Play an AI Text Adventure Offline on Your Phone

You can play an AI text adventure offline by running a language model on your phone and giving it a game-master system prompt. The model describes a scene, you type what you do, and it narrates what happens next, like a choose-your-own-adventure book that never runs out of pages. Because the model runs on the phone, it works on a plane or anywhere without signal, and your story stays private.

A phone-sized model is a capable storyteller, but it has a short memory. The tips below keep long adventures on track.

What you need #

1. A local AI app that lets you set a system prompt for each chat. Personal LLM does, and it’s free on iOS and Android with no account.
2. A model that fits your phone. A 3B to 4B model is the sweet spot for storytelling on most phones. In Personal LLM, Qwen 3.5 4B and Gemma 4 E2B are good starting points; Gemma 4 E4B and Qwen 3.5 9B write richer prose on phones with more RAM.
3. About five minutes to write a system prompt, or copy one below.

Download the model on Wi-Fi before you travel. After that, no connection is needed.

How to set up an AI game master #

Step 1: Start a new chat for your adventure #

Keep each story in its own chat so its setting and rules don’t leak into your other conversations.

Step 2: Paste a game-master system prompt #

A system prompt is standing instructions the model follows for the whole chat. This one works well with small models:

You are the game master of a text adventure. Setting: a fog-bound coastal town in 1890 where ships keep vanishing. I play a young harbor clerk.

Rules:

• Describe each scene in 3 to 5 sentences, in second person (“You see…”).
• End every turn by asking what I do. Don’t act for me.
• Keep track of my inventory and remind me of it when relevant.
• Let my choices have consequences, including failure.
• Keep content suitable for a general audience.

Start with me arriving at the harbor office on a stormy morning.

Step 3: Pick a creative setting #

Stories benefit from a bit of randomness. Personal LLM’s Creative preset raises the temperature to 1.2, which gives more surprising descriptions and plot turns. If the story starts to ramble or contradict itself, drop to Balanced (0.7).

Step 4: Keep thinking mode off #

Thinking mode makes the model reason step by step before answering. That’s great for puzzles and slow for storytelling. Leave it off for a snappier game.

Step 5: Play #

Type actions in plain language: “I check the ledger for the missing ships’ names,” “I ask the old sailor about the lighthouse.” Short, specific actions give the model a clear next beat.

More system prompt ideas #

Swap the setting paragraph to change genre:

• Space mystery: “A mining station orbiting Jupiter. The station’s AI has gone silent. I’m the night-shift engineer.”
• Fantasy quest: “A kingdom where magic is fading. I’m an apprentice cartographer sent to map the last enchanted forest.”
• Detective noir: “Los Angeles, 1947. I’m a private investigator hired to find a missing jazz pianist.”
• Survival: “A small island after a shipwreck. I have a knife, a flare and a torn sail.”
• Cozy: “I’ve just inherited a run-down bakery in a village where everyone has a secret.”

For a more game-like feel, add rules such as “Give me three numbered choices at the end of each turn, plus the option to do something else,” or “Track my health from 10. Dangerous actions can cost points.”

We keep a broader set of reusable prompts in system prompt examples.

Why does the AI forget what happened earlier? #

A language model can only take in a limited amount of text at once, its context window. In a long adventure, the early turns eventually fall outside what the model can see, and it may forget a character’s name, an item you picked up, or a promise it made. On a phone, the practical window is smaller than on a big cloud service, because a larger context uses more memory.

Fixes that work:

1. Keep a “story so far” note. Every 15 or 20 turns, ask: “Summarize the story so far in 5 bullet points, including my inventory and key characters.” Copy it.
2. Start a new chapter. Open a new chat with the same system prompt and paste the summary as your first message.
3. Put permanent facts in the system prompt. Your character’s name, the main villain, the world’s key rules.
4. Correct it gently. “Remember, I already have the brass key” usually gets things back on track.

In Personal LLM, the chat header shows how much of the context is used, so you can see when a chapter break is due.

How good are local models at storytelling? #

Better than you might expect for short scenes, and weaker over long arcs. A 4B model writes vivid descriptions and reacts well to what you do. It’s weaker at long-term plotting, keeping many characters straight and planting clues that pay off later. A 9B model does noticeably better if your phone has 6 GB of RAM or more.

Small models also tend to push the story forward on their own, or let you win too easily. Rules like “Don’t act for me” and “Let my choices fail” help.

Offline AI vs AI Dungeon and other cloud apps #

Cloud storytelling apps such as AI Dungeon run their models on servers. That usually means larger models, longer memory features and extras built for play, and it requires a connection. Your stories are stored on the service’s servers.

If you mostly play on commutes and flights, or you’d rather your stories stay on your phone, a local model is a good fit. If you want the longest, most coherent campaigns, a cloud app with bigger models may suit you better.

Tips for a better game #

• Set the tone in the first message. “Make it eerie and slow-paced” changes the whole feel.
• Use the regenerate button when a turn falls flat. You’ll get a different take on the same moment.
• Edit your action instead of arguing with the story. Personal LLM lets you edit and regenerate.
• Play in short sessions. Long sessions warm the phone and fill the context. Our post on AI and battery drain has more.
• Try a different model for a different voice. You can switch models mid-chat.

If you’re picking an app first, our roundup of the best offline AI apps compares which ones support custom system prompts.

Frequently asked questions #

Can I play AI text adventures without internet? #

Yes. Download a local language model on Wi-Fi, give it a game-master system prompt, and play in airplane mode. The model runs on your phone’s chip.

What’s the best model for AI storytelling on a phone? #

A 3B to 4B model is the best balance of speed and quality on most phones, such as Qwen 3.5 4B or Gemma 4 E2B. If your phone has 6 GB of RAM or more, a 9B model writes more coherent long stories, at slower speed.

Why does my AI story forget details? #

The model can only see a limited amount of recent conversation. Summarize the story periodically, start a new chat with the summary, and put permanent facts in the system prompt.

Is an offline AI text adventure private? #

Yes, if the app runs the model locally and has no server. The story never leaves your phone. Test by playing in airplane mode.

Can local AI play a character, not just narrate? #

Yes. Tell it in the system prompt: “You are Captain Reyes, a gruff smuggler. Stay in character and speak only as her.” Small models hold a character well for a few dozen turns before needing a reminder.