You’ve got the transcript. Maybe Craig recorded the session and a transcription tool turned it into text, maybe your VTT spat one out, maybe you typed it up like a Victorian clerk. Either way it’s Sunday night, there’s a wall of text on your screen, and the move is so obvious it barely counts as an idea: paste it into ChatGPT and say “summarize this D&D session.”

And it works. Let me say that clearly before I say anything else, because I build a paid tool that lives one shelf over from this workflow and you should know my angle. Paste one session into a chatbot and ask for a recap, and you will get a decent recap. Ask it who took the deal with the smuggler, and it will usually tell you. For one session, in one window, on one Sunday, the free DIY version genuinely delivers. If your campaign were one session long, I’d have nothing to sell you.

Campaigns are long. That’s the entire plot of this post. Three things go wrong between week one and week six, they go wrong quietly, and each one lands on the same person: you.

Week one: the recap is good, and you did four jobs to get it

Count the steps you just did by hand. You exported the audio. You ran it through a transcriber. You read enough of the output to fix the worst damage, because the transcript came back calling your warlock’s patron “Sir Rathis” in one scene and “zero of this” in another. Then you pasted it, waited, read the summary, and fixed the two places where it attributed the ranger’s kill to the bard.

None of those steps is hard. That’s exactly what makes the trap comfortable. Each step is ten minutes, the whole pipeline is under an hour, and the result really is useful, so you do it again next week. You’ve volunteered for a recurring shift, and like most recurring shifts, it was easiest to accept in the week you had the most enthusiasm. Week one is always the week you have the most enthusiasm.

The names drift, and the chatbot papers over the drift

Here’s the failure that surprises people, because it hides inside a strength. Chat models are agreeable readers. Hand one a transcript where the lich’s name arrives as “Vel Karoth,” “Velkorath,” and “the whole karath,” and it will smooth all three into one villain without being asked. Convenient, except it has to guess which spelling is canon, and it guesses fresh every week, because next Sunday is a new chat window with no memory of last Sunday’s guess.

So session four’s recap stars Velkorath and session seven’s stars Vel Karoth, and six weeks later your notes doc contains two liches, or one lich under a name nobody at the table actually says. There’s no glossary anywhere in this pipeline. The model in the window can’t keep one, the transcriber doesn’t know your campaign exists, and the one place the canonical spelling lives is your head. Every proper noun the DM invents mid-sentence, and the DM invents them constantly, gets re-adjudicated weekly by a reader who wasn’t there.

We fight this exact problem at Loracle and I can tell you from the inside: the fix lives outside the transcript. You check each new mangle against a persistent record of the campaign, where Vel Karoth already exists as an entity with a spelling, so the next phonetic casualty gets reconciled to him before it can found a rival lich. The fix is memory. A chat window is the one place in this workflow that’s guaranteed to have none.

Every window is session one

Which brings us to the structural problem, the one that no amount of careful pasting gets you past. A chat conversation holds what you put in it. Your campaign’s best material lives between the sessions: the favor from session three that comes due in session eleven, the fisherman who’s been the duke since session two, the exact wording of the prophecy from the night half the table was giddy on cough syrup.

Your rogue’s Wednesday question is almost never about last session. It’s “do we still owe the smuggler, or did the barge thing cancel it,” and the answer is split across three transcripts in a folder. Now your options are: start a fresh chat, paste in all three (assuming you can tell which three), and hope the answer survives a context window that fits four hours of table talk the way an overhead bin fits a cello. Or do what every table without a record does: argue from memory and let the loudest version win.

Some of you will now invent the obvious patch, because you’re the kind of person who’s read this far: keep a running “campaign so far” document, paste that in alongside each new transcript. It helps. It’s also a fifth manual job, it grows every week, and you’ve just rebuilt a campaign wiki with yourself as the sync mechanism. At that point you should read what a campaign wiki is actually for, because you’re building one on hard mode.

The honest answer

Yes, you can paste the transcript into ChatGPT, and for some tables that’s the right call. Here’s the profile: your campaign is short or episodic, your questions are mostly “recap last session” rather than “trace this thread across the fall,” and someone at the table genuinely likes being the archivist. One-shots, six-session arcs, a West Marches table where every night is self-contained. For that table, DIY is free and fine, and I mean both words.

The table it quietly fails is the long campaign, where the questions compound and the pipeline runs every single week whether you’re tired or not. The money cost stays zero. The cost is an hour of your Sunday, a glossary that lives in your head, and a stack of summaries that can’t talk to each other. You weren’t quoted that price up front, which is the only dishonest thing about the whole workflow.

Loracle is the version of this where the pipeline is nobody’s job. It sits in your Discord call, you type /record, and after the session the transcript shows up speaker-labeled in whatever language your table plays in, the recap posts itself to your channel, and every session feeds one connected record of the campaign, the NPCs and debts and prophecies linked across every session they touch. When the smuggler argument starts, anyone types /chat and the answer comes back from the actual scene.

If you’re unsure which table you are, run the DIY pipeline for a month; it’s the cheapest possible experiment. Then notice which Sunday you start putting it off. That Sunday is your answer arriving.