Who loves to play as yuan-ti (snakes)? This adventure makes a perfect 3 hour one shot! #dendar#merrshaulk
https://t.co/0uekDcPls3
[I'm working on a sequel]
#SelfPromoSaturday#dnd#ttrpg
The LLM should also summarize and provide this information as quick reference to a DM with auto-rolls.
Narrator's Cue
The fight breaks all at once. Six hobgoblins snap longbow fire into Cam. The Hell Hound lunges in with a crushing bite at Aria.
Rolls / Resolution
Six Longbow attacks vs Cam, 1d20+3 to hit, 1d8 + 1 piercing damage each. Sample rolls: to hit 9, 11, 20, 10, 15, 11; damage 6, 2, 9, 2, 9, 9.
Bite vs Aria, 1d20+5 to hit, 1d8 + 3 piercing damage on hit. Sample rolls: to hit 19; damage 11.
Revisiting the tactics model to compare #TTRPG GPT vs. rule-based tactics. I'm moving toward a hybrid model.
Step 1: Calculate line-of-sight and pathfinding for all creatures acting simultaneously.
Step 2: Consider cognition (mindless to tactical genius) and role (skirmisher, caster, brute).
Step 3: Score candidate group actions by alignment with tactical role and cognition. Penalize non-role/non-preferred actions.
Step 4: Use an LLM to supervise group action, selecting a top-10 candidate. Tip: Do not always select the very top candidate -> improvisation and built-in difficulty slider.
Here we are simulating group tactics for the Marauder Attack encounter from Red Hand of Doom.
-Hobgoblins attack from range and hold position.
-Hellhounds rush in.
#dnd#ttrpg#vtt
i added blocking and intelligent group tactics to TTRPG encounters!
- generates a table of candidate eligible group moves/actions by tokens using SRD stat blocks
- ranks moves/actions based on optimality (move into ideal ranged and attack vs. move into melee and attack vs. dodge/retreat
- implement tile reservations so tokens recognize same space
- supervisor ai improvises a group strategy from the list of candidates
- tactical decision reasoning is fully visible to GM ("Did this because"...)
- it is fun as holy hell to play
#TTRPG #VTT #indiegame
this is a fun map used for group ranged combat tactics involving 6 bandits and pillars obstructing line of sight
bandits move into positions and fire ranged weapon (ideally, they should keep moving behind pillar after firing- need to work on that)
paths show movement, hashed lines show ranged attack LOS
map courtesy Dyson Delves (Shrine of the Mosswater)
An AI goblin costs about $0.0003 per turn.
I built an open-source VTT where LLMs control monsters, so I benchmarked the cost of letting AI run combat.
Turns out goblin brains are surprisingly cheap.
I experimented with the rules-in- prompt at first, but that was unwieldy as you said. I took the rules and made a RAG system, then reserved the prompt for board state and stat blocks (ideally only changes to stat blocks should go in prompt, templates should remain in RAG). The output is structured as json to be interpretable as an updated board state. Like chess.
A current limitation is that I am not using blocking or object data, so open maps or open rooms are best.
@archon I tested on 5.5e SRD and MorkBorg, but it's intended to be interoperable and rules agnostic. It's fun to see the emergent behaviors of multiple monsters taking "optimal" tactical actions (explicitly suboptimal tactics is fun too ...).
@Nerdcognito The name "session zero" is the problem. It implies a bunch of admin tasks to be completed together with the group. If you end your session zero and don't play, everyone leaves disappointed. The buildup has to end in fighting, looting and roleplaying drinking at the bar.