Custom GPT: Build a Firm-Voice Drafting Assistant for Client Documents

What This Builds

Every demand letter starts the same way: a blank screen, a vague memory of how the last one read, and ten minutes spent reconstructing your own firm's tone before you even get to the facts of the case. A Custom GPT removes that reset. Once it is loaded with your firm's letterhead style, standard clauses, and preferred phrasing, you open it, describe the matter, and get a first draft that already sounds like your firm instead of a generic chatbot.

Prerequisites

• A ChatGPT Plus, Business, or Enterprise account (Custom GPTs are not available on the free tier)
• Three to five past documents you can use as style examples, either redacted or built from fictionalized facts
• A short written style guide, even an informal one: preferred terms, formatting habits, how the firm signs off
• Total ongoing cost: $20/month per attorney at minimum. If the firm needs conversations excluded from model training and admin-level audit controls, the Business tier costs more per seat. Check chatgpt.com for current business pricing before committing.

The Concept

Think of a Custom GPT as a new associate on their first day who already read the firm handbook. A normal ChatGPT conversation starts from zero: you explain your tone, your formatting quirks, and your standard clauses every single time. A Custom GPT stores that context permanently in its Instructions and Knowledge files, so it starts already knowing how your firm writes. You still review and sign off on everything it produces. It just skips the re-explaining.

Build It Step by Step

Part 1: Gather Your Style Materials

Before opening the builder, collect what the GPT will learn from:

1. Pick three to five documents that represent your firm's voice well: a demand letter, a client update email, maybe a short motion. Redact client names, addresses, case numbers, and any other identifying detail, or rewrite the facts entirely using a fictional client.
2. Write down your non-negotiable style rules. Common ones: "shall" instead of "will," how you capitalize defined terms, how you format dates, how you sign correspondence.
3. List your two or three most-used standard clauses (a basic indemnification clause, a standard scope-of-representation paragraph) as short text snippets, not full documents.

Part 2: Build the GPT

1. Go to chatgpt.com and open the GPT builder from the sidebar (look for "Create a GPT" or "My GPTs").
2. Switch to the Configure tab for direct control over each field rather than the conversational builder.
3. Name it something internal and specific, like "[Firm Name] Drafting Assistant." Skip a public-facing description since this GPT will stay private.
4. In Instructions, paste a system prompt structured like this:

You are a legal drafting assistant for [Firm Name], a small firm
handling [practice areas].

Tone: formal and professional. Conservative on legal risk language.
Use "shall" rather than "will" in operative sentences. Capitalize
defined terms consistently (e.g., "the Client," "the Agreement").

Before drafting, ask for: the client's name, the matter type, and
the jurisdiction. Ask one clarifying question at a time if
information is missing.

Default to the firm's standard clauses when the document type calls
for them, unless told otherwise.

Every draft must open with: "DRAFT - For attorney review before use."
This is not legal advice and has not been reviewed by an attorney.

Flag anything unusual or non-standard in a separate note after the
draft, rather than burying it in the document body.

1. Under Knowledge, upload your redacted style examples and clause snippets (up to 20 files, 512 MB and roughly 2 million tokens each, which is far more than a firm style library needs).
2. Leave Capabilities like web browsing and code interpreter off unless a specific use justifies them. A drafting assistant does not need internet access to your firm's confidential templates.
3. Under Actions, add nothing unless you have a specific integration in mind and have reviewed what data it would send externally.

Part 3: Test, Set Sharing, and Refine

1. Use the live preview pane to test with three or four realistic (but fictional) prompts: "Draft a demand letter for a slip-and-fall matter, client is Jane Doe, incident at a grocery store."
2. Read the output against your actual style examples. If the tone drifts, add a line to Instructions naming exactly what went wrong ("Avoid contractions" or "Do not use bullet points in correspondence").
3. When you click Create, choose Only me. Do not select the public GPT Store option. A shared link works if you want a specific associate to use it, but keep it out of the public directory entirely.
4. Revisit the Instructions after a week of real use and tighten anything that keeps needing manual correction.

Real Example: Building a Demand Letter Assistant

Setup: A two-attorney personal injury firm loads three redacted demand letters, a one-page style guide, and instructions requiring the "DRAFT - For attorney review" header on every output.

Input: "Draft a demand letter for a car accident matter. Client: fictional name for testing. Liability is clear, injuries include a fractured wrist, treatment ongoing, no settlement demand figure yet."

Output: A full first-draft letter in the firm's established structure and tone, with a placeholder for the demand figure and a flagged note asking for medical billing totals before finalizing.

Time saved: Drafting from scratch with proper firm formatting typically runs 30-45 minutes. This build cuts that to roughly 10 minutes of review and editing, across the dozens of similar letters a busy practice sends each month.

What to Do When It Breaks

• Output drifts back to generic chatbot tone after a few sessions → The Instructions field is too vague. Add specific "avoid" examples, not just "be formal."
• The GPT invents a clause or citation that isn't in your Knowledge files → Add a line telling it to say "no matching firm template found" instead of guessing when a clause type isn't in its knowledge base. Never send drafted language to a client or court without checking any cited authority yourself.
• A colleague you shared the link with pastes in real client details → Set a firm policy before sharing: redacted facts only, or fictionalized names, until you've confirmed your plan tier keeps data out of training. Business tier or higher is the safer default for anything client-adjacent.

Variations

• Simpler version: Skip the Knowledge file uploads and put your two or three clause snippets directly in the Instructions field. Less powerful, but faster to set up and easier to audit.
• Extended version: Build a second GPT scoped to a single practice area (estate planning, for example) with deeper knowledge files specific to that document type, rather than one generalist assistant for the whole firm.

What to Do Next

• This week: Run every new correspondence draft through the GPT for a week and note where it needed correction.
• This month: Fold those corrections back into the Instructions so the GPT stops making the same fix twice.
• Advanced: Pair this with the Level 3 discovery summarization workflow so case facts pulled from documents feed directly into a drafting prompt for the GPT.

Advanced guide for small firm attorneys. These techniques use more sophisticated AI features that may require paid subscriptions.