I was standing in the driveway of a $539,900 house in Gallatin, Tennessee, watching my meticulously crafted scoring system crumble.
On paper, this house was an 80 out of 100. Brand new construction. Gas cooktop. Dedicated office. Level lot. My spreadsheet practically glowed with green cells. I’d spent three days refining my Claude prompt to evaluate listings, creating weighted categories for everything from commute times to HVAC age.
Then my wife stepped out of the car, looked at the house, and said: “It faces the wrong way. The backyard feels cramped. I don’t like it.”
The Spreadsheet Delusion
As an engineer, I thought I could math my way through house hunting. We’re trained to break complex problems into measurable components. So naturally, I built what any reasonable developer would build: a 100-point scoring system with six weighted categories.
Price (25 points): < $520k gets full points
Age (15 points): Brand new gets full points
Flooring (15 points): Mostly hardwood, little carpet
Kitchen (15 points): Gas stove + vent hood
Lot & Yard (15 points): 0.25+ acres, fenced
Bonus Features (15 points): Level driveway, etc.I created a project on Claude Desktop and fed a carefully crafted prompt into the project instructions. I copied the entire listing from Zillow or Redfin and pasted it into a new conversation each time. One conversation for each house. The AI gave me a score and detailed breakdown for each house. I felt like I’d solved house hunting.
The first crack appeared when we visited a 75-point house. On paper, it was nearly perfect. In person? The “spacious” backyard was a narrow strip overlooked by three neighbors. The “open floor plan” meant you could hear the toilet flush from the dining room.
My wife’s immediate “no” wasn’t irrational. She was processing a thousand micro-details that my scoring system completely missed.
The CFO vs COO Problem
Here’s what took me too long to understand: I was being the CFO, running numbers and checking boxes. My wife was being the COO, thinking about the actual daily experience of living in these spaces.
She was evaluating different data:
- Will I feel happy drinking coffee here at 7 AM?
- Can I imagine hosting friends in this layout?
- Will this backyard make me want to go outside or stay in?
These are the difference between a house and a home. But they’re also not random—they can be measured if you ask the right questions.
The Two-Key Solution
Instead of fighting about spreadsheets versus feelings, I rebuilt my prompt with what I call the “Two-Key System.” A house needs two keys to unlock consideration: my objective score AND her subjective score.
Here’s the prompt structure that finally worked:
## Scoring System (170 Points Total)
### Key 1: Objective Analysis (140 points)
1. Commute Score (25 points)
2. Price (25 points)
- Under $520k = 25 pts
- $520-535k = 20 pts
- [continues...]
[... all the technical categories ...]
### Key 2: Feel Factor (30 points) - INTERACTIVE
After objective analysis, ask the buyers:
Natural Light & Orientation (10 pts)
"On a scale of 1-10, how does the natural light
make you feel? Energizing or cave-like?"
Layout & Flow (10 pts)
"Can you imagine your daily routine here? Where
would the coffee machine go? Rate 1-10."
Backyard Atmosphere (10 pts)
"Does this feel like a sanctuary or just grass?
Would you want to spend evenings here? Rate 1-10."The magic wasn’t in the specific questions. It was in making the AI pause after the objective analysis and explicitly ask for subjective input. No house moves forward without both keys.
The Prompt That Actually Works
Here are the actual instructions I give the AI now:
# House Hunting Scoring System
You are evaluating real estate listings. When provided with
a listing, analyze it in THREE phases:
## Phase 1: Objective Analysis (140 points)
[All the measurable criteria with specific point values]
## Phase 2: Feel Factor Questions
STOP after objective analysis. Ask the buyers to rate:
- Natural light and orientation feel (1-10)
- Daily living flow and layout (1-10)
- Backyard atmosphere and privacy (1-10)
## Phase 3: Combined Score & Recommendation
Only after receiving Feel Factor scores, provide:
- Final score out of 170
- Tier rating:
- Must-See: 135+ points
- Worth Considering: 115-134
- Proceed with Caution: 95-114
- Skip: Below 95What This Actually Accomplished
The two-phase approach did something unexpected: it made my wife trust the spreadsheet more, not less.
The system was now our system.
Even better, the scale forced us both to articulate why a house felt wrong. “The backyard feels bad” became “The backyard is a 3/10 because it’s narrow and directly overlooks the neighbor’s deck.”
Our scoring system has predicted our mutual reaction with a fair degree of accuracy. Houses scoring 130+ we both loved. Houses under 100 we both immediately rejected. The 110-125 range is where we negotiate.
The North-South Revelation
That original house in Gallatin that started this whole journey? When we applied the Two-Key System, it dropped from 80/100 to 92/170. The objective score was still strong, but the Feel Factor was 12/30:
- Natural Light: 3/10 (north-south orientation)
- Layout: 4/10 (felt compartmentalized despite being “open”)
- Backyard: 5/10 (overlooked by neighbors)
The system revealed what my wife knew instantly: great specs don’t overcome bad daily living experience.
What I’m Still Figuring Out
The Weighting Problem: Is 30 points for Feel Factor enough? The balance between objective and subjective is still more art than science.
The Photo Limitation: We’re scoring Feel Factor based on listing photos, but photos lie. That “sun-drenched living room” might only get light for 30 minutes at sunset.
The Preset Bias: Once the AI gives an objective score, does it influence how we subjectively rate the house? Would we rate differently if we saw the house first?
The Unexpected Win
The biggest surprise wasn’t that the system worked; it was how it changed our house hunting dynamic. We became a team with complementary evaluation methods.
Your Turn: What’s Your Second Key?
How are you handling the objective versus subjective challenge in your AI prompts? Have you found ways to make AI capture “gut feelings” in a systematic way?
Drop me a note at hello@ashishacharya.com. I’m especially curious if anyone has figured out how to make AI better at evaluating photos for subjective qualities.