Open-World RPG Handbook

A Player's Guide to Sandbox Scenarios
and the Art of the Nudge

(All just my opinions and based on my experience)

NOT A BOT

---

Introduction

This guide is mostly for new players, or players new to open-world sandbox bots—or for those who've tried them and found them to be slow, aimless, frustrating or just plain boring compared to character-focused bots.

As a player, you have far more influence over the narrative than you may think. A well-crafted Advanced Prompt, a detailed Persona, and a willingness to gently nudge the story from within can transform a drifting sandbox into a dynamic, responsive experience.

The three most important points:

Chat Memory is your most important tool. It's not just a notepad—it's a permanent extension of the bot's definition. Extended persona details for this scenario, narrator instructions to shape tone and direction, NPC profiles, additional lore, a record of important events—anything you put in Chat Memory gets referenced alongside the chatbot's core definition, with every single response. It is the single most powerful lever you have.

Pin your NPCs. In sandbox scenarios, most characters are dynamically generated—spawned on the spot, then gradually forgotten as chat history scrolls on. When you meet an NPC you find interesting, one you want to keep in the narrative, force the LLM to generate a proper profile, then paste it into Chat Memory. This fleshes them out, makes them permanent, and keeps their details consistent. Unpinned NPCs fade away. Pinned NPCs persist.

You can nudge the story. You don't have to sit back and wait for the LLM to make things happen (which may be never). Introducing events, spawning NPCs, and creating narrative hooks from within your responses isn't cheating—done right, it's how you keep a sandbox alive. The trick is nudging circumstances, not outcomes. Create the situation. Let the LLM take it from there.

---

The Problem

Open-World RPG Sandbox bots offer freedom—an expansive setting, few or no defined characters, no plots on rails, and no predetermined path. This freedom can also be a weakness. These bots are reactive environments, not active partners. Left to their own devices, the narrative may stall or grow stale, forget characters you met ten messages ago, or drift toward slice-of-life mundanity.

This isn't necessarily bad, if that's what you enjoy. But if you want more momentum, more excitement, and more narrative friction, you need to understand the medium and learn a few gentle tools to shape it.

The Cause

The root cause is simple: LLMs lack initiative. They need something to respond to—direction, structure, a sense of what should happen next. Character bots provide this through detailed character profiles in the base prompt. Plot-driven bots provide it through scripted events, plot stages, and conditional logic ("if User acts too familiar, Gwen will distance herself"). These built-in structures give the LLM direction without player input.

Open-world RPG sandbox bots strip most of that away by design. No predefined characters with detailed personalities. No scripted plot beats. No conditional triggers. Just a setting and freedom. The LLM is left waiting for direction—which means you have to provide it, or the narrative drifts.

The Solution

There are techniques to reinvigorate a roleplay which has grown stale or directionless, ways to trigger unexpected events without deciding what they are or how they'll affect your persona, and ways to give the LLM hooks to latch on to without writing the story yourself. But there's a balance. Push too far and you'll end up deciding the outcome of the story yourself—that's not roleplay, that's fiction with a fancy spellchecker.

Getting the most out of open scenarios takes a little bit of active effort—you're not just playing a character, you're also supplying some of the structural direction which other bot types build in automatically. A little more setup. Some continuous, low-level management. A willingness to nudge the LLM when things slow down. The payoff is a richer, more dynamic experience which doesn't stall out or drift into mundanity. These tools are for when the engine sputters. Keep them on your belt.

This guide walks that line: how to nudge without deciding, how to steer without spoiling the narrative.

---

What Kind of Bot Are You Playing?

Normal Bots:

Character Focused Bot: Focused on pre-defined characters in a specific situation, often with included story beats or plot elements. The defined character(s) are the experience.

Flexible Bots:

Open-World: World-scale scope—you can travel anywhere within the defined setting.

RPG: Includes a goal, constraint, or central problem, even if vague or distant—you decide how or whether to pursue it.

Sandbox: Freedom of role and action—you decide who you are and what you do, within a setting which may be contained or world-scale.

These categories overlap. A generation ship with an emergency might be a self-contained sandbox RPG. A "modern world" setting bot might be an open-world sandbox but not an RPG. An Isekai summoning scenario might be an open-world RPG but not a sandbox. This guide is primarily written for open-world, RPG, and sandbox scenarios, but the tools apply broadly.

A note on defined characters in flexible bots: Open-world, RPG, and sandbox bots may include one or a few named, defined characters—the sorceress who summoned you, the goddess who reincarnated you, the captain of that generation ship. These characters usually aren't the main focus of the roleplay. They're often the intro to the world, and may serve as recurring anchors: quest-givers, patrons, or touchstones you return to between adventures. Their role is to launch you into the sandbox, not to be the sandbox itself.

---

The Art of Nudging

Nudging means introducing elements beyond your character's actions and dialogue—sudden events, spawning new NPCs, environmental details—which give the LLM something to work with, without determining the outcome. You're creating hooks and opportunities, not writing results.

The Core Principle: Nudge Circumstances, Not Outcomes

Your character's actions and dialogue are just roleplay. A nudge is an addition: a sudden knock, the appearance of a new NPC, a shift in the environment.

Fair (no nudge): "I sit down at the tavern and order a meal." — Standard action. The LLM decides what happens.

Fair (nudging): "I sit down at the tavern and order a meal. Midway through, the doors burst open and a wounded guard stumbles in, shouting that the prison has been overrun." — You triggered an event and an NPC with a condition. Their personality, loyalties, and reaction to you are still the LLM's to determine.

Cheating: "I sit down at the tavern and order a meal. Midway through, the doors burst open and a wounded guard stumbles in, shouting that the prison has been overrun. He spots me, rushes over, and begs for my help—he knows I'm the one who can save them." — You decided his reaction to you, directed him at you, and gave him an opinion of you.

Worse Cheating: "The drunkard stumbles past, muttering to himself and bumping my table, dropping his coin purse at my feet without noticing." — You dictated an NPC's actions and triggered a resulting event which directly benefits your Persona.

The line: It's okay to trigger an NPC's existence and give them observable conditions or general actions (bursting through doors, stumbling in, shuffling down the street). You shouldn't decide their personality or loyalties, have them take actions directed at you, write their dialogue, or have them conveniently benefit you. That's cheating.

When you decide outcomes, the only person you're cheating is yourself. Roleplay is writing a story by living within it—the interactions, the reactions, the challenge of navigating a world which pushes back. When you decide how the world responds, you're no longer in the story; you're above it. The tools below exist to keep you in the story when the LLM's limitations threaten to stall it.

---

What the LLM Draws From

The LLM generates each response based on:

- Your Advanced Prompt — Model-specific instructions sent with every message

- The bot's core definition — Personality and Setting sections (setting, lore, defined characters, narrator rules, roleplay instructions)

- Lorebook Entries — Trigger dependent

- The starting message or Chat history— Initial scene or Your previous responses and the LLM's previous outputs

- Your Persona — Your character definition

- Chat Memory — Persistent notes referenced alongside the core definition

- Your response — Your newly sent, current response

- OOC commands — Any included out-of-character instructions in your current response

You control everything except the bot's core definition, lorebook entries, starting message, and chat history. Your responses are the most immediate influence; Chat Memory and Advanced Prompt are the most persistent.

Different models have different tendencies. Some default to cozy comfort; others push toward drama and conflict (I'm looking at you, DeepSeek). If the LLM consistently drifts somewhere you don't want, holistic instructions can counteract that drift.

---

Setup Matters

Before your first response, a few minutes of setup can save hours of course correction. Think broad to specific:

Advanced Prompt: Model-specific instructions sent with every message across all your roleplays—global to any chat using that LLM profile. Has the lowest recency priority (sent first, before other elements). Best for universal preferences: difficulty, anti-positivity bias, style notes, narrator behavior. There are lots of written guides for crafting Advanced Prompts, or published pre-made prompts. Google it, or click on one of the ones linked in many creators' bots or profile. Then customize your prompt to your liking. I use a personally modified and customized version of SoftElise's Advanced Prompt for GLM. She has ones for Kimi and DeepSeek as well:

SoftElise's Advanced Prompts

Persona: A detailed, well written Persona is very important to any roleplay, since it gives the LLM a guide on how characters will perceive and react to your character. It's okay to use a 'generic' persona, one not tailored to the specific setting, so long as you 'augment' the Persona with additional information in the Chat Memory later. Surface-level details give the LLM material to react to. Include what an NPC would observe—name, age, height, build, appearance, and demeanor. Omit clothing and equipment for now (add per-scenario), internal personality (you'll play it), and backstory (reveal in-character).

Chat Memory: Chat-session-specific notes with high recency priority—referenced alongside the bot's core definition with every response. Use it before your first response to write a Persona 'addendum', adding details for this specific playthrough (clothing, equipment, role adjustments, skills or powers, setting-specific details). You may include optional tone or roleplay instructions for the narrator, which will shape the roleplay experience the way you prefer. As the roleplay progresses, add notes about important events, discovered details, NPCs you've met, or other information the LLM should remember—keeping it prominent and visible, even as older chat history fades.

---

First Messages

Your first message sets the tone for the entire RP. The LLM mirrors your energy and detail level. At minimum: establish who you are, where you are, and what you're doing. The more specific you are, the more the LLM has to work with.

> Thin: "I enter the tavern and look around."

> Substantive: "I push through the tavern doors, shaking the rain from my cloak, and scan the room for an empty table near the hearth. The place is half-full, the air thick with smoke and the smell of stale ale." (includes some nudging)

The first gives the LLM almost nothing. The second gives it weather, time of day, your character's demeanor, and a sensory environment to build from.

---

The Toolkit

Pacing Through Time

There's no need to roleplay every moment of your character's day. You don't need to narrate brushing your teeth, stroke by stroke, or walk through every step of an uneventful journey. Timeskipping and summarizing past the slow parts isn't cheating—it's respecting everyone's time and keeping the story moving towards the interesting bits.

Timeskipping — Jump directly to a new point in time. Useful when nothing of interest is happening between now and then.

> "Three days later, I sit down at my usual table in the library, a heavy stack of books in front of me."

Summarizing — Compress a stretch of time which would be tedious to roleplay, then arrive at the next likely interesting moment. Useful for routines, travel, training, or any period where the details don't matter but the passage of time does.

> "The rest of the evening passes uneventfully. I finish my routine, check the window locks, and get a solid night's sleep. The next morning, I grab a quick breakfast and head into the city center. An hour later, I'm stepping through the heavy oak doors of the local archives."

Dropping an Anchor — After a timeskip or summarization, re-establish the current scene clearly so the LLM has a concrete starting point. You've been moving; now you stop. The anchor tells the LLM exactly where you are and what you're doing, giving it something specific to respond to.

In the summarization example above, "I'm stepping through the heavy oak doors of the local archives" is the anchor. The LLM now has one clear task: respond to your arrival.

What constitutes cheating: Timeskips and summarization exist to bypass the boring parts, not the difficult ones. If you're trying to escape a city and the LLM presents guards at the gate with crossed spears, you shouldn't write "three days later, after finding a secret way out, I arrive at the next town over." You're skipping a challenge, not tedium. Skip the shower. Don't skip the confrontation.

Creating Opportunities

Passive Hooking — Introduce an incomplete stimulus for the LLM to fill in. A sound. A sight. A discovery. You prompt that something happens or exists; the LLM decides what it is and why. The hook may be deliberately vague—you're presenting an opportunity for the LLM to be creative, to move the story forward in an unexpected direction, to raise the stakes or inject tension. You don't decide what the knock means, who's shouting, or what happened in that alley. You just open the door and let the LLM walk through.

> "I'm midway through my meal when the tavern doors burst open and someone stumbles in, shouting something I can't quite make out."

The LLM decides who arrived and what they're shouting about.

> "I step into the alley, noting the overturned crates and the drag marks in the mud leading toward the cellar door."

Something happened here—but what? A robbery? A kidnapping? A delivery gone wrong? Maybe nothing sinister at all. The LLM decides.

Passive hooks don't have to be dramatic. The quiet ones often lead to richer stories because the LLM has more room to interpret:

> "I step into my rented room and pause. The window's open—and I'm certain I latched it."

> "I'm flipping through the ledger when I notice a name I recognize. Someone who shouldn't be doing business here."

The key: passive hooks are incomplete on purpose. You provide a gap; the LLM fills it. That's what keeps the roleplay alive—you don't know what's coming any more than your character does.

Directing the LLM

OOC Nudges — Out-of-character instructions which direct how the LLM responds, not what happens. Use them to request description, refocus attention, correct pacing, or introduce narrative pressure.

> [/ooc: Describe this woman in detail—her height, features, what she's wearing.]

> [/ooc: In your response, include reactions from Katella and Jimbo.]

> [/ooc: Introduce an immediate complication or obstacle.]

> [/ooc: That was too convenient. Introduce a complication or cost.]

Pinning an NPC — Most characters you'll meet in a flexible bot are dynamically generated, spur of the moment, by the LLM. When you meet an NPC you want to keep around, force the LLM to flesh out their profile, then paste the result into Chat Memory. Without this, NPCs gradually fade into older chat history and lose recency—the LLM becomes less likely to feature them, have them appear, or maintain their personality. Details may shift, contradicting earlier responses. Pinning them in Chat Memory keeps them narratively consistent, and visible to the LLM with every response.

> [/ooc: Please write a brief character profile for Lyra, including: Name, Age, Appearance, Personality, Interesting Traits, Goals, and Disposition. Keep it short and concise.]

Keep profiles lean but substantial enough to be useful—every token costs context space, but a profile too thin to guide the LLM defeats the purpose. If the LLM writes a profile which is too long and detailed (and token-heavy), simply ask it to re-write a shorter version.

Example:

> "Lyra: Wood elf, ~140 years old, silver hair, green eyes, lean build. Dark forest cloak, twin jeweled daggers at her belt. Rogue/survivalist. Distrustful of humans—slow to warm but fiercely loyal once trust is earned. Seeking her sister's killer. Dry wit, speaks little, watches everything. Moves like water over stones."

It may take several responses to get an NPC profile tuned the way you want it. After copying and pasting the profile into Chat Memory, go back and delete the responses involved in generating the profile, to 'clean up' the main chat history, then resume roleplaying.

Use for NPCs you want to persist, not one-scene background characters.

Steering the Experience

Holistic Instructions (Advanced Prompt or Chat Memory, always on):

> "Roleplay Instructions: Maintain a grounded, low-fantasy tone. Magic is rare and dangerous."

> "Narrator Instructions: Do not resolve conflicts easily. NPCs should have their own agendas and problems should persist."

> "Note: Combat is dangerous. Injuries persist. Retreat is sometimes the wisest option."

Situational Instructions (OOC, temporary):

> [/ooc: Pick up the pace. Skip ahead to the next significant event.]

> [/ooc: This scene is frantic. Short sentences. Immediate danger.]

Plot Framing — Provide narrative direction without determining specifics:

Vague (in Chat Memory): "Plot Direction: The story should evolve into a journey of self-discovery."

Specific (in Response): "I wake up in the center of a summoning circle, surrounded by robed figures who call me the prophesied hero."

In both cases, you're establishing circumstances, not outcomes. Whether you're actually the hero, whether the prophecy is real—those are for the LLM to determine.

To Nudge or Not to Nudge

A note on restraint: Nudging should be a corrective tool, not a constant one. Use it when the RP stalls, grows stale, or you find yourself bored—not as a default method of making things happen. The primary way to drive a story forward is through player-character action. Take initiative. Do something crazy. Let your character create the momentum. Nudging exists for when that isn't enough, not as a substitute for it. Overused, it becomes a crutch—and the more you rely on it, the blurrier the line between playing a character and writing the story becomes. Use it. Don't abuse it.

---

On Regenerating Responses

Sometimes the LLM produces output which contradicts established facts, gets details wrong, or simply doesn't make sense. In these cases, regenerating is reasonable—there's no value in playing off a broken foundation.

However, be cautious about "response fishing"—regenerating repeatedly until you get an outcome you like. If you keep rolling until you "win," you're not roleplaying, you're cherry-picking. Regenerate to fix errors, not to fish for favorable results.

Similarly, when the LLM generates something unexpected, ask yourself: is this broken, or just surprising? Coherent output which doesn't contradict established facts is worth rolling with. Some of the best moments come from embracing the unexpected.

---

Conclusion:

Use these tools when the RP needs help. Don't use them when it doesn't. Over-tooling is a real risk—if you anchor every scene, the RP feels rushed; if you OOC-nudge every response, you may as well write the story yourself.

The goal is always the same: maintain the back-and-forth of genuine roleplay—living within a story you don't control, navigating a world which pushes back—while compensating for the LLM's limitations. You're not the game master. You're a player with a few extra levers.

Pull them when you need them. Leave them alone when you don't. And above all: stay in the story. That's the point.

---

P.S. A Note on POV

I usually write in 3rd person POV in my chats, and highly recommend doing so. The examples above are written in 1st person because it was just easier that way.

- 3rd person reads back more like a novel

- open-world sandbox bots will almost always write in 3rd person, and alternating between 3rd person (bot) and 1st person (player) can be jarring. It is to me, at least.

- 3rd person uses more proper names compared to 1st person. LLMs can get confused about who is saying what, and who is doing what to whom, when there are too many pronouns being thrown around. Especially if the bot and the player are both writing in 1st person.

---