A friend of an advocate we work with spent about ₹2,000 a month on a general AI assistant and drafted roughly three times as many files as he managed with us for the same money. He was not exaggerating and we are not going to pretend he was.

So this is not a piece telling you to stop. Plenty of lawyers have moved some of their work to a general assistant this year and many of them are better off for it.

It is a piece about the fact that a lot of them are using it at a fraction of what it can do, and a few are using it in a way that will eventually cost them money or a complaint. Both problems have the same cause, and it is not the model.

Why "I represent the petitioner" gets you nothing

Type those four words into a chat window with a scanned record attached and you will get a polite question back. Type them into a legal tool and you get an analysis.

The difference is not intelligence. It is that the tool was told, before you ever arrived, who it is meant to be, what jurisdiction it is working in, what a legal analysis is supposed to contain, in what order, in what register, and what it is forbidden to invent. A general assistant knows none of that about your work until you tell it.

You can tell it once. Set up a project and put standing instructions in it, and every chat inside that project starts from them. Projects are on every plan, including the free one. Do that. It is the single biggest improvement available to you and it takes ten minutes. What remains yours is the brief itself: writing it, and revising it when a draft comes back wrong. Nobody else is going to maintain it.

The brief you are not giving it

Here is a starting point. Put it in your project's instructions, or keep it in a note if you would rather paste it. It is dull and it is the entire difference between a usable output and a plausible one.

You are a senior advocate with twenty five years of practice before
Indian courts.

Task: <one line. "Draft a reply to the notice attached.">

Work only from the attached documents. Do not supply facts, names,
dates or figures that are not in them. Where the record is silent,
say so rather than filling the gap.

Do not cite any judgment that does not appear in the attached
material. If you think authority exists on the point, describe the
proposition and tell me to find the case rather than naming one.

Output: <the shape you want. Numbered paragraphs, cause title, facts,
grounds, prayer, verification.> Plain text, no tick marks.

Run it on the strongest model your plan offers. The smaller ones are cheaper for a reason, and a pleading is not where to save.

Two things sit outside the brief.

The first is that the controls move. If a course taught you to turn the improvisation settings down for legal drafting, those settings are gone on current models, replaced with a different control. The provider will keep making decisions like that. What survives every such change is the instruction in words, which is why the stay-inside-the-record lines above matter more than any setting: prefer saying nothing to filling a gap.

The second is the document, and it is worth being precise about which part of it is a one-time cost and which part is not.

Whatever comes back is prose in a chat window. Keep a blank template of your own in Word with the paper size, margins and font already set, and paste into that rather than starting from a fresh document. That is a one-time job and it takes the page setup off your list for good.

What the template does not fix is the paragraph numbering, the annexure references and the cause title, because those belong to this draft and not the last one. Never let the assistant number the paragraphs for you. Typed-in numbers break the first time you insert a ground, and every reference to "Annexure P-4" now points at the wrong exhibit. Use Word's own numbering, applied after you paste, and check every annexure reference against the actual list. That is the per-draft cost. In chambers with juniors it is theirs; if you are the whole chambers, it is yours.

The two habits that matter more than the prompt

Everything above affects quality. These two affect whether you are exposed.

Know what your settings actually say

Anthropic's published policy for its consumer plans is better than most people assume. Your chats are used to improve Claude only if you switch on Model Improvement in Privacy Settings, and Incognito chats are excluded even if you have switched it on. So the default is not the problem.

The problem is not knowing which way your own toggle is set. Go and look. If you have client material in there, look today. And for anything genuinely sensitive, use an Incognito chat so the question does not arise.

Memory is a separate setting, and it is better designed than people assume: each project keeps its own memory, so what the assistant picks up while planning a holiday does not follow you into a client's matter. What is worth checking is whether it is on at all, because the default is not the same everywhere. On the individual plans it is on; on the team and enterprise plans it is off until an owner turns it on. If you moved from a personal account to a firm one, or the other way, the setting you think you have may not be the setting you have.

Different products answer all of this differently, and some consumer AI tiers do train on conversations unless you opt out. The habit worth building is checking, on whatever you use, rather than assuming.

Never cite what you have not opened

This is the one where the cost is not an afternoon.

LiveLaw's reporting on phantom precedents sets out where this has already gone in India. The Supreme Court issued notices to the Attorney General and the Bar Council of India after a trial court relied on AI generated rulings that did not exist. Justice Surya Kant criticised the growing practice of filing AI drafted petitions without verification. Justice B. V. Nagarathna encountered a fabricated case called Mercy v. Mankind during PIL proceedings.

In January 2026 the Bombay High Court imposed ₹50,000 in costs on a party that filed fake citations, noting the submission carried the obvious giveaways of raw AI output, green tick marks and repetitive formatting included. In September 2025 a Delhi High Court petition was withdrawn after opposing counsel found invented paragraphs 73 and 74 attributed to Raj Narain v. Indira Nehru Gandhi, a judgment that runs to 27 paragraphs. None of these incidents is attributed to Claude. They are attributed to lawyers who did not open the judgment.

The mechanism is worth understanding, because it explains why this will not simply be fixed. A language model on its own does not look anything up. It produces the most probable next piece of text, and a citation that does not exist can be exactly as probable looking as one that does. Confidence is not evidence of retrieval.

So far the consequences have been costs orders and withdrawn petitions. Those are the mild end. Bar Council rules on an advocate's duty to the court sit behind them, and Section 35 of the Advocates Act, 1961 treats professional misconduct as a matter for suspension and removal. Nobody has been taken that far over a fabricated citation yet, and nobody should want to be the test case.

So the rule is simple and it applies to every AI tool including ours. Open the judgment. Read the paragraph. If you cannot find it, it is not there.

Where the window runs out

A practical limit people hit without planning for it. A consumer plan caps you by time: a rolling five-hour session window, with weekly limits on top for paid plans. Attachments count against it. When you reach it the reset comes on its own and costs nothing. Whether there is a way to keep working sooner depends on your plan and your settings, so find out which options you have before the evening you need them.

That is a reasonable design and most days you will never touch it. The day you will is the day before a filing, when the record is long and the draft is late. Plan for it: start the heavy attachments early, not at nine in the evening, and keep the last window for corrections rather than first drafts.

And before you assume the model ignored something in a long record, ask it plainly what it received and what it is missing. It will say so if asked. It will not always say so unasked.

What no chat window is going to do

None of the above is a criticism of the model, and this section is not either. It is just a boundary.

A chat window cannot tell you what is listed tomorrow. It cannot read a cause list into a diary, check a matter against eCourts overnight, or send you a morning email with what is listed and before which bench. It cannot keep a ledger of your bills with what is paid and what is not. These are not prompting problems. They need a standing connection to court infrastructure and somewhere to keep a record between conversations, and a conversation has neither.

If drafting is genuinely all you need, that costs you nothing. If you are running a practice, it is most of the job.

The honest close

If you are willing to set up a project with a proper brief and keep it current, keep your own Word template and do the numbering in Word, check your privacy and memory settings, and open every judgment before you cite it, a general assistant will do good work for you. Genuinely. We would rather you did that than use it badly.

We built Legal Desk AI because we did not think that should be the lawyer's job. The brief is already inside each tool and we revise it against real drafts. When you turn retrieval on, citations come from your own library or from actual judgments, with the source shown to you. The draft downloads as a Word document with live numbering and annexure references, laid out to sensible Indian filing defaults, or matched to the paper, margins, font and spacing of a .docx of your own. When the allowance runs out you top up and carry on. We do not use your prompts, responses or documents to train AI models, and there is no setting that changes that.

That is a real difference and it is also not magic. Verify our output too.