Claude Vs Gemini For Coding: A Local Breakdown
The open-source Claude skills repo for real estate agents answers a question most AI tool reviews never ask: where does the data come from? If you are already using Claude Code in your business, the skills in that repo divide cleanly into two categories — skills that run on your licensed data and skills that run on public records. Understanding that split is not a technical detail. It is where your compliance line lives.
This post goes narrower than the overview page. If you want the full list of what the repo contains and how to install a skill, start at the open-source Claude skills repo for real estate agents. This post is for agents who are already in the repo and want to know exactly which workflows are pulling from agent-supplied inputs versus which ones run autonomously against public data — and why that distinction protects your license.
The Two-Lane Data Split Every Agent Needs to Understand
Before you run any skill from the Claude skills repo, you need a mental model that the repo’s overview page does not spell out in granular terms: every skill in the library belongs to one of two data lanes.
Lane 1 — Agent-supplied data. These skills do not run until you provide input. They accept your MLS export, your comp selections, your listing remarks draft, or your client CRM notes. The skill formats, analyzes, or assembles based on what you bring. It cannot do anything without you.
Lane 2 — Public records data. These skills query open government portals — permit databases, GIS parcel data, Register of Deeds records — and return structured results without needing any input from your licensed workflow. They are fully autonomous on the public-data side.
The real power — and the compliance-safe design — comes from combining both lanes in sequence. That combination is the system. Neither lane replaces the other.
Which Skills Require Your Input to Run?
Several skills in the Claude skills repo for real estate agents are input-gated. They literally wait for you. Here is what that category includes and what each one does with your data when you supply it.
The Price-Analysis Skill
This is the skill most agents ask about first, and it is the one where the data-split matters most. The price-analysis skill runs in two sequential steps. Step one is yours entirely: you log into your MLS, run your comp search, apply your filters, and export your results. In the workflow Victoria runs, that export lands as a local CSV file the script reads directly. The MLS itself has no API connection to this system — she pulls and she chooses.
Step two is the automation: the skill reads your export, then goes out to Pitt County’s permit portal and open data sources to pull permit history on the subject property. It checks whether additions or conversions have a closed certificate of occupancy. It cross-references the Register of Deeds for covenants and leasing restrictions. Then it runs three valuation methods against your comps and builds the packet.
Your comp selection drove the output. The automation handled the retrieval and assembly. That sequence is intentional and non-negotiable.
The Listing-Remarks Skill
The listing-remarks skill is purely agent-input-driven. You supply the property details — features, recent updates, lot characteristics, any seller-provided notes. The skill drafts remarks. You review before anything touches the MLS. There is no data pull here; the skill’s job is language, not data retrieval.
The Listing-Presentation Skill
This skill takes your pricing opinion — which you derived from your own comp analysis — and builds the presentation packet around it. It pulls in the math you already ran, formats the net-to-seller calculation, and assembles slides or a Gamma document for the seller conversation. Again: your judgment first, assembly second.
Which Skills Run Autonomously on Public Records?
A separate category of skills in the repo runs on a schedule without waiting for you to trigger them manually. These are the cron-based public-records skills. They operate on Lane 2 data exclusively — nothing that touches MLS, nothing that requires a licensed input, nothing that produces an opinion of value.
The Permit-Pull Workflow
On a schedule, this workflow queries the Pitt County EPL portal and the Pitt County Permits open data source. It returns permit history for a specified parcel — what was pulled, what was closed, what is still open. An agent reviews this output; the skill does not interpret it. The output is raw permit data structured for easy reading, not a recommendation.
The Public-Records Pull
The public-records skill queries Pitt County OPIS GIS parcel data and ArcGIS open data layers. It surfaces ownership records, lot boundaries, and parcel-level characteristics. The Register of Deeds pull — handled by a separate documented script — retrieves deed covenants and amendments, including any owner-occupancy or leasing restrictions that might affect an investor’s purchase decision.
The Market Intel Cron
Victoria’s daily market intel cron runs at 5:50 PM and writes a structured JSON file with market-level signals. That file is the only source from which numbers appear in published content. If a number is not in the file, it does not get published. This is the specific reason any post on this brand that cites market data is traceable — not because the AI is confident, but because the pipeline enforces a citation gate.
How the Two Lanes Combine in a Real Transaction
Here is a concrete walkthrough of how the two lanes operate together during a listing appointment prep sequence. This is the actual workflow, not an aspirational example.
The day before the listing appointment, Victoria logs into NCRMLS and runs her comp search. She applies her own filters — square footage range, age, location, condition adjustments she knows from market experience. She exports the results as a CSV. That file lands in a local directory the scripts read.
The price-analysis skill reads her CSV. It then queries Pitt County’s permit portal for the subject property’s address, checks for any open permits that could affect the closing timeline, and pulls deed records from the Register of Deeds to flag any restrictions the seller may need to disclose. It runs three valuation methods against her comps. It assembles the packet.
She reviews the packet before it goes anywhere. The Canon-Lock gate in the workflow means nothing ships until she says go. That is not a best-practice suggestion — it is enforced in the system architecture.
The result: a packet that took hours to assemble manually now takes minutes. And her name is on it because her judgment is in it.
Want to see the full workflow map? We broke down how we built our AI stack — and what it actually does versus what most AI tools claim — in the AI Prompts for Agents resource guide.
Skill-by-Skill Data Source Comparison
The table below maps each documented skill to its data source, trigger type, and the agent action required before output is usable. This is the reference sheet the overview page does not include — bookmark it for your own implementation planning.
| Skill Name | Data Source | Trigger Type | Agent Action Required Before Output |
|---|---|---|---|
| price-analysis | Agent MLS export + Pitt County permits + GIS + Deeds | Manual (agent supplies CSV) | Comp selection, final pricing review, Canon-Lock approval |
| listing-remarks | Agent-supplied property notes | Manual (agent supplies input) | Review and edit before MLS upload |
| listing-presentation | Agent pricing opinion + assembled math | Manual (agent supplies pricing conclusion) | Review before seller conversation |
| pull-permits (cron) | Pitt County EPL portal + open data | Scheduled / autonomous | Agent interprets results in context of transaction |
| pull-pitt-records | Pitt County OPIS GIS + ArcGIS + Register of Deeds | Scheduled / autonomous | Agent reviews for disclosure relevance |
| market-intel cron | Public market data sources, structured JSON output | Scheduled daily | Content team cites only from the verified JSON file |
| deed-check | Register of Deeds, covenants, amendments | Manual trigger with parcel input | Agent determines disclosure implications |
| hoa-check | Public HOA records and county data | Manual trigger with address input | Agent verifies against seller-provided HOA documents |
Notice what is not in this table: there is no skill that logs into an MLS on your behalf, no skill that selects comparables for you, and no skill that produces a CMA without your comp data as the input. Those capabilities do not exist in the repo — and their absence is the compliance design, not a limitation.
For a deeper look at how the repo stacks up against building raw prompt chains from scratch, the agent resource guide on brokerage systems and tools covers the tradeoffs in plain terms.
Where Your License Line Sits — and Why It Matters
This section is the one most AI tool marketing skips entirely — and it is the one that should anchor every decision you make about your AI stack.
NAR Code of Ethics Article 11 places the opinion of value on the licensee personally. Not on the tool. Not on the brokerage. On you. That means comp selection is yours. Adjustments are yours. The final price recommendation is yours. Automation that performs those steps on your behalf is not a productivity tool — it is a liability transfer you did not sign up for.
The Claude skills repo for real estate agents is designed around that constraint. The price-analysis skill does not select your comps. It reads the comps you already selected. It does not recommend a list price. It runs valuation math against the comps you brought and presents the result for your review.
The permit-pull and public-records skills operate entirely outside the licensed opinion-of-value workflow. They retrieve factual public data — permit status, deed covenants, parcel boundaries — that you would otherwise chase by hand through county portals. They do not interpret that data for you. They present it.
That design is also why the Canon-Lock gate exists in the price-analysis workflow. Nothing ships until the agent reviews and approves. A citation gate blocks any draft from reaching a live brand surface unless the numbers are traceable to the verified data file. These are architectural constraints, not suggestions.
If you are evaluating any AI tool for your real estate business — whether from the Claude skills repo or anywhere else — run this test: can you trace every number in the output back to a specific, verifiable data source? If the tool cannot answer that question, you should not put your name on its output.
What This Means in the Broader AI Tools Conversation
Real Trends published a roundup of twenty AI tools for real estate agents this week. Lists like that are useful for awareness — but they rarely answer the data-source question. And in 2026, the data-source question is the one that separates tools that help you from tools that create exposure.
The agents who are winning with AI right now are not the ones running the most tools. They are the ones who built one documented workflow that runs on their own data — their MLS export, their county’s permit portal, their own CRM notes — and who can explain exactly what the automation did and did not do at any point in the process.
That is what the open-source Claude skills repo for real estate agents was designed to enable. Not a magic button. A documented, auditable, agent-controlled workflow that does the assembly work so you can spend more time on the analysis and the relationship.
Billionaires are flooding into AI infrastructure right now — data centers, training compute, inference capacity. That capital is moving because AI is becoming the operating layer of every industry. Real estate is not exempt. The agents who understand their tools at the data-source level will be the ones who use that infrastructure without becoming dependent on it.
Related: if you want to understand how the SEO and content side of our AI stack runs — specifically how the keyword tracking cron feeds our publishing decisions — the YouTube Starter Kit walks through how we use scheduled automations to drive consistent content output without manual daily decisions.
How to Start Running These Skills Today
If you are an agent who is already producing and you want to add the Claude Code skills workflow to your business, the entry point is simpler than most technical content makes it sound. You do not need to be a developer. You do need to be willing to run a skill the first time, see what it returns, and decide what it means for your transaction — exactly the way you would evaluate any other data source.
Here is the practical sequence for getting started with the price-analysis skill specifically:
- Export your comps from your MLS. Your filters, your selections, your export. Save the CSV locally.
- Point the skill at your export file. The skill reads the path you specify. It does not access your MLS directly.
- Supply the subject property address. The permit-pull and public-records pulls run against that address on Pitt County’s public portals.
- Review the assembled packet. Check the valuation math against your market knowledge. Adjust your pricing opinion as needed. Approve before anything goes to the seller.
The first time through, budget an hour. By the third transaction, the assembly step is genuinely minutes. The time you recover goes back to comp selection — which is where your expertise actually lives.
For agents outside Pitt County: the public-records and permit pulls are written for the specific portals available here. The price-analysis math and the listing-remarks and listing-presentation skills are market-agnostic — they run on whatever comps you supply and whatever property details you provide. The public-data skills require adaptation for your county’s specific portals, which is part of why the repo is open-source. You can see exactly what each script queries and modify it for your market.
The CEO Time Audit resource is a useful companion here — before you add any automation layer, knowing exactly where your hours are going tells you which skill will have the highest immediate impact on your production time.
And if you want to walk through the full stack — what runs on a cron, what runs on manual trigger, how the compliance gates work, and how we have adapted these skills across multiple market conditions — Al and Victoria Pinder are available for a direct conversation. Book a strategy call at calendly.com/theprosperityagent or grab the Blueprint at theprosperityagent.com/resources/blueprint/.
You spent years building your market knowledge. The Claude skills repo for real estate agents is designed to make that knowledge move faster — not to replace it.
Frequently Asked Questions
What is the open-source Claude skills repo for real estate agents?
It is a library of documented procedures — called skills — that real estate agents run inside Claude Code. Each skill handles a specific task: pulling public permit records, assembling a pricing packet from agent-supplied comps, drafting listing remarks, or retrieving deed covenants from county databases. The repo is free, MIT-licensed, and built around a compliance-safe design where agents supply the licensed data and the automation handles public-records retrieval and assembly.
No. Claude Code does not log into any MLS, Flexmls, or listing platform in this system. The agent runs the MLS themselves, exports their selected comps as a local file, and the price-analysis skill reads that file. There is no API connection to the MLS. The agent’s comp selection and judgment are never delegated to the automation — that separation is the core compliance design of the repo.
Which Claude skills for real estate agents run without manual input?
The permit-pull cron, public-records pull, and daily market intel cron run on a schedule without manual agent input. They query open government portals — Pitt County EPL, OPIS GIS, ArcGIS, and the Register of Deeds — and return structured public data. They do not produce an opinion of value, select comparables, or recommend a price. The agent reviews their output in the context of each transaction.
How does the Claude skills repo protect an agent’s license compliance?
The repo is designed so that the opinion of value — the step NAR Article 11 places on the licensee personally — always requires agent input before it runs. The price-analysis skill reads agent-supplied comps; it does not select them. A Canon-Lock gate blocks any pricing output from shipping until the agent approves. Permit and public-records pulls return factual data for agent interpretation, never recommendations. Nothing about the system bypasses the agent’s professional judgment.
Can agents outside Pitt County use these Claude skills?
The price-analysis math, listing-remarks, and listing-presentation skills are market-agnostic — they run on whatever comps and property details the agent supplies. The permit-pull and public-records scripts are written for Pitt County’s specific portals. Agents in other markets can modify those scripts for their own county’s open-data portals, which is one reason the repo is open-source: you can read exactly what each script queries and adapt it for your jurisdiction.
What is the difference between the Claude skills repo and a general AI prompt library?
A prompt library gives you text you copy into a chat window. The Claude skills repo gives you documented procedures — code and workflow logic — that run in Claude Code as a structured skill, accept specific data inputs, call external public-data sources, and produce structured outputs. The difference is the difference between a recipe and a kitchen: the repo is a system, not a collection of starting points.
How long does the price-analysis workflow take once the agent exports their comps?
Based on how the workflow runs in practice: after the agent exports their comps and points the skill at the local file, the permit pull, public-records retrieval, valuation math, and packet assembly typically complete in minutes. The first time through the setup takes longer — expect roughly an hour for initial configuration. By the third transaction, the assembly step is a background task. The time the agent recovers comes directly from the research and formatting work that used to be done manually.