Daggerheart works here, and there is exactly one thing to set up first

Nothing in the pipeline is rules-aware

This is the part that matters and it is easy to check. Transcription turns speech into text. Extraction pulls out proper nouns and sorts them into categories. Recaps summarise what was said. None of those steps reads a rulebook, knows what a statblock is, or cares whether you rolled d20s or Duality Dice.

So there is no Daggerheart support to switch on, and equally nothing that breaks. Your sessions transcribe, your NPCs and locations land in the wiki, the recap covers what happened, and you can ask questions about your own campaign afterwards. The same is true of Pathfinder, Call of Cthulhu and anything homebrew.

What the product does not do is also unchanged by system: no character sheets, no cards, no dice, no maps, and no speaker labelling — a transcript does not say who spoke.

Two prompts say "D&D", and it is worth knowing where

We would rather show you this than let you discover it. In two places the AI is told, in so many words, that it is looking at a D&D game.

The first is transcription. Every run is primed with a short glossary built from your campaign, and that glossary is prefixed with the sentence "A Dungeons and Dragons tabletop session." The prefix is there because it genuinely helps the model expect fantasy proper nouns rather than meeting vocabulary, and Daggerheart is a fantasy game, so most of that benefit carries. What it will not do is bias the model toward Daggerheart’s own terminology.

The second is entity extraction, which asks the model to pull named entities from a "D&D session transcription" and then — the important half — to categorise what it finds dynamically, based on context, rather than against a fixed D&D list. That is why a Daggerheart campaign does not come back with everything crammed into D&D-shaped buckets.

The glossary is the fix, and it takes one sitting

Each campaign has a glossary you control, and its contents are fed into transcription ahead of every session. Put your Daggerheart-specific proper nouns in it — the places, the factions, the NPC names, anything invented at the table — and they start coming back spelled consistently instead of phonetically.

It also takes pronunciation hints, which is the part most people miss. An entry can carry how the word actually sounds, rendered like "Wendelbrie (WEN-del-bree)", which is far more use to a speech model than the spelling alone. For a system where a lot of the vocabulary is invented rather than borrowed from D&D’s well-known bestiary, this is the single highest-value thing you can spend twenty minutes on.

There is a budget: the glossary that reaches the model is capped at around 850 characters, because the transcription API only honours roughly the last 224 tokens of a prompt. So it is a list of your most-confused names, not a dictionary. The Whisper guide explains where that limit comes from if you want the underlying detail.

Where the Hope and Fear part actually bites

A system that generates consequences through dice rather than prep produces continuity nobody wrote down. That is not a criticism of Daggerheart — it is the point of it — but it does mean the gap between "what we decided" and "what is written anywhere" is wider than in a module-driven campaign.

The practical effect is that the recap and the wiki are doing more work for you than they would for a group running published adventures, because there is no book to check when someone asks what the innkeeper was called. It also means the question answering has more to answer from: what you improvised in session four is in the transcript whether or not anyone remembered to write it down.