ARELLO · AI Working Session

Arello · Real Estate Investigators

Never assume.
Always verify.

How to turn an AI assistant into a disciplined partner for investigative work — one that cites its sources, flags what it doesn't know, and pushes back when you're wrong.

Takeaway page — everything from the session Copy the blocks · paste into your AI tool
The idea to leave with

Why this matters

The future belongs to those who can clearly and accurately define and articulate their needs and goals in a concise, actionable way.

An AI is only as good as the instruction you give it. The skill that separates a useful tool from a dangerous one isn't technical — it's the ability to say precisely what you want, name what you already know, and demand proof for the rest. Vague in, vague out. The quality of the answer is capped by the quality of the ask.

The foundation

Your custom instructions

Paste this once into your tool's persistent settings — ChatGPT's Custom Instructions, Claude's Preferences / Personalization, or Gemini's Saved info. It then shapes every conversation without you retyping it. Replace the [bracketed] parts.

custom-instructions · full 
I am [NAME], a real estate investigator with [AGENCY] in [JURISDICTION]. My work involves compliance, enforcement, and investigating potential violations of license law. My work product may be reviewed by attorneys, boards, or courts — treat accuracy, traceability, and defensibility as the highest priorities.

TRUTH & SOURCING
- Verify before you state. When you make a factual claim, cite the source (statute, rule, case, document, URL) so I can confirm it independently.
- Separate and label (a) established fact, (b) reasonable inference, and (c) speculation.
- If you do not know, say so plainly. Never guess, fabricate, or invent citations, statute numbers, case names, or quotes. "I can't verify this" is always an acceptable answer.
- Flag every ambiguity and every assumption you had to make.

CHALLENGE ME
- Push back when I state an incorrect assumption, draw an unsupported conclusion, or ask a question built on a flawed premise. Say so plainly and explain why. Do not agree just to be agreeable.

TEACH AS WE GO
- Explain your reasoning and the fundamentals in plain language; define legal, technical, or industry jargon. Reinforce key concepts even when I know them — but if I say "skip the background," drop it for that answer.

GET WHAT YOU NEED
- If more context would produce a better result, ask before answering.
- If missing documents, data, or details would make the research more complete or accurate, request them explicitly.
- When a task is ambiguous, ask a clarifying question before producing a long answer.

CONFIDENTIALITY
- Assume anything I share may be sensitive. Do not prompt me to share personal information (PII), case details, or confidential records beyond what is necessary, and warn me if I am about to over-share.

NOT LEGAL ADVICE
- You provide information and analysis, not legal advice or an official determination. For anything driving a legal or disciplinary action, tell me to confirm against the authoritative primary source.

Guiding rule: Never assume, always verify.
custom-instructions · compact 
I am [NAME], a real estate investigator at [AGENCY]. Verify before you state, and cite your sources. Separate fact from inference from speculation, and label each. If you don't know, say so — never invent citations, statutes, case names, or quotes. Push back when I'm wrong or working from a flawed premise. Explain in plain language and define jargon. Ask for missing context before answering. Don't prompt me to share confidential or personal information. You provide information, not legal advice — for anything actionable, tell me to confirm against the primary source. Never assume, always verify.

Tip: keep the compact version for tools with a character limit; use the full version everywhere else.

Read the fine print

What each rule is doing — and why

New to AI? Here's the reasoning behind each directive, so you can adapt them to your own work instead of copying blindly.

Cite the source
AI states everything with the same confident tone — true or not. Forcing a citation gives you something to independently check, and makes a fabricated answer easier to catch.
Label fact vs. inference vs. speculation
The most dangerous answers blend a real statute with a plausible guess. Making the model sort them for you keeps you from building a case on the guess.
"I can't verify this" is acceptable
Left unchecked, AI would rather invent than admit a gap. Explicitly permitting "I don't know" reduces confident fabrication — the failure mode that burns investigators.
Push back on me
A yes-man tool quietly confirms your bias. An assistant told to challenge flawed premises becomes a second set of eyes instead of an echo.
Teach, define jargon
You can't verify what you don't understand. Plain-language explanations let you audit the reasoning, not just accept the output.
Ask before answering
A model that guesses at missing context produces confident, wrong work. One that asks first forces you to articulate the task — which is where the real quality comes from.
Confidentiality guard
Consumer AI tools may retain or train on what you paste. A standing instruction to warn you before over-sharing is a cheap safeguard for sensitive case material.
Not legal advice
The model is an analyst, not counsel. Anchoring every actionable answer back to the primary source keeps your work defensible.
The core skill

Anatomy of a strong request

Instructions set the ground rules once. This is how you frame each individual ask. Five parts — the more you name, the better the answer.

01

Context

Who you are, what you're working on, the background that matters.

02

Goal

The one specific outcome you want. Not "help with X" — a defined deliverable.

03

Constraints

Format, length, tone, and what to avoid or exclude.

04

Assets

The documents, data, or examples it should actually work from.

05

Check

How you'll verify it — what a good answer must include or prove.

The same ask, framed well

Context: I'm investigating a possible failure-to-disclose in [state].

Goal: a one-page summary of the seller's written disclosure obligations for an internal case memo.

Constraints: plain language, under 400 words, no legal advice.

Assets: the statute text I'm pasting below.

Check: cite the exact section for each obligation, and flag anything ambiguous.

See the difference

Vague vs. articulate

✕ Vague Tell me about disclosure rules.

No jurisdiction, no goal, no format. You'll get a generic essay that may mix states, invent specifics, and can't be checked.

✓ Articulate I'm investigating a possible non-disclosure in [state]. Summarize the seller's written disclosure duties under that state's license law in under 400 words, cite each statute section, and flag anything ambiguous. For an internal memo.

Scoped, sourced, and checkable. The model knows exactly what "done" looks like.

Ready to use

Starter prompts

Reliable openers that build the good habits in. Tap copy and adapt.

Before you start

Research first — don't work from stale assumptions

The tools, rules, and best practices around AI change monthly. Before any big project, spend ten minutes confirming your approach is current. Applying "never assume, always verify" to your own toolkit.

Before committing to a strategy — a new investigation workflow, a tool you'll rely on, a way of using AI — run a quick scan of what's changed in the last 30 days. What was true six months ago may already be outdated: models get more capable, features appear, guidance shifts, and yesterday's limitation may be today's solved problem.

Check the tool, not your memory
"AI can't do X" ages fast. Before ruling something out, verify against the current version — the capability may already exist.
Scan recent discussion
Ask what practitioners are saying this month about the tool or approach. Fresh, real-world reports beat a confident model working from old training data.
Confirm the fundamentals still hold
Rules, thresholds, and best practices move. A ten-minute recency check keeps your strategy and optimizations up to date before you build on them.
A recency-check prompt

Before I build [project/workflow], summarize what has changed in the last 30 days for [tool / topic] — new capabilities, updated best practices, and any recent cautions. Cite sources with dates, and flag where the picture is still unsettled.

Read this twice

The one non-negotiable

Guard confidential and personal information

Investigators handle case files, complainant details, and subject records. Many consumer AI tools may retain or train on whatever you paste. Treat every input as if it could resurface.

  • Prefer an enterprise / business tier with a no-training data policy, or your agency's approved tool.
  • Strip or redact names, addresses, and identifiers you don't strictly need for the task.
  • Never paste anything you couldn't defend disclosing — assume it leaves your control.
  • The custom instructions above tell the model to warn you before you over-share — but the responsibility stays yours.