How to work

Our AI oracle in Copenhagen, Alex, can be booked Saturday from 9-14 for a 15-minute slot. Just click the link and book your preferred time. First come, first served.

Booking link: https://calendar.app.google/vRxZdB9xLnfoRoZYA

4 Elements of Output in AI Solutions

1. Large Language Models (LLM)

 Foundation: The LLM serves as the core of the AI solution.

 Capabilities and Limitations: Each model has unique strengths and limitations, affecting

factors like reasoning, creativity, and factual accuracy.

 Token Limits: Models like GPT-4 (128k tokens per chat) and Claude (200k tokens per

chat) have different token capacities. Tokens generally represent chunks of text, such as

words or characters, which limits the amount of information the model can process at

once.

 Model Selection: Choosing the right model for the task is essential, as different models

excel in specific types of output, such as creative writing vs. factual analysis.

2. External Sources

 Purpose: Attached files expand the model's knowledge and align it with specific subject

fields, helping control the data the model bases its responses on.

 Typical Attachments: Common sources include court cases, legal articles, firm

templates, book chapters, examples of previous work, and inspirational materials.

 Quality and Relevance: High-quality, relevant attachments improve the accuracy and

specificity of the output.

 Privacy and Security: Out Solution is GDPR safe

 Source Validation: Check the accuracy and credibility of attached sources to ensure

reliable responses.

 Attachment Formatting: Ensure files are formatted in a way the model can process (e.g.,

clear text, standard file types), as formatting issues may impact comprehension. Our

model does not read pictures/video/audio today.

 Scope Limitation: Be mindful of the scope of attached materials to avoid overwhelming

the model with too much context, which may dilute focus.

3. System Prompt

 Often referred to as the model’s “personality,” the system prompt controls how the

model presents its knowledge.

 Customizability: The system prompt can be adjusted to fit specific use cases or domains,

making outputs more consistent and aligned with desired tones or styles.

 Flexibility: System prompts can range from simple to advanced

 Benefit: A tailored system prompt increases the likelihood of stable and predictable

outputs.

 Consistency Across Tasks: Maintaining a consistent system prompt for similar tasks can

lead to more reliable and uniform results across interactions.

 Experimentation: Testing different variations of the system prompt can help identify

which configurations yield the best results for specific needs.

4. User Prompt

 Function: The user prompt specifies the task given to the model. Its complexity and

creativity are limited only by the user’s imagination.

 Prompting Techniques:

- One-shot Prompting: Providing a single example to guide the model.

- Series Prompting: Using a sequence of prompts to guide the model through multi-step

tasks.

 Optimization Tips:

- Use frameworks like COSTAR as a foundation for structuring prompts.

- Ask the model for feedback on how to improve prompts and clarify why the output

may not meet expectations.

 Clear and Concise Language: Use straightforward language to help the model

understand complex tasks without ambiguity.

 Layered Instructions: For complex tasks, break down instructions into clear steps or

phases to improve comprehension and output quality.

 Iterative Feedback Loop: Regularly update the prompt based on model performance

and output feedback for continuous improvement.

GOOD LUCK!