Skip to main content
Start free
Start free

Custom instructions for your landlord AI: a template for your rules

DoorKeeping team

3 min read

Write your rules once and the assistant follows them in every chat: your state, how you word things, what it must ask before doing, and what it should never include. In Claude, project instructions do this for a whole project. The template below covers tone, tenant privacy, state law, fair treatment and approvals, and is short enough to keep current.

Why standing instructions help

Without them, you repeat yourself in every chat: "I am in Ohio", "keep it short", "do not use the tenant's full name". Mistakes creep in when you forget one. A standing set of instructions fixes the context once. Anthropic describes projects as self-contained workspaces with their own chat histories and knowledge bases, and says you can define project instructions for each project to tailor Claude's responses, such as a more formal tone or a role (Anthropic: What are projects?). ChatGPT has its own places for standing instructions, and their names change, so check OpenAI's help for the menu today.

What belongs in the instructions

  • Who you are and where: "I am a landlord with 14 units in [state]. Follow [state] and [city] rules, and say when you are unsure which applies."
  • Voice: short sentences, polite, no threats, no sarcasm, no legal jargon.
  • Facts only: "Use only facts I give you. If a fact is missing, ask. Never invent amounts, dates or laws."
  • Citations: "When you state a legal rule, name the statute and quote the sentence, or say you do not know."
  • Privacy: "Do not ask for Social Security numbers, bank account numbers or ID photos. Use unit numbers instead of names."
  • Fair treatment: "Apply the same standards to every tenant and applicant. Never suggest treating anyone differently because of a protected class."
  • Approvals: "Never send, sign, pay or file anything. Show me a draft and wait for my yes."

A template to copy

Example instructions: "You help me, a small landlord in [state], with my rentals. Write short, calm, polite messages with no threats. Use only facts I give you and ask when something is missing. When you state a law, name the statute and quote the sentence; if you are not sure, say 'I don't know'. Never ask for or repeat Social Security numbers, bank details or ID numbers. Use unit numbers in place of tenant names when you can. Apply the same standards to every tenant and applicant. Never send, sign, pay or file anything: show me a draft and wait for my yes. Keep answers under 200 words unless I ask for more."

Edit it for your own habits. If you do not want tables, say so. If you want every answer to end with "what I could not check", add that.

What instructions cannot do

Instructions are a request, not a lock. A model can still make an error, drift in a long chat or misread a rule. They do not replace the controls in the software behind a connector. If a tool can change your records, leave the assistant's tool approvals on and let the software require its own approval for money and legal steps.

Review the instructions every few months and after any mistake. If the assistant did something you did not want, add a line that forbids it. If a line is never used, delete it, because long instruction lists get harder for anyone to follow, people and models alike.

Test the instructions before you rely on them

Start a new chat and try three things: ask for a message to a tenant with a name in it, ask for a legal rule, and ask it to send something. A good set of instructions makes the assistant use a unit number, quote a statute or admit it does not know, and offer a draft instead of sending. If one fails, make that line more specific and test again. Keep the test questions in a note and run them after each change.

How DoorKeeping handles it

Your own instructions and DoorKeeping's rules work together. DoorKeeping's tools are grouped as reads, drafts you say yes to, and items that wait for your tap on doorkeeping.com, so "never send, sign, pay or file" matches how the software already behaves. A connected assistant can reject an item in your approvals queue but cannot accept one.

Sources

  1. Anthropic: What are projects?, read Oct 30, 2026
  2. DoorKeeping: Connect Claude or ChatGPT (steps and tools list), read Oct 30, 2026

Questions

Where do I put the instructions in Claude?
In a project, under its instructions. Anthropic's page describes defining project instructions for each project.
Will the assistant always obey them?
Not always. Treat them as a strong request and keep approvals switched on for anything that changes records.
Should I include tenant names in the instructions?
No. Instructions apply to every chat. Keep names and private details out of them.

Keep reading