To build a custom GPT or Claude Project for your firm, you upload your own documents — precedent letters, style guide, standard operating procedures — write a set of plain-English instructions that tell the AI how to behave, test it with a few staff, then share it with your team. No coding. The whole first version takes an afternoon. A custom GPT needs a ChatGPT paid plan (Plus for yourself, or Team/Enterprise to share properly); a Claude Project needs a paid Claude plan (Pro for one person, or Team to share across staff).
This guide is for the practice manager or partner who wants one reusable assistant that already knows how the firm writes a client letter, runs an onboarding, or drafts a file note — instead of re-explaining the same context to a blank chat every time.
What is a custom GPT or Claude Project, in plain terms?
Both are the same idea with different badges. You take a general AI chatbot and bolt on three things: a permanent set of instructions, a small library of your own reference files, and a name. From then on, everyone who opens that assistant starts the conversation with the firm's context already loaded.
Think of it as the difference between a temp who needs the brief re-explained every morning, and one who already has the office manual on their desk. The assistant does not become smarter — it becomes consistent. That consistency is the whole point for a firm where ten people otherwise get ten slightly different letters.
Claude Projects vs custom GPTs for business: which should you pick?
For most professional services firms the decision comes down to what you already pay for and how much reference material you want the assistant to hold.
| Factor | Custom GPT (ChatGPT) | Claude Project |
|---|---|---|
| Plan needed | Plus to build; Team or Enterprise to share across staff | Pro to build; Team to share across staff |
| Uploaded files | Good for a modest set of reference docs | Handles larger bodies of reference text comfortably |
| Writing quality for letters | Strong, flexible tone | Often preferred for longer, nuanced drafting |
| Sharing model | Published to your workspace members | Shared within a Team workspace |
| Best when | You already run Microsoft or ChatGPT tooling | You want a document-heavy drafting assistant |
A practical rule: if your firm already standardises on Microsoft 365, look first at whether Copilot inside your tenant does the job before adding another subscription — see our comparison of Microsoft Copilot vs ChatGPT for accounting firms. If you are choosing between the two consumer-grade builders, pick the one your drafting staff prefer after a week's trial. Features on both move constantly, so treat any specific limit as a thing to check on the day, not a fixed fact.
What to put in the project instructions
The instructions are the brief you would give a capable new hire. Write them as if that person is sitting across the desk. Here is a starting template you can paste into either tool and adapt — replace the bracketed parts with your firm's reality.
You are a drafting assistant for [Firm Name], a [law firm / accounting practice / conveyancing / real estate] business in [city]. You help staff produce internal drafts only. Everything you produce is reviewed by a qualified person before it leaves the firm.
Tone: [plain, warm, professional — no jargon unless I ask]. Match the voice in the uploaded style guide. Australian English and Australian spelling always.
When I ask for a letter or email, use the matching template from the uploaded files as the structure. Keep our standard headings, sign-off and disclaimers exactly as written.
When you do not have enough information, ask me for it rather than inventing details. Never guess a date, a figure, a name or a legal/financial position.
Do not include real client personal information in your examples unless I have provided it in the chat. Flag anything that looks like it needs partner sign-off.
If a request falls outside drafting — advice that must be right every time, final figures, anything client-facing without review — say so and tell me to escalate it.
That last paragraph does a lot of quiet work. It trains the assistant to flag its own limits, which is exactly the behaviour you want staff to see modelled.
For the uploaded files, keep it tight and current. A bloated library makes answers worse, not better. Start with:
- Three to five precedent letters or emails that represent your best work
- Your one-page style guide (tone, banned phrases, how you address clients, sign-offs)
- The SOP for the task the assistant covers — the onboarding checklist, the file-note format, the engagement sequence
- Any standard disclaimers or compliance wording that must appear verbatim
Custom GPT for a law firm: worked examples
Abstract instructions are hard to picture, so here are assistants a small firm could realistically build in an afternoon.
- Client letter drafter (law firm). Loaded with precedent letters for common matters, it produces a first draft in the firm's voice. The lawyer edits and approves. Pairs well with the ready-made starters in our ChatGPT prompts for conveyancers.
- File-note and attendance assistant. Staff paste rough notes; the assistant formats them into the firm's standard file-note layout. If you capture meetings, read how this connects to AI meeting notes for professional services.
- Onboarding document-request builder. Give it the matter type and it lists exactly which documents to request, using the firm's wording.
- Accounting explainer. For a practice, an assistant that turns technical answers into client-friendly plain English without changing the substance.
Notice the common thread: every one produces an internal draft that a human finishes. None of them sends anything, files anything, or makes a decision.
How to test it with staff before you roll it out
Do not announce the assistant firm-wide on day one. Build it, then run it past two or three people who do the task daily — they will spot where it drifts off-voice or invents detail faster than you will.
- Run ten real-ish tasks through it. Use realistic but de-identified inputs, not live client files. Compare each output against how your best person would do it.
- Tune the instructions, not the files first. If the tone is wrong, strengthen the tone line. If it keeps dropping a disclaimer, add a sentence making that disclaimer mandatory.
- Deliberately feed it a bad request. Ask it for something it should refuse — a final figure, a piece of advice. Check it escalates instead of bluffing. If it bluffs, tighten the limits paragraph.
- Write a one-line "what this is for" note. Staff need to know the assistant drafts and never sends, and that every output gets reviewed.
For the wider habit-building that makes any firm-wide tool stick, our guide on training staff to use AI in a professional services firm covers the rollout side.
How to share a custom AI assistant with staff
Sharing is where plans matter. A personal Plus or Pro subscription lets you build an assistant but not properly distribute it to the team. To share:
- Custom GPT: you need a ChatGPT Team or Enterprise workspace. You publish the GPT to the workspace, and members open it from their sidebar.
- Claude Project: you need a Team workspace, where a Project can be shared with other members.
Resist the temptation to share one login around the office. Shared logins break the audit trail, violate most terms of service, and make it impossible to switch someone off when they leave. Pay for seats on a Team plan — it is cheaper than the problem it prevents. For how these subscription costs sit against the alternatives, see how much AI automation costs for a small firm in Australia.
What goes in, and what never does
This is the part most "build your own GPT" articles skip, and it is the part that matters most in professional services. The OAIC's guidance on commercially available AI products is clear that the Privacy Act applies the moment personal information goes into an AI system, and that publicly available chatbots should not be fed personal information. That shapes what belongs in a firm assistant.
Safe to upload as reference: precedent letters with client details removed, your style guide, SOPs, standard disclaimers, checklists. These are firm templates, not client data.
Never upload as reference files: actual client files, lists of client names, matter details, financials, anything identifying a real person. If a task genuinely needs client information, that belongs in the chat on an appropriate enterprise or business tier where your data is not used for training — not baked into a shared Project everyone can read.
Where client information must be processed at all, use only the fields the task needs, redact identifiers where you can, and keep a short written record of what data flows where. If any processing happens overseas, treat it as a cross-border disclosure under Australian Privacy Principle 8 and document it. This is general guidance, not legal advice — our full explainer on AI and client data under the Australian Privacy Act goes deeper.
Who owns it, and keeping it out of shadow IT
An assistant that no one owns quietly rots. Someone updates a letter template in the document system, nobody updates the Project, and six months later it is drafting from a stale precedent. Decide three things up front:
- An owner. One named person — usually the practice manager — maintains the instructions and refreshes the files when templates change.
- A review date. Put a quarterly reminder in the calendar to re-run a few test tasks and check the outputs still hold up. AI products change their behaviour without warning, so periodic testing is not optional.
- A line in your AI policy. The assistant should be a sanctioned tool, not a thing someone built on a personal account that the firm depends on without knowing. If you do not have a policy yet, our AI policy template for law firms in Australia gives you the checklist to slot this into.
When to stop and get help: if you find yourself wanting the assistant to pull from live systems, send emails, or make decisions that must be right every time, you have outgrown a custom GPT. That is a workflow automation project, not a chat assistant, and it needs proper tooling and review.
Start small — one assistant, one task, one owner. Prove it on internal drafting where a human always has the last word, then build the next one. If you want a shortlist of where a firm assistant pays off fastest, our free guide to 10 AI workflows to save 10+ hours a week maps the tasks worth starting with.
Common questions
Do I need a paid plan to build a custom GPT or Claude Project?
Yes. You can build a custom GPT on a ChatGPT Plus plan and a Claude Project on a Claude Pro plan, but to share either properly across staff you need a Team or Enterprise workspace. Sharing a single personal login around the office breaks terms of service and your audit trail, so budget for seats rather than passing one account around.
Can I upload real client letters as reference files?
Only after removing client-identifying details. Precedent letters stripped of names, matters and figures are firm templates and are fine to upload. Actual client files, names and financials should never be baked into a shared assistant everyone can read. If a task genuinely needs client data, enter only the required fields in the chat on an enterprise tier where data is not used for training.
Is a Claude Project or a custom GPT better for drafting client letters?
Both draft well. Many firms find Claude handles longer, nuanced documents and larger reference libraries comfortably, while a custom GPT fits neatly if you already run ChatGPT or Microsoft tooling. The honest answer is to trial both for a week with the people who actually do the drafting and let their preference decide, since features on both change constantly.
Who should own the assistant once it is built?
Name one person, usually the practice manager, to maintain the instructions and refresh the uploaded files whenever templates change. Set a quarterly reminder to re-test a few tasks, because AI products change behaviour without notice. List the assistant in your firm's AI policy so it is a sanctioned tool rather than something the firm quietly depends on but does not govern.
When should I stop using a custom GPT and build something else?
Stop when you want the assistant to pull from live systems, send emails, process payments, or make decisions that must be correct every time. A custom GPT or Claude Project is for internal drafts a human finishes and approves. Anything that acts on its own is a workflow automation project needing proper tooling, review and, usually, a developer or consultant.
Want these ideas working in your firm? We build controlled AI workflows for Australian professional services firms — starting with a free automation audit.