The Future of AI Prompting: 5 Context Levels

Tiago Forte

31,778 views • 1 month ago

Video Summary

The key to unlocking better AI performance isn't more complex prompts, but rather providing richer context. This involves understanding the different layers of context available to AI models, starting from their vast training data and moving towards more specific and personalized information. By leveraging system prompts, user preferences, and project-specific knowledge, you can transform an AI from a general assistant into a highly effective team member.

The process involves four primary levels of context. Level two, the system prompt, is uneditable but can be "exploited" by using specific trigger words to elicit deeper analysis. Level three, user preferences, allows you to dictate how the AI communicates, such as using bullet points or stating its certainty level. This personalizes the AI's output to your specific needs and communication style.

The most transformative level, level four, is project knowledge. By uploading relevant files and documents into a project workspace, you provide the AI with detailed information about your products, customers, and business operations. This allows for highly specific and actionable outputs, dramatically improving the AI's effectiveness and saving significant time in generating strategic content.

Short Highlights

The effectiveness of AI is determined by the context provided, not just prompt complexity.
System prompts, though uneditable, can be influenced by specific trigger words for deeper AI analysis.
User preferences allow for personalized AI communication, such as preferred output formats and certainty levels.
Project knowledge, by uploading relevant files, provides AI with detailed business-specific information for highly tailored outputs.
A roadmap to "context mastery" involves researching system prompts, setting user preferences, and organizing project files within a week.

The Foundation: AI's General Knowledge [00:26]

Every AI and LLM has processed billions or trillions of documents, giving them a broad "business degree."
The primary limitation of AI is its lack of specific knowledge about you or your business.
A general prompt for a marketing strategy can yield a comprehensive-sounding but ultimately inaccurate response because the AI lacks specific business context.
The AI's initial response was incorrect because it made assumptions about audience segments, and didn't know that SEO or partnerships were not priorities.

The problem is that it knows nothing about you.

Level 1: Training Data [01:30]

The initial, less effective AI response is based solely on "level one" context, which is its training data.
Obsessing over the size or sophistication of AI models (e.g., number of parameters) is less important than the context provided.
The sophistication of the model is not what matters most; it's how much context you give it.

What matters now at this moment is not the size or the sophistication of the model. What matters is how much context you give it.

Level 2: System Prompts [02:02]

Every LLM has a hidden "system prompt," often a 120-page document, that shapes its responses.
These system prompts are created by the AI developers and cannot be edited directly.
However, system prompts can be "exploited" by using specific trigger words.
For example, Claude has seven trigger words that force the LLM to perform deeper analysis and thinking.
Using even a few of these trigger words can lead to a significant improvement in AI output without adding more explicit context.
Adding a few strategic words like "in-depth" and "comprehensive" to a prompt resulted in a deeper, more strategic marketing strategy that considered market size, growth, and competitive landscape.

You can't edit it. You can't change it, but you can exploit it.

Level 3: User Preferences [04:19]

This is the first level where users can take direct control of the AI's output.
By adding details about the product or service (e.g., "an online training program that helps business owners implement AI"), the AI's response becomes more tailored.
User preferences can be set in the AI's settings or preferences menu.
These preferences are remembered and applied to all future conversations within that AI.
Examples of user preferences include: answering in bullet points and sub-bullet points, being conservative in thinking, and stating the level of certainty.
This level also leads to the AI being more conservative in its estimations and avoiding unrealistic claims.

You can tell the LLM how you want it to communicate with you. You can set those communication preferences once and they'll apply everywhere all the time.

Level 4: Project Knowledge [06:32]

This is considered the most transformative and underappreciated level of AI context.
By uploading project-specific documents into a dedicated workspace, the AI gains deep knowledge about your products, pricing, customer segments, past testimonials, and competitors.
The AI can recall specific details like product names, company abbreviations, customer ranges (10 to 50 employees), and preferred marketing channels (email newsletter, LinkedIn).
This allows the AI to generate highly specific and actionable strategies that align perfectly with your business context.
The AI essentially acts like a fully onboarded, experienced team member rather than an intern.

What you see here is something that is so far advanced beyond what I what I think most people are using LLMs for.

Level 5: The Prompt Itself [09:48]

This level combines the previous four layers of permanent context with the final prompt entered into the chat box.
Even with advanced context engineering, prompt engineering remains crucial for effective AI interaction.
When the right context is preloaded, prompt engineering becomes vastly more powerful and effective.
The AI can even assist in creating sophisticated prompts by analyzing existing information and desired outcomes.
For instance, the AI can generate a detailed prompt for a marketing strategy by analyzing project data and specific requests for content like YouTube video ideas and team assignments.