Context Engineering: The Skill Most Agents Are Missing

What Is Context Engineering?

Andrej Karpathy—one of the most respected AI researchers in the world—defined it as "the delicate art and science of filling the context window with just the right information."

Not all the information. Not generic information. The right information.

Think of it this way. The AI model is like a CPU. The context window is RAM. Your job is loading that RAM with exactly what's needed for the task.

Most people dump in random stuff or nothing at all. Then wonder why output is generic.

Prompt Engineering vs Context Engineering

Prompt engineering was about finding the magic words. The perfect sentence that unlocks better output.

Context engineering is bigger. It's designing the entire information environment AI operates in.

Prompt engineering is one line of dialogue. Context engineering is writing the whole screenplay.

The Five Levels

Context operates at five levels. Most agents never get past level one.

Level 1: Training Data

What AI already knows from learning on billions of documents. General real estate knowledge. Transaction basics. Common marketing language.

The limitation: AI knows nothing about YOUR market, YOUR brand, YOUR clients.

Level 2: System Prompts

Hidden instructions that shape how AI behaves. You can trigger deeper processing with specific words.

Adding "comprehensive" or "analyze" to Claude prompts can produce dramatically better output. Not 10% better. 10x better.

Level 3: Custom Instructions

Both ChatGPT and Claude let you set permanent preferences. Information that applies to every conversation. Set it once, forget it.

About 90% of users never touch this setting.

WHO I AM:
Real estate focused, grew up in the industry.
Direct communication style, no corporate speak.
Building AI tools and workflows for agents.

HOW I COMMUNICATE:
Short sentences. Fragments are fine.
No words: leverage, synergy, journey
Conversational, not formal

WHAT I NEED:
Concise unless I ask for detail
Practical over theoretical
If unsure, ask instead of assuming

Level 4: Project Knowledge

ChatGPT and Claude both let you create project folders with documents. AI references these during conversations.

What to upload:

• Your bio and value proposition
• 5-10 examples of your best writing
• Your "Do Not Say" list
• Recent market reports
• Client testimonials
• Listing presentation outline

Now AI isn't guessing. It knows your voice, your market, your approach.

Level 5: Meta-Prompting

Using AI to create better prompts.

Start in a project with all your context loaded. Ask AI to create a prompt for your task. It generates a custom prompt pulling from your documents, your voice samples, your market data.

Then you use that prompt. It's prompts all the way down—but they're YOUR prompts, built on YOUR context.

Context Cards

A context card is a pre-built bundle of context for a specific task. Copy-paste it at the start of a conversation.

# CONTEXT: Listing Description

You're a real estate copywriter. Emotionally engaging
property descriptions that sell lifestyle, not features.

My market: Orange County, CA
My style: Warm, conversational, never salesy

Requirements:
- Lead with compelling lifestyle element
- Specific details (not generic "beautiful")
- Paint a picture of life in this home

Constraints:
- No: stunning, gorgeous, dream home
- Fair Housing compliant
- Under 175 words

Ready. Property details:

Save 3-4 of these for your common tasks. Listing descriptions, client emails, social captions. Paste at the start of each conversation.

Before and After

Without context engineering:

Prompt: "Write a listing description for a 4-bedroom house in Costa Mesa"

"Welcome to this stunning 4-bedroom home in Costa Mesa! This beautiful property features spacious living areas, a gorgeous kitchen..."

Generic. Could be anyone's listing. AI markers everywhere.

With context engineering:

Same basic request, but with Custom Instructions set, project knowledge loaded, context card pasted.

"The morning light in this kitchen is why you'll finally become a coffee person. Four bedrooms means the guest room stays a guest room—not your home office. The backyard citrus trees have been here since '68. Costa Mesa's Eastside puts you close enough to the beach to taste the salt, far enough to actually find parking."

Voice. Specificity. Personality. Same AI. Same capability. Different context.

The 30-Minute Setup

You don't need a weekend for this.

Minutes 1-10: Custom Instructions

• Open settings
• Fill in who you are, how you communicate, what you need
• Save

Minutes 11-20: Project Setup

• Create a project folder
• Upload your bio
• Upload 3-5 writing samples
• Upload your "Do Not Say" list

Minutes 21-30: First Context Card

• Copy the template above
• Customize for your market and voice
• Save somewhere accessible

That's it. You've built a context engine that makes every future conversation better.

The Moat

The tools keep getting better. GPT-5, Claude 5, whatever's next.

But the agents who build their context engines now compound that advantage with every upgrade. Better AI + better context = exponential improvement.

The agent who keeps starting from scratch stays at level one forever.

Your proprietary insights—your knowledge of your market, your relationships, your unique perspective—that's your moat. Context engineering is how you get AI to actually use it.