TL;DR
On this page
For the comprehensive overview, see the Using AI to Generate Professional Freelance Documents: The Complete Guide (Contracts, Proposals & Invoices).
Your client asked you to sign an NDA, or you need to send one to a subcontractor before handing over access. You ask ChatGPT, and ten seconds later you have a clean non-disclosure agreement with proper-looking clauses. It reads like a real legal document.
It is also the product of four decisions the AI made without asking you. It chose a one-way agreement. It picked a vague survival period. It defined confidential information with a catch-all phrase. And it almost certainly skipped governing law. Each of those is a real choice with real consequences, and the model defaulted on all four. This post gives you a prompt that forces the questions, then walks through the four decisions you confirm before you send.
The prompt that drafts a real NDA
Paste this into ChatGPT, Claude, or Gemini. The rules force it to ask the questions it normally answers for you.
You are an expert at drafting freelance NDAs. Draft a non-disclosure
agreement from the details below.
PARTIES:
- Disclosing party: [who shares the confidential info]
- Receiving party: [who must keep it secret]
- Is this MUTUAL (both sides share secrets) or ONE-WAY? [answer]
PROJECT: [one line on the engagement]
Rules:
1. If I did not say mutual or one-way, ask me before drafting.
2. Define "Confidential Information" with a SPECIFIC, enumerated list
(e.g. source code, client lists, pricing, designs), not a catch-all
like "any information disclosed."
3. State the survival period: how long confidentiality lasts after the
agreement ends. Ask me for the number of years if I did not give one.
4. Include standard carve-outs: information that is public, already
known, or independently developed is not confidential.
5. Add a governing-law line and ask which state or country applies.
6. Plain English. Flag any clause that depends on my jurisdiction.
Output the NDA, then list the decisions you made that I should confirm.
AI is genuinely good at the drafting itself. As the team at Spellbook describes the appeal:
Today, AI-powered NDA generators turn the hours-long NDA drafting marathon into a 10-minute sprint.
Source: Spellbook, "AI for Drafting Customized NDAs"
The speed is real. The judgment is what you still supply, in the four places below.
Decision 1: one-way or mutual
Ask for an NDA and ChatGPT hands back a one-way (unilateral) agreement that protects whoever is disclosing and binds whoever is receiving. That is correct when a client is sharing trade secrets with you and nothing flows the other way. But many freelance engagements are two-way: the client shares its customer data, and you share your methods, pricing, or proprietary process. In that case a mutual NDA fits, and a one-way agreement leaves your own confidential material unprotected. A common freelance case: a fractional CMO who receives the client's customer list while bringing a proprietary campaign framework of their own needs the agreement to bind both directions, not just protect the client. The prompt above forces the model to ask. If you are unsure which applies, the mutual versus one-way breakdown walks through the common freelance scenarios.
Decision 2: how long confidentiality survives
The survival period is the clause AI is vaguest about, and it is the one a reviewing lawyer flags first. There is no single right answer. Per Corporate Finance Institute, an NDA is generally in force for 1 to 2 years, sometimes ending on completion of the transaction. Template providers set a wider band: PandaDoc notes parties commonly agree to preserve confidentiality for 1 to 10 years depending on sensitivity. Source code and trade secrets justify the long end; routine project details sit at the short end. Pick a number deliberately rather than accept whatever the model inserted. The duration guide covers how to choose by project type.
Decision 3: what counts as confidential information
A weak NDA defines confidential information as "any information disclosed." Courts and reviewing lawyers prefer a specific, enumerated definition. CFI describes confidential information as data or information that cannot be obtained from public sources, and good templates list the actual categories: business plans, pricing, source code, client lists, designs. The prompt forces an enumerated list, but check that the list matches your project. Equally important are the carve-outs, which say that information already public, already known to the receiving party, or independently developed is not covered. AI sometimes drops these, and without them the definition is overbroad. A concrete example: if you are a developer, the enumerated list should name source code, repository access, and architecture notes, while the carve-out makes clear that a general programming technique you already knew is not suddenly the client's secret. Specificity on both sides is what makes the clause enforceable rather than decorative.
Decision 4: governing law
ChatGPT almost never asks which jurisdiction's law governs the agreement, and it matters the moment you and the client are in different states or countries. A confidentiality clause that is enforceable in one jurisdiction may be read differently in another. A US-based freelancer and a UK client, for instance, should name one governing law rather than leave a court to decide later, because terms written for one legal system do not always map onto the other. The prompt adds a governing-law line so the document names a single applicable law instead of leaving it blank.
One caution applies across all four decisions: drafting an NDA in a chatbot can expose the very thing the NDA exists to protect. One analysis found that sensitive data makes up 11% of what employees paste into ChatGPT, so paste a generic description while you draft and add the real subject matter privately afterward.
Here is what the AI defaults to versus what you set:
| Decision | AI default | What to set |
|---|---|---|
| Direction | One-way (discloser only) | Mutual, if both sides share secrets |
| Survival period | Vague or missing | A deliberate term, one to ten years by sensitivity |
| Confidential information | Catch-all "any information disclosed" | An enumerated list plus public/known carve-outs |
| Governing law | Omitted | The state or country whose law applies |
Read it once, because it is still a draft
The four fixes are the predictable gaps. The standing rule with any AI legal text is to read the whole thing before relying on it. As Pactly puts it:
Any text generated by an AI, no matter how sophisticated, must be treated as a draft and requires final legal oversight.
Source: Pactly, "Can I Use ChatGPT to Write a Legal Contract?"
For a routine NDA, the prompt plus the four confirmations plus one read is enough for most people. For a high-stakes confidentiality arrangement, get one lawyer review and reuse that version.
pro tip
Save the prompt with your standard clauses pre-filled. The only parts that change per NDA are the parties and the project line, so a reusable prompt that already specifies your preferred survival period, your enumerated definition of confidential information, and your governing law turns each new agreement into a quick fill-in rather than a fresh round of four decisions.
Or skip the defaults entirely
The prompt works, and the four-question version above is far better than a bare request. The friction is the same as every AI-document workflow: you answer the questions, fix the draft, paste it into a document, and reformat, on every NDA.
If you would rather start from a document where the direction, survival period, definition, and carve-outs are already set up correctly, FreelanceDesk builds the agreement with those defaults in place, and because it generates locally in your browser, the confidential subject matter never travels to a third-party model. It is free. The NDA fundamentals guide covers when you need one at all, the red-flag clauses post covers what to catch in any draft, and once confidentiality is handled, the service contract is the document that covers scope and payment.