When Prompt engineering and system prompts fits
- You need to shape model tone, format, scope, and refusal behavior. System prompts are exactly the right tool for that.
- You are deploying an AI product and need to define the model's persona and constrain its response patterns.
- You want to prevent the model from reproducing sensitive data in its outputs. System prompts can help with that.
When Locke fits
- Your concern is data in transit: sensitive fields leaving the device in the first place.
- You need a control that is resilient to prompt injection, which can override system-prompt instructions.
- You need auditable evidence of what was detected and how it was handled, not just a behavioral guideline.
Side-by-side
| Dimension | Locke | System prompts |
|---|---|---|
| When it acts | Before the prompt is sent: data is masked on the device at composition time. | After the prompt is sent: the instruction is part of the transmitted payload. |
| Does sensitive data leave the device? | No. Masked content is substituted before transmission. | Yes. The full prompt, sensitive fields included, is transmitted. |
| Resilience to prompt injection | Not relevant. Data is stripped before reaching the model. | Vulnerable. Injected instructions can override system-prompt constraints. |
| Auditable evidence | Local detection log of what was flagged and how it was handled. | None. Behavioral, not evidential. |
| Best for | Preventing specific data from reaching any model context. | Shaping model behavior once the model has received input. |
Bottom line
System prompts and Locke solve different problems and belong in the same stack. A system prompt can tell a model not to repeat an SSN, but the SSN already left your device in the prompt. Locke removes it before transmission, so it never enters the model's context.
