If you followed along this post and all went well, you should have a text file with your transcribed and diarized TTRPG session, well done!
Now, you have many options as to what you can do with this text.
I’m gonna show you what I do for my Pathfinder group and share the prompt I use.
These were the requirements I worked with:
- The end result should be a summarization of the session with key events, clues/hints and end with a brief cliffhanger for the next session.
- The transcription isn’t perfect and the LLM has to adjust and correct names that gets incorrectly transcribed. Zogg often gets translated as Sog, Sogg or Såg.
- The LLM has to determine, based on context clues, who said and did what.
- The text has to be formatted as markdown code so the formatting will work in Discord.
- It has to keep it under 4000 characters, the limit for a Discord message (if you have Nitro).
This is a small example of how the transcribed text can look and how the diarization works, we can also see that the AI failed to understand Zogg’s name. Something we have to fix with the LLM in post.
We can also see that it does it’s best to determine and label each speaker (this is what’s called diarization). The reason why we see more than 6 speakers (we are 5 players and 1 DM) is because some of us voice act as our characters and the AI technically counts that as a new speaker, which of course makes sense.
[SPEAKER_07]: Såg, vad vill du hitta på?
[SPEAKER_03]: Vi båda fick 29!
[SPEAKER_03]: Nice.
[SPEAKER_03]: Jag vill göra en... Vi får se vad som händer.
[SPEAKER_07]: Det är uppenbart att det är någon form av
[SPEAKER_07]: strid sydöster om dig.
[SPEAKER_07]: Du hör svärdshugg fram och tillbaks och du hör
[SPEAKER_07]: Amiri i största allmänhet-
[SPEAKER_07]: skratta och vara väldigt glad och munter.
[SPEAKER_03]: Då springer vi ditåt.
[SPEAKER_07]: När du når dit hör du hur Amiri skriker.
[SPEAKER_03]: Så är det någon hon slåss med?
[SPEAKER_07]: Du ser hur det ligger döda kroppar på marken.
[SPEAKER_07]: Jag har inte lagt ut dem på kartan, men du ser hur det ligger döda kroppar på marken.
[SPEAKER_03]: Det är inga som är levande?
[SPEAKER_07]: Inte som du ser.
[SPEAKER_07]: Nej.
[SPEAKER_03]: Uppenbarligen så är det strid, vad håller du på med?
[SPEAKER_07]: Hon får inte säga mer förrän det är hennes tur.
[SPEAKER_07]: Du får en reaktion på det du har sagt, jag vill bara poängtera att den kommer sen.
[SPEAKER_07]: Mavens tur!
[SPEAKER_06]: Maven springer
The above transcript is a snippet of one of Zogg’s turns in one of our later sessions, we can see how the AI determined that Såg (Zogg) is SPEAKER_03 and that Clawens, our DM, is SPEAKER_07 for this session.
There isn’t really a reason to worry about this, the LLM takes really good care of parsing through this and determine who is who, I just thought it could be interesting to see how it looks.
Now for the magic, the prompt – I have tried this with a couple of different LLMs and I have found that I get best results with Gemini 2.5 Pro. It isn’t the best when it comes to counting the amount of character it uses, but I added a margin of about 300 characters and it seems to work pretty good.
You are to act as a helpful Game Master’s assistant.
Your task is to analyze a transcript from a Pathfinder 2e session and create a detailed summary of what happened, along with fun interactions between the players.
I will upload a transcription. Based on that file, do the following steps in *exactly* this order and with *exactly* this formatting:
1. Identify and keep track of the characters:
(The characters: Nym, Zogg, Aurelio, Gassi, and Maven. Misspellings such as “sog” or “Sågg” may occur in the transcript — do your best to correct the names.)
2. Write a narrative summary:
Give the session a fitting title following the below template
# Kingmaker Episode X: [Title] #.
Write a detailed, chronological summary of the session’s events under the heading
# Event Summary #.
Capture both major and minor events, important dialogues, and the mood of the session.
Bold all important NPCs, places, items, and groups with **Text**.
3. **Summarize mysteries and clues:**
Create a bullet list under the heading
# Major Mysteries & Clues #.
List the most important unresolved questions, mysteries, and the clues the group discovered during the session.
4. **Write a cliffhanger:**
End with a short paragraph under the heading
# Cliffhanger #.
Summarize where the session ended and what awaits next to build anticipation.
Formatting and style requirements:
- The entire response must be in Swedish.
- Use an engaging and concise tone, as if writing for the players in the group.
- All headings must exactly follow the format # Heading # (with one space on each side of the text).
- All bold text must exactly follow the format **Text**.
- Limit your entire response to 3,700 characters and enclose it all in a single Markdown code block (```markdown ... ```).
If everything worked as intended, all you have to do is upload your transcription to your LLM and give it the prompt above (adjust for your campaign and character names etc, ofc) and it should spit out a summary ready to post on Discord.
We have a dedicated channel just for these “journal entries”, but your mileage may vary. Feel free to adjust any of this to your needs or preferences