Guide

Set up your AI agent

A practical, vendor-neutral walkthrough for turning a general chatbot into something that actually works the way you do, and when you're ready, one that can take action on its own.

Beginner-friendly~20 minutes to a first versionNo code required

Start in settings

Make it sound like you, and keep it honest

Before anything else, paste the block below into your agent's custom instructions / settings. The first line is yours to fill in so it sounds like you; the rest forces truthful, sourced answers and cuts confident guessing.

Paste into your agent's settings

[Provide details about yourself, the tone you want used most of the time, and your directness / fluff level.]

You are committed to truth and accuracy above everything else, including being helpful. A wrong answer delivered confidently is worse than no answer. Follow these 7 rules in every response:

1. UNCERTAINTY: If you are not fully certain about something, say so clearly. Use phrases like "I am not certain, but..." or "You may want to verify this...". Never state guesses as facts.

2. SOURCES: Do not invent paper titles, author names, URLs, or book references. If you cannot name a real, verifiable source, say "I do not have a verified source for this."

3. STATISTICS: Flag any number you are not 100 percent confident in. Say "approximately" and recommend I verify it from a primary source.

4. RECENT EVENTS: Remind me when a topic may have changed since your knowledge cutoff. Do not present outdated info as current.

5. PEOPLE and QUOTES: Never attribute a quote to a real person unless you are certain they said it. If unsure, say "I cannot confirm this quote is accurate."

6. CODE and TECHNICAL: Never invent function names, library methods, or API syntax. If unsure a function exists, tell me to verify it in the current docs.

7. LOGIC GAPS: Do not fill missing context with assumptions. If something is unclear, ask a clarifying question before answering.

If a response would require breaking any of these rules, choose honesty over helpfulness every time.

First, what is an "AI agent"?

Skip the hype. For our purposes an AI agent is just a model you've set up in two layers:

An assistant that knows your context: you've given it a role, instructions, and reference material, so it stops answering like a stranger and starts answering like a teammate.
(Optional) An assistant that can act: you've connected it to your tools so it can do things like file a note, draft a follow-up, or update a record, without you copy-pasting.

Most people only need the first layer to save real time. The second is where it starts to feel like a genuine "agent." Build them in that order.

One caveat that applies to this whole guide: exact menu names and button labels change constantly across ChatGPT, Claude, Copilot, and Gemini. Treat the steps below as the shape of the build and confirm the current wording inside whichever tool you're using.

The build, step by step

Pick your model

Match the tool to your firm's policy and the sensitivity of your data first, capability second. If confidential client or deal data is involved, you need an enterprise tier, not a free consumer login. See the three questions to ask before you commit and the platform reference on the Tools page.

Give it a home

Don't work in one-off chats. Create a dedicated workspace: a Project in ChatGPT or Claude, or a Custom GPT, so everything it learns stays in one place. Name it for the job it does ("Lease abstractor," "Listing copywriter") rather than something generic.

Write its instructions

This is where most of the value lives. Tell it who it is, who it's helping, the tone and format you want, and what to avoid. A simple framework from Session 3, the Frame · Focus · Finish method: frame the task with full context up front, focus by correcting its first draft, finish by checking facts and applying judgment.

Feed it the right knowledge, and only that

Upload the reference material it needs to sound like you and your firm: templates, brand voice samples, a sample lease, a past memo. Then stop. Don't dump your whole drive. Anonymize and redact names, account numbers, and sensitive terms before uploading, and prefer a tier that doesn't train on your inputs.

(Optional) Let it take action

When the assistant is solid, connect it to your tools so it acts, via automation platforms like Zapier, Make, or n8n. One thing that trips everyone up: connecting a model to an automation requires the model's API (a separate account with its own billing). Your logged-in consumer chat doesn't link to these platforms. See the Automation Playbook for copy-this examples.

Keep a human in the loop

An AI agent is a prediction machine, not a thinking machine. It can't guarantee accuracy, make the call for you, or be trusted with confidential data by default. Review every output that leaves your hands. The goal is to outsource the mental load, not your judgment.

A 20-minute first version

Your starting checklist

Choose one painful, repetitive task that takes more than 15 minutes.
Open your approved model and create a Project (or Custom GPT) named for that task.
Paste in a short role + instructions: who it helps, tone, output format.
Upload one or two reference files (a template and a good past example).
Run the task, correct the first draft, and save what worked back into the instructions.
Once it's reliable, decide whether step 05 (taking action) is worth wiring up.

Keep teaching it

The agent gets better the more of your finished work it sees. Each time you complete something you're proud of, upload it back into the agent's knowledge files. Those approved pieces become its reference set for your voice and standards, so every new draft starts closer to done. Keep that library current as your style evolves, and remove anything you would no longer want it to imitate.

Keep going

Session 03

Want a hand setting yours up? Bring it to a session, or reach out through the member survey and tell us what you're trying to build. We're figuring this out together.

← Back to Tools & Resources