Chapter 04 · Section III · 16 min read
Letters and structured correspondence
A small Nepali practice writes dozens of letters a week — demand notices, writ-warning letters, settlement covers, government correspondence — and the single highest-leverage instruction you can give the model is to put the action on its own line, in bold, at the bottom.
The volume of letters that comes out of a small Nepali practice is genuinely surprising the first time you try to count it. Demand letters on overdue invoices and contractual payments. Notice letters before filing a writ. Cover letters with settlement offers. Correspondence with banks asking for statements, with municipalities asking for certified copies, with opposing counsel proposing dates, with government departments seeking clarifications. In a busy week a single advocate may put their signature on twenty such letters. Each one is short, formulaic, and consequential — and each one used to take fifteen minutes to write properly. AI changes that arithmetic, but only if you use it the way the work actually demands. This section is about the discipline that turns the model into a reliable letter-drafter rather than a fluent generator of risk.
Fact-first prompting, always
The same rule that governs client emails and intake memos governs letters: the model never supplies the facts, and you never let it. Before asking for a single sentence of letter prose, you assemble the factual skeleton — matter number from the firm’s filing system, party names and roles, the date the relevant event happened, the deadline the letter is enforcing or proposing, the specific action required from the recipient, and the tone the situation calls for. Only then do you ask the model to draft.
The cost of breaking this rule, in correspondence, is not a slightly inaccurate paragraph. It is a letter on your firm’s letterhead, signed by you, that asserts a date that is wrong, a sum that is wrong, or a deadline that does not match the underlying contract. Opposing counsel reads it, notices the error, and now your client’s position on the merits is harder to maintain because your letter overstated or understated the facts. The model did not put your firm’s name on the letter; you did. The factual discipline at the prompt is the protection.
A working pattern for any client letter looks like this. Use the same template across every matter; only the contents change:
Draft a short letter on behalf of the client. Use only the facts below. Do not invent any date, amount, contract reference, statute, or party name not provided. Do not add new legal claims beyond what the facts support.
Matter reference: [matter no.] Sender: [the client, described by role and business type only] Recipient: [the counterparty, described by role only] Background facts: [the dated, neutral chronology of what happened] Action required: [the single thing the recipient must do, by when] Tone: [warm / measured / firm / formal — pick one and add one sentence of context] End the letter with the required action on its own line, in bold.
The model wraps your facts in serviceable letter prose. You spend two or three minutes checking that no figure or date drifted in transit. The signature is yours; the facts were always yours.
Three real examples, three different shapes
The fact-first structure adapts to almost every letter type a Nepali practice produces. Three common shapes show the variation.
One: the demand letter on an overdue contractual payment. A client supplied goods or services under a written contract, the agreed payment date passed six months ago, three reminders went out from the client’s own accounts team, and now the matter has come to you. The facts block names the contract by date and parties, the amount due, the original due date, the dates of the prior reminders, the applicable interest provision in the contract if any, and the deadline you are setting for payment before formal legal action. The tone is measured and firm, not aggressive — the relationship may still be salvageable, and a hostile letter rarely produces faster payment than a clear one. The action paragraph at the bottom is one line: “Please remit Rs. X, together with contractual interest of Rs. Y, to the account specified below by [date].” The model produces this in seconds. You verify the figures against the actual ledger before sending.
Two: the notice of intention to file a writ petition. A government decision or an action by a public authority has affected the client adversely, and you intend to seek constitutional relief if it is not reconsidered within a reasonable time. The facts block names the decision, the date it was communicated, the specific legal infirmity you allege, the relief that would render a petition unnecessary, and the deadline beyond which you will file. The tone is formal and precise — this letter may be quoted back at you in court, and any flourish you add will be your own writing under cross-examination. The model produces a clean draft. You verify that every statute reference and every authority mentioned is real and correctly cited, applying the same verification discipline the chapter on legal research insisted on.
Three: the settlement-offer cover letter. A negotiation has reached the point where the client is prepared to make or accept a concrete offer, and the letter is the formal record of that offer. The facts block names the underlying matter, the offer itself in clear figures and terms, the conditions attached, the period for which the offer remains open, and the consequence of non-acceptance (the matter proceeds, the client withdraws the offer, costs are sought). The tone is measured and constructive — the offer is meant to settle, not provoke. The action paragraph is one line: “Please confirm acceptance of the offer set out above by [date], failing which the offer will be withdrawn and the matter will proceed.” The model handles this register precisely when told.
In each case the input takes a minute to assemble and the output takes two minutes to verify. The full letter cycle drops from fifteen minutes to three, and the consistency across letters from your firm improves noticeably.
Tone register and the audience problem
Nepali correspondents are not one audience. A family-run business client wants warmth — a sentence acknowledging the relationship, a polite opening, a willingness to find a way through. A government department expects formality — measured paragraphs, specific section references, a request rather than a demand where possible, and titles used correctly. Opposing counsel expects measured firmness — neither friendly nor hostile, professional in register, with the legal position stated cleanly and without rhetorical flourishes. A bank’s compliance team expects bureaucratic precision — your letter will be read by someone whose job depends on being able to file it cleanly. The same factual content, four different envelopes.
The model handles each of these the moment you tell it. The instruction is one sentence at the bottom of the facts block — “Tone: warm and acknowledging the long relationship, this is a fifteen-year family-business client” or “Tone: formal, government correspondence to a department head, use the official title throughout” or “Tone: measured and firm, this is opposing counsel in a contested matter and the letter may be quoted in court.” The output adjusts completely. What you must not do is leave the tone unspecified and hope. An unspecified tone defaults to a kind of generic professional English that lands wrong on every Nepali correspondent — too cold for the family business, too informal for the government, too soft for opposing counsel.
The action paragraph rule
Every letter you send ends with one ask. Not two. Not “please remit the amount and also send the bank reconciliation and we should probably catch up about the next contract.” One. If you let it, the model will end every letter with three or four parallel asks and the recipient will do none of them — or, worse, will do the most convenient one and ignore the rest.
The instruction is a single sentence in the prompt: “End the letter with the required action on its own line, in bold.” The recipient who scrolls past the body and reads only the bottom — which is most recipients, most of the time, especially on a phone — sees exactly what they must do and by when. The follow-up rate visibly improves. The number of clarifying replies drops. The legal record, if the matter ever reaches a court or an arbitrator, contains a letter whose ask is unambiguous on the face of it.
If you genuinely need two things from the recipient, send two letters, or structure one letter with two clearly labelled sections and one bold action line for each. The cost of one extra letter is much smaller than the cost of one ignored letter containing two requests, and the cost of an ambiguous record in litigation is larger still.
Confidentiality, the same line every time
This rule has not changed across the chapter and will not change. Do not paste your client’s full name, citizenship number, PAN, bank account number, the opposing party’s full identifiers, or any other identifier into a public chatbot. Free tiers may use the input for training. Paid tiers route data through servers outside Nepal that the firm cannot audit. The Nepal Bar Council confidentiality obligation does not pause because the tool produces faster letters.
Redact at the prompt. “Sender: the client, a mid-sized cement dealer in Birgunj” rather than the client’s full name. “The agreed payment of Rs. 18,40,000” rather than the contract scan. “Recipient: the counterparty, a sole proprietor in the same ward” rather than the opposing party’s name and citizenship details. The model needs the shape of the matter to draft a serviceable letter. It does not need the identifiers that let a stranger reconstruct who the parties are.
For firms that handle volume — corporate, banking, family-law matters at scale — the right answer remains an enterprise account with a no-training data agreement, or a locally-hosted model. For everyone else, redaction discipline at the prompt is the protection. The letter the model drafts is the same; the risk profile is not.
What never goes in a letter, regardless of how the model phrased it
A final discipline. Some things belong in a letter only after you have personally verified them, regardless of how confidently the model produced them. A statute section the model cited. A case authority the model referenced. A figure the model derived rather than received. A factual assertion about what the opposing party did or did not do. Each of these, if the model produced it without your having verified it, is a hypothesis dressed as a statement. A statement of that kind, on your firm’s letterhead, is your assertion in law. The model carries none of the consequence; you carry all of it. The few minutes of verification are not optional, and they are the difference between AI as a letter-drafting tool and AI as a quiet source of professional embarrassment.
Check your understanding
Quick check
—An advocate is using AI to draft a demand letter on an overdue contractual payment. Which single instruction in the prompt is the highest-leverage change they can make to the letter's effectiveness?
What comes next
Letters are the steady, daily workload of a small practice. The chapter that follows turns to the higher-stakes drafting — the pleadings, briefs, and submissions that carry your client’s case in front of a court. The discipline does not change, but the verification bar rises, and the consequences of a careless AI-assisted draft rise with it.