AI Guides › Step-by-step guides
By Nigel Guy · 8 min read
The usual mistake is to treat a benchmark headline as a licence. You see that a Claude model scores around 90% on a legal test, paste in a contract, and ask "is this OK to sign?" You get a confident, tidy answer that feels like advice. It is a fluent reading of a document by a tool that does not know your circumstances, your jurisdiction unless you say, or what the other side will actually do.
The rule: use Claude to read, structure and draft, never to decide. Every output ends as a list of questions and a draft that you check against a primary source or a professional.
The figure that circulates is "90.9% on BigLaw Bench". I could not find that exact number on a primary source. What I could verify: Harvey, the legal AI company that built the benchmark, published Claude results of 91.1% for Opus 4.8 on its own site, and search results attribute 91.3% to Claude Sonnet 5 and a 90.2% baseline to earlier Claude models. Treat the exact number as unconfirmed and moving.
More useful is what the test measures. Harvey's own description says BigLaw Bench is built from real attorney time entries across transactional and litigation tasks, scored against rubrics written by Harvey's lawyers, with penalties for hallucination. The authors state it emphasises tasks models can or should be able to do and falls "far short" of everything lawyers do. It is a vendor-run test of lawyer-style work products, not a measure of whether the model is right about your tenancy, your supplier or your UK-specific rights. A high score means the tool is good at the shape of legal work. It does not mean its answer on your document is safe to rely on.
Each job below has a fixed output, so you can see what was invented.
Start a new chat. Paste or attach the contract, then use this prompt. Fill in the bracketed parts.
You are a careful contract reviewer helping a non-lawyer in the UK
understand a document before they sign. You are not their solicitor and must say
so in your first line.
Context: I am the [ROLE: e.g. freelance supplier / tenant / customer]. The other
party is [OTHER_PARTY]. I am in [JURISDICTION: England and Wales / Scotland /
Northern Ireland]. What I care about most: [PRIORITIES: e.g. getting paid on time,
ending the contract, liability cap].
Task, in order:
1. Summarise what the contract commits me to in plain English, in under 150 words.
2. List every clause on payment, term and renewal, termination, liability and
indemnity, intellectual property, and anything automatic (auto-renewal,
penalties, price rises). Quote the clause number and the exact wording.
3. Flag the clauses that are most one-sided against my role, and say why.
4. Note what appears to be missing that I would normally expect.
5. Give me a numbered list of questions to put to the other party or a solicitor.
Rules: do not say whether I should sign. Do not invent clause numbers or law. If a
point depends on a statute or case, name it as something for me to verify and say
you are unsure. If any input above is blank or the text looks incomplete, ask me
before answering.
Before replying, check that every quoted clause appears in the text I gave you.
Output: a plain summary, a clause table and a question list. Take the question list to the other party.
This job suits an unpaid business invoice. The UK late-payment rules apply to business-to-business supply of goods and services. Gov.uk says you can claim statutory interest of 8% plus the Bank of England base rate unless your contract sets another rate, and that you should send a new invoice if you add interest. Section 5A of the Late Payment of Commercial Debts (Interest) Act 1998 adds fixed compensation of £40 (debts under £1,000), £70 (£1,000 to £9,999) or £100 (£10,000 or more). Check both on gov.uk and legislation.gov.uk before you quote them, because rates and rules move.
You are a credit controller drafting a firm but courteous payment chase letter on
behalf of a UK small business.
Facts (use only these, do not add any):
- My business: [YOUR_BUSINESS]. Customer: [CUSTOMER].
- Invoice number [INVOICE_NO], dated [INVOICE_DATE], for £[AMOUNT], due
[DUE_DATE]. Payment terms in the contract: [TERMS].
- Contact so far: [DATES_AND_SUMMARY_OF_CHASES].
- Contract states a late-payment interest rate: [YES_WITH_RATE / NO].
Write: a letter of no more than 300 words that states the debt, the due date, the
exact amount, a clear payment deadline of [DEADLINE_DATE], and what I will do if it
is not paid (for example, claim statutory interest and compensation under the Late
Payment of Commercial Debts (Interest) Act 1998, and consider a court claim). Tone:
calm, specific, no threats you cannot carry out.
Also give me a short list of figures and rules in the letter that I must verify on
gov.uk before sending. If any fact above is missing, ask me first.
Self-check: confirm that no date, amount or legal claim appears that I did not give
you or that you have not marked as "verify".
Output: a letter and a verification list. Do the verification list, then send it by a method that gives you proof of delivery.
This is the safest of the three, because you are organising your own facts rather than asking for law.
You are a neutral note-taker helping me write a dated record of events for my own
file. I will give you rough notes in any order.
Topic: [SITUATION_ONE_LINE]. People involved: [NAMES_AND_ROLES].
My raw notes: [PASTE_NOTES].
Produce: (1) a chronological timeline table with date, what happened, who was
present, and the evidence I hold (message, email, photo, document); (2) a separate
list headed "My interpretation or feelings" so opinion never mixes with fact;
(3) a list of gaps, meaning dates or details you could not place; (4) a list of
documents I should gather.
Rules: add nothing I did not say. Do not guess dates. Keep my wording where I give a
quote. Ask me about gaps rather than filling them.
Save the output with the original notes. Notes written soon after the event, and kept alongside the originals, are far more useful to anyone helping you later than a polished account written months on.