The skill behind this guide: AI Legal Assistant. A senior legal assistant for demand letters, motions, correspondence and day-to-day legal admin, including paralegal-tier filing prep and court forms - $19, one payment, yours permanently.
View the AI Legal Assistant →A fabricated citation is indistinguishable from a real one by reading. The case name is plausible, the reporter cite is plausible, the year is plausible, the holding is plausible and on point. Your instinct - does this look right? - is precisely the instinct that fails, because the output is optimised to look right. That single fact reorganises everything about using an AI chat for legal work, and it does not only apply to case law. It applies to quotations from the contract sitting in front of you.
This page is not legal advice, and neither is the skill
Nothing here, and nothing a model produces, is legal advice or a substitute for a qualified lawyer in your jurisdiction. The techniques below are for document work - drafting, summarising, organising, preparing questions - and every one of them ends with a human who is qualified and accountable. If a decision carries legal consequence, it belongs to a solicitor or attorney, not to a chat window and not to this page.
The verification instinct that does not work
In June 2023 the Southern District of New York sanctioned two lawyers and their firm over a brief containing six judicial opinions that did not exist. They had been produced by ChatGPT, complete with internal citations to further cases that also did not exist, and presented as published in the Federal Reporter and the Federal Supplement. Judge P. Kevin Castel imposed a $5,000 penalty and required the lawyers to notify the real judges whose names had been attached to the fabricated opinions. He described the episode as promoting "cynicism about the legal profession and the American judicial system".
The instructive part is not that it happened. It is how far it got. The brief was read, filed, and defended before anyone pulled the cases. Nothing about the text announced itself as invented, because a language model producing a citation is doing exactly what it does everywhere else: generating the most plausible continuation. Plausible is the whole problem.
So the defence cannot be a reading defence. It has to be mechanical: every authority is located in a real database and read, or it does not go in the document. Not skimmed for plausibility. Located. There is no shortcut that operates on the text itself, because the text is where the failure lives.
The part people miss: it also misquotes your own documents
Most people assume this risk lives with outside knowledge - cases, statutes, things the model half-remembers. It does not stop there.
Hand a model a sixty-page agreement and ask what the termination provisions say, and it can hand back a clause quotation that is subtly not what the document says. A shall that was a may. A notice period from a different sub-clause. A carve-out that is not there, or one that is there and has quietly gone missing from the summary.
This is more dangerous than an invented case, because you trust it more. You supplied the document. It feels like reading rather than recall. It is not: the summary is still generated, and generated text can drift from its source in ways that are invisible without going back to the source.
The operational rule follows: no statement about a document without a locator. Clause number, section, page. Nothing is quoted unless you can put your finger on it, and anything the model cannot locate is flagged rather than smoothed over. A summary with locators takes thirty seconds longer to produce and can actually be checked.
Jurisdiction is not a preference, it is the answer
These tools have read overwhelmingly more American legal text than anything else, and they do not announce it. Ask an English-law question and you may get American assumptions delivered in English-sounding prose, with the vocabulary drifting silently: discovery for disclosure, statute of limitations for limitation period, consideration doctrine applied where it does not belong, at-will employment assumed in a jurisdiction that has never had it.
And within one jurisdiction, the law has a date. A model does not know what changed last month, and it will not say so unprompted. State the jurisdiction and the date at the top of every conversation, and require it to say which assumptions it is making and where it is unsure.
The boundary is not something the model can police
It is common to hear that a well-built legal skill "knows the line between admin and advice". It cannot, reliably. Whether your question is administrative or advisory depends on facts it does not have and consequences it cannot see.
What you can do is put the burden in the right place, in how you ask. Never ask whether something is safe to sign, whether you have a claim, or what your rights are. Ask what the document says, what is unusual in it, and what you should put to a lawyer. The second kind of question has an answer the model can actually produce, and it converts its output into a brief for a human instead of a substitute for one.
Prompt 1 - intake, before anything else
Prompt 2 - summarise with locators
A senior legal assistant for demand letters, motions, correspondence and day-to-day legal admin, including paralegal-tier filing preparation and court-form completion. It holds your jurisdiction, your document conventions and your house rules between sessions, which matters here more than anywhere: the rules above only protect you if they are applied every time, and a fresh chat applies none of them.
View the AI Legal Assistant →Prompt 3 - the adversarial read
A summary tells you what a document says. This tells you where it can be used against you, which is a different question and the one that matters when something goes wrong.
Prompt 4 - turn the output into a brief for a human
Prompt 5 - drafting, with nothing asserted
Prompt 6 - the citation quarantine
Run this whenever any authority has appeared in the conversation at all, including ones you did not ask for.
Confidentiality, privilege, and who is allowed to do this
If you are a lawyer, your professional duties apply to this the way they apply to everything else. The American Bar Association's Formal Opinion 512, issued on 29 July 2024, is its first ethics guidance on generative AI: competence under Model Rule 1.1 includes understanding the benefits and risks of the technology you use; client information is protected under Rule 1.6 and needs informed consent before disclosure; Rule 1.4 brings a duty to consult clients about the methods used on their matter; and on fees, you may bill for time spent inputting material and reviewing output, but generally not for the time you spend learning the tool. Read the opinion itself, and your own jurisdiction's guidance, which may be stricter. Source: ABA.
If you are not a lawyer, using this on your own documents is your business. Using it to prepare documents or give guidance for other people may be practising law without a licence, which is regulated in most jurisdictions regardless of whether you charge for it. That line is worth checking before you are helpful to a friend.
Either way, settle the confidentiality question before the workflow rather than inside it. Client material, opposing-party documents and anything under a confidentiality undertaking are governed by your duties, your engagement terms, and the terms of the specific tool and tier you are using. Working from a redacted extract rather than the whole file removes most of the question.
Where it fails
| Failure | What happens | What to do about it |
|---|---|---|
| Fabricated authority | Plausible case names, reporters and holdings that do not exist, sometimes citing each other | Mechanical verification only. Every authority opened and read, or it does not go in |
| Misquoting your own file | A shall that was a may, a notice period from elsewhere, a carve-out that is not there | Locators on every line. Nothing quoted without a clause number |
| Silent jurisdiction drift | American doctrine and vocabulary applied to a question that is not American | State jurisdiction and date up front, and require its assumptions to be declared |
| The law has moved | It does not know what changed since training and will not say so unprompted | Treat every legal statement as of unknown date until confirmed |
| Absence is invisible | It reads what is there. Most contract risk is in what is missing | Ask explicitly what a document of this type normally contains that this one does not |
| Confident conclusions | Asked whether something is enforceable, it will answer | Never ask. Ask what it says and what to put to a lawyer |
| Archaic drafting | Hereinafter, witnesseth, and other decoration mistaken for precision | Ask for plain drafting and treat old-fashioned phrasing as a smell |
| No accountability | It carries no duty, no insurance and no licence. In the sanctioned case, the lawyers did | The qualified human is not a formality in the process. They are the process |
How do you install the AI Legal Assistant?
The download is a ZIP with SKILL.md at the root of the archive, not inside a nested folder - that folder structure is the usual reason an upload fails. In the Claude desktop app, open Customize → Skills, upload the ZIP and toggle it on.
Skills need code execution enabled, under Settings → Capabilities. Anthropic's help centre currently lists Skills on Free, Pro, Max, Team and Enterprise, while its Academy tutorial lists Pro, Max, Team and Enterprise - so if you are on the free plan, check Settings → Capabilities for your own account rather than taking either page's word for it. The full walkthrough is in the skill installation guide.
In ChatGPT or Gemini there is no upload step: open SKILL.md, copy the contents, and paste them into custom instructions. You lose automatic triggering and keep the method.
Who is this for?
Legal assistants and paralegals clearing document work, small firms without support staff, in-house teams, and founders who read contracts without a lawyer in the room and need to know what to ask. It works in Claude, ChatGPT or any AI chat.
The legal roles are split into their own skills rather than bundled:
- Vera - Legal Document Reviewer ($19) - review as its own discipline, clause by clause
- Leon - In-House Counsel ($19) - the commercial view, where legal meets the business decision
- Rex - NDA Specialist ($19) - the document everyone signs without reading
- Nina - GDPR Compliance Officer ($19) - data protection, which has its own vocabulary and its own deadlines
- Karolina - Employment Lawyer ($19) - the most jurisdiction-sensitive area of all
- Cass - Legal Workflow Specialist ($19) - matters, deadlines and the system around the documents
The wider set is the legal and counsel skills and the legal, HR & people ops skills.
In summary:
Treat plausibility as worthless, because a fabricated citation reads exactly like a real one and that is what got two lawyers sanctioned in 2023. Verify mechanically or not at all: every authority opened and read, none marked verified by the model itself. Extend the same suspicion to quotations from your own documents, which drift in ways that are invisible without a locator, and require a clause number on every line of every summary. State the jurisdiction and the date up front, and make the model declare its assumptions and its weak spots. Ask what a document says and what to put to a lawyer, never whether it is safe or enforceable. Then use the output as a brief for a qualified person rather than in place of one. For the jurisdiction, conventions and house rules held between sessions, the AI Legal Assistant ($19). Works in Claude, ChatGPT and any AI chat, with a 30-day money-back guarantee. None of it is legal advice.
AI Legal Assistant
Demand letters, motions, correspondence and day-to-day legal admin, including paralegal-tier filing preparation and court forms. Holds your jurisdiction and your house rules between sessions, so the checks get applied every time rather than when you remember. Drafts for qualified review, never in place of it.
KissMySkills is a marketplace of 2,000+ AI skills, 300+ prompt packs, 100+ agents & free tools for Claude, ChatGPT & any AI chat.
