can describe. They cannot catch a cloud of spores that no ability ever creates, or a trait named Steam Shroud that produces no steam.

So every generated monster now gets one extra API call: an audit that reads the finished block and looks for problems a Dungeon Master would hit at the table. Can every effect be adjudicated? Does every imposed condition have a way out? Does anything contradict anything else? It also grades the monster’s name and flags a creature type that the content does not support.

The important part is that the audit has to show its work. Every finding must quote the exact text it is complaining about, verbatim. If the quote does not appear in the block, the finding is thrown away before it reaches anything. That one rule makes a nondeterministic judge safe to wire into a pipeline, because a hallucinated complaint costs nothing.

When the audit is right, its findings ride along with the linter’s into the same repair call, so a block with both kinds of problems is fixed in one pass instead of two.

Where I decided to stop

The generator still produces the occasional strange monster. A creature that would have been better classified as a monstrosity than an elemental. An ability with a condition that is technically always true. A name that is fine but not great.

I spent a while treating each of those as a bug, and eventually landed on a better test: would this break a table, or would a Dungeon Master fix it in half a second without noticing they had? Published monsters have oddities that fall in that second category all the time. That is the bar the generator is held to now, and it clears it.

What I got in exchange is worth it. The abilities coming out of the generator are dramatically more interesting than anything the template library produced, because they are being invented for a creature rather than inherited from one. The monster that eats a party member and drags them underground, the construct whose cracked bell forces it to ring, the swarm that shatters and reforms with half its hit points. None of those are variations on an SRD entry. They exist because the math was already handled and the model had nothing left to do except be imaginative about what a monster could be.