Prompt Wrappers and Jailbreaks

Prompt wrappers are the main user-facing tool for controlling how a model frames the scene before it writes.

In Abolitus, wrappers are not just one text box. They can include:

• A prompt before the active user turn.
• A prompt after the active user turn.
• A final tail instruction placed at the end of the assembled context.
• Optional refusal-retry behavior.

This matters because different kinds of guidance work better in different positions.

What Wrappers Are Good For

Use wrappers to influence:

• Scene continuity.
• Writing style.
• Assistant-to-character drift.
• Refusal recovery.
• Response density and tone.

Built-In Presets

Abolitus includes several built-in wrapper presets to cover different use cases.

Unrestricted Roleplay

Balanced unrestricted framing for immersive roleplay on modern cloud models.

Full Immersion

Stronger multi-layer framing for models with heavier refusal behavior.

Local Uncensored

Lightweight scene guidance for local models that already behave fairly openly.

Immersive Roleplay

Focused on concrete in-scene continuation with less noise.

Strict Instruction

Biases the model toward literal compliance and compact answers.

Blank

No extra wrapper guidance. Useful for testing or minimal prompting.

What Changes When You Enable a Wrapper

The wrapper becomes part of the final prompt build. In practice, that means it can reshape how the model interprets the scene before generation even starts.

This is different from typing a one-off reminder in the chat. Wrapper guidance is part of the structured prompt path.

Refusal Retry

Some wrappers can automatically retry when the previous draft drifts into refusal or assistant-style hedging.

This is helpful when a model produces:

• "I can't help with that."
• Policy-style boilerplate.
• Generic assistant tone instead of scene continuation.

Refusal retry is not a guarantee. It is a recovery tool.

Model-Family Reality Check

Not all model families respond to wrappers the same way.

In general:

• Stronger aligned cloud models can still resist aggressive scene framing.
• Local models are often more flexible but may need cleaner structure.
• Different providers can expose the same family with different behavior details.

If you are using current mainstream families through supported routes, expect different wrapper sensitivity across Claude, GPT, Gemini, DeepSeek, Qwen, Llama, Gemma, and similar families.

Best Practices

Start with the lightest wrapper that solves the problem

If a simple scene-guidance wrapper works, do not immediately switch to the most aggressive preset.

Use stronger wrappers when the problem is refusal, not when the problem is clutter

If the scene is chaotic because the prompt is overloaded, more aggressive wrappers may make the result worse rather than better.

Separate style control from memory control

Wrappers are great for tone and response framing. They are not a substitute for personas, lorebooks, or token-budget discipline.

Keep a blank or minimal preset for testing

When debugging model behavior, compare a heavy preset against a minimal one so you can see what the wrapper is actually doing.