How do I know the AI isn't just making things up?
Fair question, and the one I care most about too. I'm a GM and a builder, not an AI researcher, so I won't pretend to explain the model's internals. What I can tell you is how I test it: I've run it on more than a dozen real sessions across two campaigns, and because they're my sessions I've checked every brief against what actually happened. It's held up well on NPC dispositions, faction moves, and consequences, and it hasn't invented events that didn't happen. It's not perfect, it occasionally misspells a name or misses a small detail, which is exactly why everything is editable inline. No AI output should be trusted blind, including this one.
What are the file size limits?
File size is capped at 120MB per upload. A typical 2-hour audio recording is about 100MB. A transcript of that same 2-hour session is about 180KB.
That's the practical difference between the two upload paths. Audio is heavy and hits the cap quickly on longer sessions. Transcripts are light enough that session length effectively stops mattering: a 6-hour transcript is still small text, and a 40-hour campaign would still be small text. One upload is one session no matter how long it ran.
If your sessions consistently run long, the transcript path is the one I'd point you to. Zoom, Google Meet, Riverside, Fireflies, Otter, and most VTT platforms will export a transcript as a side effect of recording.
Why pay for this when I could just use ChatGPT myself?
Honestly, the way I think about it is taxes. Plenty of people do their own, and good for them. I don't, because doing it myself stresses me out and I'd rather pay to take it off my plate. You're not paying for access to a model, you're paying for the structure around it: transcription, a consistent brief format that tracks the same things every session, one-click application to your campaign, and the fact that it persists so you're not re-pasting your whole campaign history into a chat window every week. If you'd happily do that by hand, you genuinely might not need this. For me, the convenience is worth it.
Can I see real examples of what it produces?
Yes, and I'd rather show you the honest version than a cherry-picked one. I've published a full breakdown of a real session, the actual brief it generated, what it got right, and where I had to correct it. It's from my own homebrew Electric Bastionland game so it's all my material to share. Take a look and judge for yourself.
What happens to my audio and my players' voices?
Your uploaded audio is sent to OpenAI's Whisper for transcription, then deleted from our servers, we don't keep the recording. The text transcript and the brief are stored so you can use them. You can delete any session, which removes its transcript, brief, and player report, or delete an entire campaign, at any time. Brief generation runs through Anthropic's Claude. Since sessions usually include other people's voices, it's worth letting your table know their game is being transcribed. If you'd rather not upload audio at all, you can paste in a transcript instead.
Does it work with rules-light or heavily improvised games?
Yes, that's actually where I've tested it hardest. A lot of my own sessions are OSR and NSR games where the system is light and most of the session is improvised conversation rather than structured combat. The tool reads the session for what happened and who it happened to, so it doesn't depend on a particular ruleset. If anything, a loose improvised game is a harder test than a structured one, and it's held up.