The complete guide to AI storytelling Every reply is a collaboration.

What AI storytelling is

AI storytelling is telling a story together with an AI, one turn at a time. You decide what your character does or what should happen next; the AI writes the scene that follows, with its places, its other characters, and its consequences. Then it’s your turn again.

It takes a few familiar forms:

• Interactive fiction and text adventures, where you type any move and the world answers. See AI interactive fiction.
• Roleplay, where you play one character and the AI plays everyone else. See AI roleplay and the beginner’s guide to roleplaying with AI.
• Story generation, where you start from a premise and play the whole story out. See the AI story generator.

In every form, the AI is only half of the system. The other half is everything around it: what the app chooses to show the AI before each reply, and what it keeps from one session to the next. Most of the difference between a forgettable chat and a story you return to for months comes from that second half.

How a reply is made

A language model writes by predicting text, and it can only predict from what it is given. It keeps no memory of its own between replies. So before every reply, the app assembles a prompt. On LoreWeaver, a prompt holds:

1. The storyteller’s instructions: how to write, what never to decide for you, and how to treat canon.
2. The world: its premise, tone, and style rules.
3. Your character and notes: the persona you play and your session notes.
4. Lore cards: the pinned ones, plus any whose title or tags appear in recent messages.
5. The recap: a running summary of the whole session.
6. Memories: notes on past events that match the moment.
7. Recent turns: the latest stretch of the story, word for word.
8. Loom directions: events a world’s rules decided should happen now.
9. Your new turn.

All of it is measured in tokens and has to fit the context window, along with room for the reply. On LoreWeaver, the free plan allows up to 6,400 tokens of context and Legendary up to 100,000. The context inspector under the composer shows exactly what went into the last reply, and context windows explained covers how the space is shared.

One turn, traced from start to finish

Here is what happens to a single line in a world set in the river city of Vell:

1. You write: “I show Captain Marrow the clapper I pulled from the river.”
2. Lore wakes up. “Marrow” is a tag on Captain Ilse Marrow’s card, so her card joins the prompt: her loyalty to the church, her clipped way of speaking, the brother she never mentions. “Clapper” is a tag on the Drowned Bell’s card, so the Bell comes too.
3. Memories are found. The search pulls up the night the Bell rang and the evening Marrow warned you away from the docks, because they share names and meaning with your line.
4. The rest is packed. The premise, style rules, your persona and session notes, the recap, and as many recent turns as fit join the prompt, inside your plan’s context.
5. The Loom checks its threads. If a thread is waiting for the Bell’s card to come up, it fires now and can add an event, such as the watch knocking at the door.
6. The storyteller replies. Marrow goes very still, asks where you found it, and says nothing about her brother, because her card says she never does.
7. The story updates. A few replies later, a memory records that you showed her the clapper, and the recap grows by a line.

The context inspector shows steps two through four for the last reply: every card and memory used, and the words that triggered each card.

Why long stories drift, and how they hold

Without help, a long story slides out of the window. The model stops seeing the first scenes, then forgets a name, then contradicts a wound or a promise. Bigger windows delay that moment but don’t remove it.

LoreWeaver holds a story together with layers that each catch something different:

• Memories are written every few replies from what just happened and found again by meaning and by the words and names they share with the moment. Each plan sets how many a story can hold, and the largest plans have no limit.
• The recap keeps the whole arc in a few paragraphs and is rebuilt as memories grow. You can read and edit it.
• Lore cards hold what you wrote ahead: people, places, factions, items, and events. They step in when named.
• Loom state remembers flags, quests, and inventory in worlds that use them.

These layers follow your edits, too. Rewind or undo a scene and its memories, recap changes, and Loom effects go with it; branch a story and the branch begins from what was true at the fork. How AI memory works goes deeper.

Writing turns that get better replies

The model mirrors what it’s given, so your turns shape every reply. These habits matter most:

• Be specific. “I slip the forged letter under the bishop’s door” gives the storyteller far more to work with than “I go to the church.”
• Write the attempt, not the outcome. Decide what your character tries; let the story decide whether it works.
• Set the tone by example. Terse turns tend to get terse replies, and lyrical turns get lyrical ones.
• Name names. Names wake lore cards and memories; “she” doesn’t.
• Give the scene friction. A want, a secret, or a deadline gives every reply somewhere to go.
• Use the right input mode. Do for actions, Say for dialogue, and Custom for narration or directions you want passed through exactly.

Building the world around the story

A world is the part of the prompt that stays put from reply to reply, so it pays to build it well:

• Premise: the setting, its history, and the conflict in motion. Keep it tight, because it’s read every time.
• Tone and style rules: the mood, plus explicit instructions such as tense, point of view, reply length, and what to avoid.
• Openings: one or more starting scenes players can choose from.
• Lore cards: everything else, written as cards that appear only when they’re named.

Start small: an opening location, three to five characters, and the problem in front of the player. Add cards as the story discovers what it needs. LoreMaps draws how your cards relate, and on Plus and up, Emergent Lore Detection suggests new cards from the story itself. The worldbuilding guide and lorebooks explained cover both halves in depth.

Directing the story: notes, the Loom, and choices

Sometimes you want more than a good next reply; you want the story to bend a certain way.

• Session notes in the Author drawer are read on every turn, up to 500 tokens. Use them for tone, pacing, current goals, and secrets the storyteller should keep.
• The Loom is LoreWeaver’s no-code IF→THEN engine. A thread waits for its conditions, such as a keyword, a number of actions, a flag, a lore card coming up, or a dice roll, and then acts. It can inject an event or a line of dialogue; switch, pin, or rewrite lore cards; set flags and counters; show meters, quests, and inventory; award achievements and hints; or offer tappable choices.
• Hidden threads keep a published world’s twists out of sight while still firing for every player.

A classic first thread: once the player has taken ten actions without the cellar key, the innkeeper “accidentally” leaves it on the bar. It’s the kind of quiet authorship that makes a world feel designed. The Loom guide lists every condition and action.

Choosing a storyteller and settings

The storyteller is the AI that writes your replies, and LoreWeaver offers several with different strengths, such as dialogue, dark fiction, intricate plots, or speed. Every plan includes unlimited replies from the free storyteller. Premium storytellers have a daily allowance by plan, and on Legendary premium replies are unlimited.

Context is the most a plan allows; some storytellers work with less, and LoreWeaver uses whichever limit is smaller. Switching storytellers mid-story can change the voice, so pick one for a long session and stay with it. Ultra adds every frontier storyteller. Full details are on the plans page.

Generation settings tune how replies are written: temperature (randomness), top-p and top-k (how many word choices are considered), maximum reply length, and repetition penalty. The presets (Balanced, Cinematic, Creative, and Chaos) set them together, from steady to wild, and Cinematic writes the longest replies.

Advanced techniques

• Branch at forks. Before a big decision, branch the story, play both paths for a few replies, and keep the one you like. The branch tree view keeps track.
• Regenerate with intent. If three versions miss the same way, the cause is in the prompt: your turn, a lore card, or your notes. Fix the cause, not the reply.
• Plant secrets in lore cards. A card can hold a truth the characters don’t know yet, with a style rule asking the storyteller to reveal it slowly. A Loom thread can swap the description once the secret is out.
• Schedule payoffs. A Loom thread can fire another after a set number of actions, which is how a storm promised early arrives on time.
• Give every major character a voice. Two lines of sample dialogue on a card do more than a list of adjectives.
• Edit the recap after turning points. A sentence of your own keeps the arc straight.
• Add the senses. Narration with text-to-speech, voice calls with any character, and on Pro and up, scene illustrations and an adaptive soundtrack.