Most GMs I talk to have a browser-history problem. ChatGPT in one tab asking about NPCs, Claude in another asking about tone, a map generator quietly buffering, Dungeon Alchemist minimized from three sessions ago, and somewhere a Google Doc called "campaign notes final FINAL" that you'd love to find right now if you could remember which folder past-you put it in. You open a fresh chat, type "help me prep for tonight's session," and already the cursor is judging you.
The tools aren't the problem. The problem is that we all ended up with four AI tools and no plan for how to use them together. Every prep session turns into a small scavenger hunt: which chat had the faction notes, which one has the NPC voice you liked last week, why does this one keep suggesting you add a mysterious hooded stranger.
The GMs who get prep down to under an hour aren't using a secret tool. They're using the same tools you are. They just stopped asking each one to do everything and started giving each one exactly one job.
That's what this post is about. Not a tool review of AI dungeon master tools. We already wrote that for 2026. This is the workflow: a three-phase AI DM prep stack with the prompts, the tool assignments, and the time estimates, so you can copy the structure and tweak it for your table.
Why is your prep still taking three hours?
Real talk: the reason your prep takes three hours isn't that AI is slow or that you're bad at it. The reason is that you're making every prep decision twice. Once when you're thinking about what to prep, and again when you're asking the AI for it. Every switch between "what do I need?" and "get me that" burns context.
One of our playtesters described their old workflow almost perfectly: "My workflow consisted of switching between ChatGPT, Gemini, and Claude Sonnet. Plus also having an instance of Perplexity all running at the same time." When they needed to know how an NPC would feel about a conversation, the round-trip went: Perplexity space A for NPC lore, Perplexity space B for the boss backstory, Claude for tone, then Obsidian to stitch it together. "That's so very time-consuming."
They're not unusual. Most GMs end up with a spiderweb of half-used chats because nobody taught us how to sequence this work. The fix isn't a new tool. The fix is splitting prep into three phases and giving each phase a narrow, bounded job.
Here's the split:
- Phase 1 — Pre-session (45 minutes, day of or night before): pull last session's recap, decide the 3-5 things you'll actually need, generate them.
- Phase 2 — At the table (0 minutes of new prompting): AI chats are pre-opened with context loaded. You only tap it for fast, bounded tasks.
- Phase 3 — Post-session (15 minutes, right after): capture what happened so next week's Phase 1 starts with a recap and not a blank page.
That's the whole stack. The rest of this post is how to run each phase without fighting the tools.
How long does D&D session prep actually take?
Before we dig in, a reality check on what "normal" looks like. Sly Flourish ran a survey of 3,663 DMs asking how long their prep takes. The answer: 43% come in under an hour, 28% hit somewhere in the 1-2 hour range, and the remaining ~29% spend three hours or more per session.
The under-an-hour crowd isn't prepping less. They're prepping smarter: a handful of bullet points, a few names, a rough map, and trust in improvisation for everything else. The three-hour crowd is usually the one rewriting backstory for NPCs the party will never meet.
AI doesn't magically drop you into the 43%. What it does is shift the ceiling. If you're already prepping in 90 minutes, a sequenced AI workflow puts you reliably under 45. If you're doing three-hour marathons, the same workflow lands you around an hour, but only if you pair it with the pre-commit discipline we'll walk through in Phase 1. No tool survives contact with "prep everything in case."
Phase 1: The 45-minute pre-session block
Before you type anything into any AI, do this: write down (on paper or in a note) the 3 to 5 things you are confident your players will interact with tonight. Not might. Will. The NPC they're already hunting. The room they're definitely entering. The faction they're about to piss off.
This is the whole game. If you don't pre-commit, you'll spend 20 minutes generating six NPCs for a tavern the party walks past.
Now run the phase in this order:
Step 1: Recap retrieval (5 minutes)
You need last session's state loaded into your own head before you generate anything. This is where a campaign-memory tool pays for itself. If you already ran Phase 3 last week, open your recap note and your "unresolved hooks" list and read them. If you didn't, this is where you'll pay the hour back in panic later.
Our playtester's workflow here was brutal before they had memory tooling: "I'll take that information [after session] and feed it back into Perplexity and say, how does this character internalize this information and how does it affect what they do moving forward. And then that output gets pulled back and added to the character profile in Obsidian." Beautiful process. Also takes forty minutes.
If you have a RAG-powered campaign library (this is the core of ScriptoriumGM and what Archivist and StormScape do for recaps), just ask the Campaign Assistant: "What were the unresolved hooks from last session, and which NPCs did the party interact with? Pull from my session notes." You read the answer. You move on.
If you don't have one, open last week's notes, skim them in five minutes, and move on. Do not try to make the AI re-read your whole campaign from a cold context window. It will hallucinate and you will chase it.
Step 2: NPC generation for the 3-5 scenes you committed to (15 minutes)
Here's the prompt I actually use, and the structure matters:
I'm running a D&D 5e session tonight. My party is Level 6. Tonight they will interact with: (1) Corvus, a Shadow Guild fence they've met before — voice is already established, I just need a scene goal for him; (2) a new contact named Mirela, a dock worker who knows where the contraband left port; (3) a city watch captain who's suspicious but corruptible. For each: one sentence on what they want from this scene, one sentence on what they know, one concrete mannerism, one thing they'd say in a stressful moment.
Four constraints on that prompt that make it work:
- It names the NPCs and their roles. No "generate three NPCs for a port city."
- It specifies what you need per NPC: not a backstory, a scene function.
- It asks for a mannerism and a stress line, because that's what you'll actually use at the table.
- It caps the output by asking for sentences, not paragraphs.
Claude or ChatGPT Pro both handle this fine. For bulk scaffolding (stat blocks for mooks, loot tables, tavern patrons) a dedicated generator like LitRPG Adventures is faster because the output is already formatted for a sheet.
Another playtester, described their experience with the Scriptorium Campaign Assistant well when the AI is context-loaded: "You've got cluttered cobbled streets and you can hear merchants yelling out selling their wares. It picks it up. It's like, you have a merchant selling their wares — do they need a stat block?" That kind of inline offer only happens if the AI already knows the scene. Which is why step 1 matters.
Step 3: Encounter scaffolding (15 minutes)
For the one combat or skill challenge you're sure will happen, ask for scaffolding, not a finished encounter.
Level 6 party of four (fighter, cleric, rogue, sorcerer). They're breaking into a warehouse guarded by Shadow Guild enforcers. Give me an encounter skeleton: 4-6 enemies with CRs that add up to Hard, one environmental element (not generic crates), one mid-combat complication that could trigger on round 2-3, and one thing the rogue specifically could exploit. Don't write narration — just the bones.
You'll do the narration at the table. You just need the mechanical skeleton built so you're not rolling up stat blocks during initiative.
For maps, Dungeon Alchemist is still the fastest path from "I need a warehouse" to "here's a grid-lined warehouse I can paste into Roll20." Generate it while the NPC prompt is running in the other tab.
Step 4: Load, don't chat (10 minutes)
Last step of Phase 1: take every useful thing you generated and drop it into the place you'll actually read it during play. Not a chat window you'll close. A note, a one-page prep sheet, a ScriptoriumGM note, an Obsidian page. Wherever your eyes go at the table.
If you're using ScriptoriumGM, this is what the Campaign Assistant's note-creation and drag-and-drop linking does for you: every NPC becomes a searchable note, and you can drag them into tonight's session note so it all lives in one place. If you're on ChatGPT or Claude alone, copy-paste into a doc. Do not plan to "just open that chat again at the table." You won't find it.
Time check: 45 minutes, end to end, if the scenes were pre-committed.
Phase 2: At the table — the AI should be invisible
This is the phase most GMs get wrong. They've got AI open, it saved them an hour in prep, so they reach for it every time something surprises them. Forty seconds later they're typing "generate a dwarf merchant quickly" while four players watch them type.
Don't do that. The rule is: AI is for bounded tasks under 30 seconds, or it stays closed.
What stays open at the table:
- Your prep sheet or session note (the one you built in Phase 1).
- A stat block reference. D&D Beyond, the SRD, your PDF, whatever.
- One AI chat, pre-seeded for quick Q&A only. Priming prompt: "I'll ask short questions during play. Answer in 1-3 sentences. No preamble." Keep it minimized. Surface it only for the 30-second asks below, never as your prep doc.
What the AI handles at the table:
- Loot table fills. "Give me 3 unremarkable items a warehouse guard would have on them." Ten seconds. Done.
- NPC reactions when players talk to someone off your prep sheet. "Mirela was just accused of lying by the party paladin — one line of how she reacts, staying in character." Fifteen seconds.
- Rules adjudication, only if your table is cool with it. "Can a grappled creature cast a somatic spell in 5e 2024?" Faster than flipping the book.
What the AI does not do at the table:
- Generate a new encounter from scratch. You don't have time to read and vet it.
- Write narration. If you're reading AI prose at your players, they can feel it.
- Answer open-ended "what should happen next?" questions. That's your job.
Move the heavy AI work out of the session and into before and after. That's the move. The table is for playing. The prep stack exists so that the AI can disappear.
Phase 3: Post-session (15 minutes, before you close the laptop)
This is the phase that compounds. Every fifteen minutes you spend here is forty-five minutes you don't spend in Phase 1 next week.
Three things to do, in order:
1. Dump the raw session (5 minutes)
Either transcribe your recording (Archivist AI and StormScape both do this from Discord; Whisper does it for free if you recorded locally) or just bullet-point what actually happened. Five bullets is enough. "Party infiltrated the warehouse, killed two guards, captured Corvus, learned that the shipments are going to Neverwinter, skipped the dock scene entirely, ended in the Shadow Guild safehouse."
If your party skipped something you prepped, write that down too. Half of next week's prep is deciding whether to rethread the cut content or let it go.
2. Extract entities and hooks (5 minutes)
Ask your AI (or your campaign-memory tool): "From these notes, extract: new NPCs introduced, locations visited, items acquired, unresolved questions the party asked, and any promises I made to the players that I need to follow up on."
That last one is the killer. "Promises I made" is where campaigns silently die. The rogue said "I want to come back here and steal the ledger," and you said "noted" — if that never shows up in a prep note, it never happens, and three sessions later the player feels unheard. This is the single highest-leverage question in the whole stack.
3. Write the recap (5 minutes)
Three sentences. Maybe five. Post it to your Discord before you close the tab. This is next week's Phase 1 Step 1 pre-built.
If you use ScriptoriumGM, this is where the Campaign Assistant's note-creation does the compounding. The recap becomes a searchable note, and the NPCs and hooks are auto-linked so next week's prep starts with context already loaded. That's the whole pitch: Phase 3 in minute 15 is Phase 1 Step 1 in minute 0. Seven days apart.
The tool assignment table
Here's the mapping. Match your workflow to your tools, not the other way around.
| Phase | Task | Tool | Time | Why this tool |
|---|---|---|---|---|
| Pre | Last session recap | ScriptoriumGM Campaign Assistant, Archivist AI, or your own notes | 5 min | Needs campaign memory; generic AI can't do this cold |
| Pre | 3-5 NPC scene sketches | Claude Pro or ChatGPT Pro | 15 min | Narrative voice, tone control |
| Pre | Stat blocks, loot tables, mooks | LitRPG Adventures | 5 min (inside the 15) | Pre-formatted, system-aware output |
| Pre | Battle map | Dungeon Alchemist | 10 min | Fastest path to VTT-ready grid |
| Pre | Load everything into one prep doc | Obsidian or Google Docs | 10 min | The note is what you'll read, not the chats |
| Table | Loot fills, NPC reactions, rules | Pre-seeded ChatGPT or Claude chat | <30 sec each | Only if bounded |
| Table | Visuals | Dungeon Alchemist + your VTT (see our VTT comparison) | 0 min | Pre-loaded in Phase 1 |
| Post | Transcription | Archivist AI, StormScape, or Whisper | 5 min | Automated from recording |
| Post | Entity + hook extraction | Campaign Assistant or Claude | 5 min | Needs memory or the raw transcript |
| Post | Written recap | Any AI, or you | 5 min | Speed matters more than polish |
The $0 workflow vs the $25/month workflow vs the full stack
Not everyone needs everything. Three realistic builds:
The $0 stack. ChatGPT free tier for NPCs and encounter scaffolding, Whisper (free, local) for transcription, a Google Doc as your prep sheet, your own brain for campaign memory. This works. It's what I ran for two years. The friction is Phase 3: extracting hooks and entities by hand takes the full fifteen minutes, and you'll lose details across sessions. Good for short campaigns or one-shots.
The $25/month stack. Claude Pro ($20) for the narrative and extraction work, LitRPG Adventures ($5) for generators, Dungeon Alchemist ($44.99 one-time, amortize it). You've now got professional-grade models for Phase 1 and Phase 3. You still don't have real campaign memory. You're pasting context into Claude conversations, which is fine for a 10-session arc but gets unwieldy past 20.
The full stack ($40-55/month, long-campaign territory). Add Archivist AI ($10) or StormScape ($10-30) for automated recaps, or ScriptoriumGM ($10-30) as the campaign-memory layer that feeds Phase 1 and captures Phase 3. This is the build for GMs running weekly campaigns that are 30+ sessions deep, where remembering "what did we establish about the Shadow Guild in session 14?" actually matters.
What the AI should never do
Quick guardrails, because the workflow only works if you keep the human jobs human:
- AI doesn't read the room. When your rogue's player has been quiet all night because something's off, that's your job. No tool sees that.
- AI doesn't make thematic calls. Whether this session needs to end on dread or on laughter — your call.
- AI writes first drafts, not final ones. Every NPC line it gives you is a starting point, not a script. One of our GM's put it plainly: the AI "creates the narrative structure that those ideas get placed in." You still bring the ideas.
- AI doesn't own consequences. If you promised the party something in session 4, you owe it to them in session 11. The tool helps you remember. It doesn't care.
- AI doesn't replace session zero. Talk to your players about AI in your prep. Most won't care. Some will. Knowing before session 12 beats finding out.
Another ScriptoriumGM summed up the honest version of this debate: "I use it for content, but I didn't make anything from it. I ran ideas. That's the point." The AI is the sparring partner. You're still the GM.
Building the stack for your play style
A few quick adjustments depending on what you actually run:
Weekly campaign GMs. The full Phase 3 loop is worth it. Every fifteen-minute post-session block saves you forty-five minutes in Phase 1 seven days later. That's 30 hours per year of prep reclaimed, minimum. This is the build campaign-memory tools are made for.
One-shot and convention GMs. Skip Phase 3 entirely. Your Phase 1 is all front-loaded. Run the 45-minute pre-session block twice as deep: once the week before for structure, once the day of for NPC refresh. At the table, same rules apply: AI stays closed unless it's a bounded 30-second ask.
West Marches and open-table GMs. Campaign memory is non-negotiable. You're running prep for parties you might not see again for six weeks. A RAG-powered library that remembers which NPCs which party interacted with is the only reason this format is sustainable without melting your brain.
Whether you're running D&D 5e 2024, Pathfinder 2e, or Call of Cthulhu, the phase structure is the same. Only the stat block generators and rules lookups swap out. The workflow is system-agnostic; the prompts are the part you tune.
Frequently asked questions
Can I really prep a D&D session in under an hour with AI? Yes, but the constraint isn't the tools. It's pre-committing to the 3-5 scenes you'll actually run before you open any AI. Once you've done that work on paper, the three-phase workflow (45 min pre-session, 0 min at-table, 15 min post-session) keeps total AI-assisted prep under an hour for most sessions. The GMs who fail at this are the ones who skip the pre-commit step and ask ChatGPT to "help me prep" from a blank mind.
What's the best AI tool for D&D session prep in 2026? There's no single best tool. Each phase has different needs. Claude or ChatGPT Pro for NPC scene sketches and entity extraction. LitRPG Adventures for stat blocks and loot tables. Dungeon Alchemist for maps. Archivist AI, StormScape, or ScriptoriumGM for campaign memory and session recaps. Match the task to the tool. See the tool assignment table above, or our full AI DM tools buyer's guide for 2026.
How do I use ChatGPT for DM prep without the output sounding generic? Two fixes. First, pre-seed the chat with your campaign context: the NPC names that already exist, the faction your party is tangled with, tonight's three committed scenes. Second, constrain the output format. Ask for "one sentence on what they want, one mannerism, one line they'd say under stress" instead of "a backstory." Specific scene-function prompts beat open-ended worldbuilding asks every time.
Does AI replace the GM? No, and the workflow above never asks it to. AI handles the admin: recap extraction, mechanical generators, NPC scene sketches, map grids. The GM owns what actually matters. Pacing, emotional beats, and the read on a player who's gone quiet for the wrong reasons: none of that survives being delegated. The tool works because it stays invisible at the table.
Try this at your next session
Pick one Phase 3 habit and run it this week. After your next session, before you close the laptop: write five bullets of what happened, and ask your AI to pull out entities and hooks. Save the file. Name it session-[number]-recap. That's it.
Next time you prep, open that file first. Notice how your 45-minute block starts with you already knowing where you are.
That's the whole stack. Pre-commit what you need. Match tools to tasks. Keep AI bounded at the table. Close the loop after the session so next week's loop is half-built already.
If you want the campaign-memory layer that makes Phase 3 turn into next week's Phase 1 automatically (the note-writing, the entity linking, the searchable campaign library), that's what we're building. You can set up a ScriptoriumGM campaign in about five minutes and have the Campaign Assistant running against your own notes by tonight.
What does your prep stack look like? Drop into our Discord and share. The GMs who've gotten their prep under 45 minutes all did it slightly differently, and we're always swapping workflow tweaks.