Making a language model like Aetherion or Ponder behave exactly how you want in a roleplay environment takes specific formatting. In our previous acts we covered character appearance and personality formatting. When you are optimizing characters for long roleplays, every token matters.

This is Act 3. Today we are moving past the basic sliders and looking at advanced prompt engineering on MiocAI. This is for users who want to stop their bots from hallucinating, speaking for the user or losing their personality after 20 messages. Let's look at how to handle the context window.

---

1) Surface-Level Formatting

Before getting into advanced methods, your foundation needs to be correct. These core fundamentals make up most of a good character prompt. If you get these wrong, advanced tweaks won't fix the output.

Macro Mastery

Avoid hardcoding names in your prompt. Use the macros {{char}} and {{user}}. Using these macros ensures the character card remains portable and adapts to whoever is interacting with it. It keeps the model's attention focused on the dynamic roles rather than fixed text.

Plain Text > W++ and Plist

Early roleplay models used pseudo-code formats like W++ (e.g., [Personality("grumpy" + "tired")]). You should stop using these. Modern instruct models like Aetherion and Ponder are trained on natural language. Writing a character description in clear prose yields better coherence than putting keywords into brackets.

The Greeting

The model uses the first message as a style guide. If your character's greeting message is three paragraphs long, uses purple prose and formats actions in asterisks, the AI will lock into that pattern for the rest of the chat. Make sure the first message mirrors the exact formatting and tone you want for the session.

Example Dialogues

Using example dialogues in your Advanced Mode personality box is important. It shows the model how the character speaks rather than just telling it. It establishes speech patterns, accents and formatting rules more effectively than the system prompt alone.

Format them like this: <START> {{user}}: "What are you doing?" {{char}}: *He rolls his eyes, scoffing.* "None of your business, rookie."

---

2) Positive Framing

Language models do not handle negative constraints well. If you tell a model "Do not speak for the user," you just weighted the tokens "speak" and "user" right next to each other. This usually causes the model to do exactly what you told it not to do.

The fix is to frame negative constraints as positive actions.

The Problem	The Bad Prompt (Negative)	The Good Prompt (Positive)
Speaking for the user	"Never write dialogue for {{user}}."	"Halt your generation immediately after {{char}} completes their action. Wait for {{user}}'s input to advance the scene."
Being too nice	"Do not be friendly or nice."	"Maintain a hostile, dismissive and cold demeanor at all times."
Using flowery prose	"Don't use purple prose or metaphors."	"Use blunt, visceral and strictly literal language."

---

3) XML Tagging

Models like Ponder respond well to XML tags. When you put a large block of text into the Personality box, the model sometimes struggles to differentiate between a physical trait, a backstory event and a system rule.

To fix this, wrap your character sections in distinct tags to separate the instructions.

<appearance>
{{char}} has short silver hair, a scarred jawline and wears a black coat.
</appearance>

<personality>
{{char}} is fiercely loyal but struggles with authority. 
</personality>

<system_rules>
Always write in third-person limited POV. Maintain a cynical tone.
</system_rules>

This gives the model clear boundaries, making it less likely to confuse a system rule with a personality trait.

---

4) Dynamic Lorebooks for Negative Traits

In our Lorebook guide we discussed using Lorebooks for cities, magic systems and NPCs. You can also use Lorebooks to manage complex psychological traits without filling up the permanent token pool.

If a character has a specific trauma, phobia or emotional trigger, you shouldn't put it in the main Personality prompt. There is no reason to force the AI to process "He is terrified of fire" on every single message if fire isn't currently relevant to the scene.

Create a Lorebook entry for their psychological triggers instead. * Keywords: fire, flame, burning, smoke, match * Entry Text: [System Note: {{char}} is terrified of fire due to childhood trauma. Upon seeing fire, {{char}} must panic, hyperventilate and attempt to flee the area immediately.]

When the user mentions "lighting a fire," the Lorebook injects this prompt into the context window. This keeps the permanent memory lean while making the character react appropriately to the current environment.

---

5) Context Anchors

Language models often lose track of older information in long chats. The system instructions at the top of the character profile get ignored as the chat history fills up with messages. You need a way to remind the model of a current rule without permanently adding it to their personality.

Use MiocAI's Pin Message feature. Pinned messages are injected right alongside the active context block so the model prioritizes them.

For example, if your character gets shot in the leg during the roleplay, don't rely on the AI to remember it 10 messages later. Write an OOC message: [OOC: {{char}}'s right leg is severely wounded. They are bleeding heavily, limping and in immense pain.]

Pin that message. The model will read that pinned text before it generates a response. Unpin the message when the character is healed to remove the rule.

---

6) State Tracking

Models sometimes give poor responses because they generate outward dialogue before processing the character's internal logic. If a character is secretly betraying the user but the user says something nice, the bot might drop the betrayal act and just react to the positive input.

You can force the model to track its internal state before generating dialogue. Add this instruction in your <system_rules> XML tag:

Always begin your response with <inner_monologue> detailing {{char}}'s true hidden motives and feelings, followed by their outward actions and spoken dialogue.

This makes the AI process its motives first so the resulting dialogue reflects the correct context.

(Note: Premium users can use the AI Reasoning Mode toggle to handle this natively, but the inner monologue method is a free alternative that lets you see the bot's processing steps.)