Five things about AI that explain the "nestled" problem
You do not need a computer science degree to get great copy out of these tools. You need five ideas, and each one explains a frustration you have probably already felt.
It predicts words, it does not know things
A language model works by guessing the most probable next word, over and over, very fast. That is why it is fluent even when it is wrong. Fluency is not accuracy, and confident-sounding copy still needs your eyes on it.
It reads in tokens
Tokens are chunks of words, roughly three quarters of a word each. This matters for one practical reason: long, rambling prompts cost more and get less attention per instruction. Short and specific beats long and vague.
It only sees what is in the chat
The context window is the model's working memory, and it is limited to the current conversation. Close the tab and your careful instructions vanish. This is the real reason your captions drift back to generic, and it is why the fix in this post involves saving something, not typing something.
It hallucinates
If you give it a listing with no fireplace, there is a decent chance it adds one anyway, because fireplaces show up in a lot of listings. The model is not lying. It is filling a gap with the most probable detail. The defense is simple: give it the facts and tell it to ask instead of inventing.
Temperature makes every answer different
Run the same prompt twice and you get two different results. That randomness is called temperature, and you can use it. Ask for three versions every time and pick the best one instead of settling for the first.
Why every AI listing says "nestled"
Clichés come from training data. These models learned to write by reading an enormous amount of text, and a huge share of listing descriptions on the open internet lean on the same forty words. "Nestled" is in a staggering number of them. So is "boasts." So is "oasis."
When the model reaches for the next word after "this home is," the statistically safest choice is the one it has seen most. That is "nestled." It is not a style decision. It is a probability.
Here is the good news hiding in that: banning those words is the fastest way to sound human. The moment "nestled" is off the table, the model has to reach for something specific, and specific is what buyers actually read.
Swap the cliché for the specific
Keep this list taped to your monitor. Each swap replaces a word the model loves with a fact only you know.
- "Nestled in a quiet neighborhood" becomes "On a cul-de-sac two streets from the park."
- "Boasts a gourmet kitchen" becomes "Quartz counters, a 36-inch gas range, a walk-in pantry."
- "A true backyard oasis" becomes "A flat half-acre with mature oaks and a covered patio."
- "Move-in ready" becomes "Roof 2022, HVAC serviced 2026, fresh paint throughout."
- "Close to everything" becomes "Four minutes to Oak Hill Elementary, fifteen to downtown."
- "Won't last long" becomes nothing. Delete it.
Notice that every replacement is harder to write because it requires a real detail. That is the point. The specific detail is the thing AI cannot fake and the thing your buyer remembers.
The master voice prompt (write it once)
Most agents try to fix generic output by writing a better prompt every single time. That is exhausting, and it is why the AI "phase" fizzles out by week three.
The better move is a master voice prompt: one paragraph that tells the model who you are, who you serve, how you sound, what you never say, and what your writing actually looks like. Fill in the placeholders, paste it in, and you are done.
I am {{name}}, a real estate agent in {{city}} specializing in {{niche}} in {{neighborhoods}}. My clients are {{avatar}}. My tone: {{three adjectives, e.g. warm, direct, a little funny}}. I write in short sentences, plain English, first person. I never use: nestled, boasts, oasis, dream home, unlock, elevate, or more than one emoji. I always include one specific local detail. Here are three examples of my writing: {{caption}} / {{email}} / {{listing description}}. Match this voice in everything you write for me unless I say otherwise. If you are missing a fact, ask instead of inventing.
A few notes on why each piece is there:
- Name, city, niche, neighborhoods give the model a lane. "Condo specialist in downtown Denver" produces different sentences than "ranch properties outside Fort Worth."
- Avatar tells it who is reading. Relocating nurses and first-time buyers do not want the same email.
- Three adjectives are enough. More than that and they cancel each other out.
- The banned list is where the magic is. Add to it every time you catch a word you would never say.
- "Ask instead of inventing" is your hallucination insurance. It will stop the phantom fireplace.
The three samples that teach it fastest
The examples do more work than every other line combined. The model learns rhythm, sentence length and vocabulary from samples far better than from adjectives. Pick these three:
- One listing description you are proud of. Not the longest one. The one where a buyer told you it was why they booked the showing.
- One caption that performed. Saves, shares or comments, whichever you track. If it worked on your audience, it carries your voice.
- One client email that sounded like you on a good day. The follow-up after a tough inspection. The "we got it" message. Real, unpolished, yours.
Three is enough to see a difference immediately. Ten is better. A hundred is unnecessary, so stop collecting and start using.
Where to save it so it actually remembers
This is the step most agents skip, and it is the one that makes everything else stick. Saved instructions survive the context window problem. Every new chat starts already knowing you.
Claude
Open Projects and create one. Paste the master voice prompt into the project instructions. Upload your best listings, templates and past emails as project files so Claude can reference them. Make one project per use case: Listings, Emails, Social. Each gets the same voice prompt plus the files that matter for that job.
ChatGPT
Go to Settings, then Personalization, and paste the prompt into custom instructions. Turn memory on so it retains details across conversations. If you want something more structured, build a custom GPT with the prompt as its instructions and your samples as its knowledge files.
Gemini
Create a Gem and give it the master voice prompt as its instructions. Use saved info for standing facts (your brokerage, your markets, your license number format). Connect Gmail and Docs so it can pull real context from your actual correspondence instead of guessing.
Whichever tool you use, the pattern is identical: instructions in the permanent slot, samples attached, one workspace per job.
The read-aloud test
You will know the prompt is working in about 30 seconds. Here is the test.
- Ask for a caption for a listing you have right now.
- Read the result out loud. Actually out loud, not in your head.
- Picture yourself across a kitchen table from a client. Would you say every one of those sentences to their face?
If a single word makes you wince, do two things: add that word to the banned list and add one more writing sample. Then run it again. Most agents need two or three rounds before it clicks, and after that it stays clicked, because the prompt is saved.
The test works because your ear is a better editor than your eyes. Clichés slide past on a screen. They stick out when you hear them in your own voice.
A faster route if you would rather not maintain a prompt
If the idea of managing three projects across three apps makes you tired, that is fair. Agentree's AI Studio was built for exactly this. Teach AI Your Voice takes the same three samples once (a listing description, a caption, a client email) and every writing tool in the studio matches it from then on, no master prompt to copy around or update. The on-voice copy then lives on your Agentree listing page, which is where a profile click should land in the first place.
Either way, the principle is the same. Introduce yourself once, properly, and the stranger goes away.