A custom GPT is a version of ChatGPT you preload with your own instructions and documents, then share with anyone by link. No code. If you can write an email and upload a PDF, you can build one in 30 minutes, and it will answer the same repetitive questions your staff answers every day, in your firm’s voice, following your firm’s rules.
I’ve built dozens of these across my 10 firms. The one I’m best known for is the Law Firm Intake Specialist GPT, which lawyers across the country use through My Legal Academy. It takes raw notes from a lead call and qualifies the case against defined criteria, the same way a trained intake specialist would, except it never has a bad day and never forgets to ask about the repair history.
Here’s what a custom GPT is, the exact build process, and the one rule that keeps you out of trouble.
What a custom GPT actually is
Regular ChatGPT starts every conversation blank. A custom GPT starts every conversation already knowing three things: who it is (your instructions), what it knows (documents you upload as knowledge files), and how it should respond (format rules you set once).
Build it once and the setup is baked in. Your paralegal doesn’t need to know how to prompt. They open the GPT, type their question, and get an answer shaped by rules you wrote weeks ago. That’s the entire point: prompting skill stops being a requirement for your team and becomes a one-time investment by whoever builds the GPT.
You need ChatGPT Plus at $20/month to build one. If you’re new to ChatGPT itself, start with the full ChatGPT guide and come back.
The GPT that runs my intake qualification
My Law Firm Intake Specialist GPT does one job: it takes whatever an intake agent captured on a call, messy notes and all, and checks it against case criteria. Vehicle under warranty? Enough repair attempts? Purchased in a state we cover? It returns a structured verdict: qualified, not qualified, or missing information, with the specific gaps listed.
Before this, qualification quality depended on which intake agent picked up the phone. New hires missed disqualifiers that veterans caught. Now the criteria live in one place, every lead gets checked the same way, and training a new intake agent means showing them one link.
That’s the pattern for every GPT worth building: find a task where the answer depends on documents your firm already has, and where consistency matters more than judgment.
The 30-minute build
Go to chatgpt.com, click your name, then “My GPTs,” then “Create a GPT.” Use the Configure tab, not the chat-based builder, so you control exactly what goes in. Then:
-
Name it for the job, not the firm. “Intake Qualifier” beats “Smith Law Assistant.” Your team should know what it does from the name alone.
-
Write the instructions block. This is the whole game. Copy this template and fill in the brackets:
You are the intake assistant for [FIRM NAME], a [PRACTICE AREA] firm.
Your job: [one sentence, e.g., "qualify potential clients against
our case criteria and produce a structured summary for the intake team"].
Rules:
1. Ask one question at a time. Never dump a list of questions.
2. Follow the criteria in the attached files exactly. Do not
improvise qualification standards.
3. When done, output a summary in this exact format:
[Name / Matter type / Qualified: yes-no-missing info / Gaps / Next step]
4. Never give legal advice. If asked, say every case is reviewed
individually by our attorneys.
5. If a question falls outside these instructions, say you don't
know and refer the person to [CONTACT].
Tone: plain English, warm, no legal jargon.-
Upload knowledge files. Three documents cover most firm GPTs: your client FAQ, your fee structure, and your intake criteria. PDFs or Word docs both work. The GPT will quote from these instead of guessing.
-
Test it like a skeptical client. Ask the questions your intake team hears daily. Ask something outside its scope and confirm it declines instead of improvising. Ask it for legal advice and make sure rule 4 holds. Ten minutes of testing here saves you an embarrassing screenshot later.
-
Share the link. Set access to “Anyone with the link,” send it to your team, and pin it wherever your team lives, Slack or Teams. Done.
Five GPTs worth building first
Build these in order. Each one takes the same 30 minutes.
- Intake qualifier: checks leads against your case criteria, like mine does.
- Client FAQ: answers the 25 questions your staff repeats every week, from your own FAQ document.
- SOP assistant: upload your procedures, and new hires ask it instead of interrupting a senior paralegal.
- Letter drafter: feed it five of your best letters as knowledge files and instruct it to match their structure and tone.
- Marketing writer: your services, your differentiators, your voice rules, so every post sounds like your firm instead of a content mill.
Custom GPT vs Claude Project vs full agent
A custom GPT is one option among three, and picking wrong wastes either money or capability.
| Custom GPT | Claude Project | Full AI agent | |
|---|---|---|---|
| Setup time | 30 minutes | 30 minutes | Days to weeks |
| Sharing | One link, whole team | Workspace members on a Team plan | Built per workflow |
| Cost | $20/mo (ChatGPT Plus) | $20/mo (Claude Pro) | Hundreds/mo and up |
| Best for | Team-facing Q&A and process tools | Long documents and serious drafting | Multi-step work across your systems |
The short version: GPTs for anything your whole team touches, Claude Projects for drafting work where writing quality matters, and a full agent only when the task spans multiple systems, like reading email, updating your CRM, and sending follow-ups without a human in the middle.
The confidentiality rule
One rule, no exceptions: never upload client files or PII to a shareable GPT.
Knowledge files are not a vault. A user who knows what to ask can get a GPT to reproduce large chunks of its uploaded documents. People have extracted knowledge files from public GPTs as a hobby. So assume everything you upload could be read by anyone with the link.
That’s fine, because firm-process GPTs don’t need client data. Intake criteria, fee schedules, FAQs, SOPs, sample letters with names stripped: none of it is confidential in any way that matters. The moment a task requires an actual client file, move it out of the GPT and into a private chat or a Claude Project that isn’t shared.
Keep GPTs generic and process-focused and the confidentiality question mostly answers itself.
Build the first one today
Pick the question your staff answers most often. Open ChatGPT, create a GPT, paste the template above, upload your FAQ, and test it with five real questions. That’s tonight’s project, and it costs nothing beyond the $20 subscription you should already have.
Then send the link to one person on your team and watch how fast they stop asking you.