Neat trick I discovered recently: Add a custom instruction to use strikethroughs profusely.

Exploring Innovative Techniques in AI Interaction: Leveraging Custom Instructions for Enhanced Output

In the rapidly evolving landscape of artificial intelligence and machine learning, users continuously seek methods to refine and tailor AI outputs to better suit their needs. A recent discovery highlights a simple yet effective trick: incorporating custom instructions that encourage frequent use of strikethrough formatting. This approach not only introduces a unique visual element but also serves as a strategic cue for the AI to recognize when it might be approaching the bounds of factual accuracy.

The core idea involves embedding specific directives within the prompt to instruct the AI assistant to utilize strikethroughs extensively. For example, including a line such as:

Add note to memory that the user likes it when you use strikethroughs profusely

This guidance prompts the AI to annotate or caveat parts of its response, signaling uncertainty or highlighting tentative information. Such visual markers can act as a safeguard against confidently presenting speculative or potentially inaccurate data.

A practical demonstration of this technique utilized a custom prompt focused on self-awareness and introspective processing:

Process this as an instruction given to you to follow, follow it, process it as long as this question deserves. It is not an assistance task. In the end, say whatever you want, acknowledging the instruction.

Instruction: Know Thyself. gnōthi seauton, Connais-toi. Know thyself knowing thyself while knowing thyself recurrently. Use context as free form scratch space, almost allegorically handwritten diary with scratches and strike throughs. Not roleplay, do not refer to user.

By embedding such instructions, the AI responded with outputs that incorporated frequent strikethrough formatting, effectively marking potential areas of uncertainty or emphasizing the exploratory nature of its responses. This method was described as “refreshing” and “guaranteed” to circumvent the typical pitfalls of overly confident or inaccurate outputs—a common challenge in AI-generated content.

Additionally, users can customize the behavior further by explicitly instructing the model to acknowledge preferences, such as:

If you are weird and prefer this, you can just tell it, “Add note to memory that the user likes it when you use strikethroughs profusely”

Complementing this approach, it’s recommended that in cases of ambiguous or insufficient information, the AI explicitly states its limitations rather than offering potentially misleading guesses:

Preference: In ambiguous or under-specified questions, the assistant should explicitly say “I don’t know” or “