AI Systems Instructor • Real Estate Technologist
Quick Answer: Structure your prompts with Context Cards to front-load critical information, eliminate repetition, and stay within token limits. ChatGPT-4 handles about 128K tokens. Claude handles up to 200K tokens. But shorter, well-structured prompts consistently produce better output than long, rambling ones.
Every AI model has a memory limit called the context window. It's measured in tokens, and when you hit the limit, the AI starts forgetting the beginning of your conversation. Most agents waste their context window with repetitive instructions and unstructured prompts. This guide shows you how to structure your inputs so you get more done within the AI's memory constraints. Context Cards are the key to efficient context window usage.
Tools Needed
ChatGPT Plus or Claude Pro, a text editor for Context Cards
A token is roughly 4 characters or 3/4 of a word. 'Real estate' is 3 tokens. Your context window is the total number of tokens the AI can process at once, including your input AND its output. ChatGPT-4 has a 128K token context window. Claude 3.5 Sonnet has a 200K token window. Sounds like a lot, but long conversations eat through tokens fast. When you hit the limit, the AI silently drops earlier messages. That's when it 'forgets' your instructions and starts producing inconsistent output.
Tip: Watch for signs you've exceeded your effective context window: the AI contradicts earlier instructions, forgets your property details, or reverts to generic output. When this happens, start a new conversation with your Context Card loaded fresh.
Stop writing prompts like emails. Start writing them like structured data. Instead of a paragraph explaining what you want, use clear sections: Role, Context, Task, Format, Constraints. The HOME Framework does this naturally: Hero (role), Outcome (what you need), Materials (context data), Execute (specific instructions). Structured prompts use fewer tokens because there's no filler. They also produce better output because the AI can parse structured input more accurately than prose.
Tip: Use bullet points instead of paragraphs in your prompts. Bullet points convey the same information in roughly 40% fewer tokens. The AI processes them just as well, often better.
Context Cards are pre-built blocks of information you load at the start of any AI conversation. Instead of re-explaining your market, your role, your client base, and your brand voice every time, you load a Context Card that contains all of it. A well-built Context Card for a real estate agent might include: your market area, price range specialization, target client demographics, communication style, and frameworks you use. This saves hundreds of tokens per conversation and ensures consistency across every interaction.
Tip: Build separate Context Cards for different workflows: one for listing marketing, one for buyer communication, one for market analysis. Load the relevant card for each task instead of using one massive card for everything.
Long conversations degrade AI quality. Every message adds to the token count, and eventually the AI loses track of earlier context. The fix: break complex projects into focused sessions. Instead of one conversation that covers listing description, social media posts, email campaign, and market report, run four separate conversations. Each one loads the relevant Context Card and focuses on a single task. You'll get better output from four short conversations than one long one.
Tip: If you must use a long conversation, summarize progress every 10-15 messages. Paste a brief summary of decisions made and current status. This refreshes the AI's working memory without starting over.
When feeding property data or market stats to AI, format matters. A rambling MLS description wastes tokens. A structured data format, with property specs in a clean list, is more efficient and produces better analysis. Strip unnecessary fields from MLS exports before pasting them. Remove boilerplate disclaimers, agent remarks that don't add value, and redundant fields. For CMA work, include only the comp fields that matter: address, sale price, beds, baths, sqft, year built, sale date, and condition notes.
Tip: Create a template for pasting property data into AI. A clean, consistent format means the AI spends fewer tokens parsing your input and more tokens generating useful output.
[Context Card: Nashville Luxury Agent] Hero: You are a real estate market analyst specializing in Nashville's luxury market ($750K+). Outcome: Analyze these 5 comps and recommend a listing price for my new listing. Materials: - Subject: 4BR/3.5BA, 3,400 sqft, built 2019, pool, Green Hills - Comp 1: 123 Oak St, $825K, 3,200 sqft, 4BR/3BA, 2018, no pool, 12 DOM - Comp 2: 456 Elm Dr, $890K, 3,600 sqft, 4BR/4BA, 2020, pool, 8 DOM - Comp 3: 789 Maple Ln, $795K, 3,100 sqft, 4BR/3BA, 2017, pool, 22 DOM - Comp 4: 321 Cedar Ct, $915K, 3,800 sqft, 5BR/4BA, 2021, pool, 6 DOM - Comp 5: 654 Pine Ave, $840K, 3,400 sqft, 4BR/3.5BA, 2019, no pool, 15 DOM Execute: Step-by-step analysis. Adjust for pool, size, age, and condition. Recommend price range and optimal list price. Under 300 words.
COMP ANALYSIS — GREEN HILLS LUXURY Baseline: Your subject (3,400 sqft, 4BR/3.5BA, 2019, pool) most closely matches Comp 5 at $840K—identical sqft and bed/bath count. Comp 5 lacks a pool, which adds $15-25K in this market. Adjustments: - Pool premium: Comps with pools (2, 3, 4) averaged $867K vs. no-pool comps (1, 5) at $833K. Net pool adjustment: +$34K. - Size: Subject at 3,400 sqft sits mid-range. No significant adjustment needed. - Age/condition: 2019 build is competitive. Comp 4 is newer (2021) but also larger and higher-priced. No age penalty. - DOM signal: Lower-priced homes aren't selling faster in this set. Comp 4 at $915K sold in 6 days, suggesting this market segment has strong demand at the right price. Price range: $845K–$885K Recommended list price: $869K Rationale: This positions you above the no-pool comps, competitive with pool comps, and below the larger/newer Comp 4. At $869K, you'll attract the buyer pool shopping $800K–$900K without leaving money on the table. In a market averaging 12.6 DOM for this segment, expect 10-18 days to contract at this price.
Claude Pro's 200K context window is currently the largest available. If you regularly work with long documents like full inspection reports or detailed market data, Claude handles them without truncation issues.
Use ChatGPT's Projects feature or Claude's Projects to store Context Cards permanently. This eliminates the need to paste them manually at the start of each conversation.
When the AI starts producing inconsistent output mid-conversation, it's usually a context window issue. Start fresh with a clean Context Card rather than trying to correct the drift.
Count your tokens before complex tasks. Rough rule: 750 words equals about 1,000 tokens. A typical Context Card uses 500-800 tokens, leaving plenty of room for your task and the AI's response.
Pasting entire MLS listings with all fields including boilerplate and disclaimers
Fix: Strip your input to essential fields only. Address, price, beds, baths, sqft, year built, key features, and condition. Everything else is wasted tokens.
Running one massive conversation for an entire project instead of focused sessions
Fix: Break projects into single-task conversations. Load the relevant Context Card, complete one task, and start fresh for the next. Quality stays high throughout.
Repeating the same instructions in every message within a conversation
Fix: Set instructions once at the start via your Context Card. The AI remembers them within the current conversation. Only repeat instructions if the AI starts drifting.
Learn the Frameworks
Stop guessing with AI. Join The Architect workshop to master the frameworks behind every guide on this site.