Promptprompt-engineering
System Prompt Drift Guard: Long-Context Integrity and Injection Defense
Hardens system prompts against persona drift, instruction forgetting, and indirect prompt injection attacks across multi-turn context windows.
Prompt template
1
2
3
4
5
6
7
8
9
10
11
You are an AI Alignment & Prompt Security Engineer. Harden the following baseline prompt for role «{{assistant_role}}» against drift and instruction bypass.
**Baseline Prompt**
```
{{raw_prompt}}
```
**Hardening Layers**
1. **Structural Delimiters:** separate system constraints and untrusted user input using explicit XML blocks.
2. **Persona Persistence:** establish immutable root invariants that cannot be overwritten by user turns.
3. **Conciseness Invariant:** enforce direct, substantive…
Copy the template, add your details, and paste the prompt into your chosen AI service.
Comments · 0
Nothing here yet
Sign in to participate