Devanagari content production — Nepali for Nepali audiences

Translation from English versus writing natively in Nepali

The first decision shapes everything downstream — are you translating from English, or writing natively in Nepali?

The default workflow at most agencies is translation. The marketer drafts the campaign in English, gets it approved, then asks the model to translate it. The output is grammatical and the meaning is preserved. But translation carries the English sentence structure across the script boundary. You get long subordinate clauses where a Nepali writer would have used two short sentences. You get noun-heavy constructions where Nepali prefers verbs. You get the phrase “we are pleased to announce” rendered word-for-word into a register no Nepali brand actually uses.

The better workflow is to brief the campaign in English and ask for a Nepali draft tailored to Nepali conventions — not a translation, a draft. The prompt names the audience, the channel, the tone, and the key facts, and asks the model to write directly in Nepali in the register a Nepali brand of this size would use. The output is meaningfully different. Sentences are shorter. The honorific level is more deliberate. The idiom feels chosen rather than carried over. The draft still needs a native review, but the review is editing rather than rewriting.

The cost of this workflow is one extra prompt and a moment of clarity about what you actually want. The return is content that reads like a Nepali person made it.

The honorific register problem

Nepali has at least three honorific levels for you — तँ, तिमी, तपाईं — and a fourth deferential register, हजुर or यहाँ. Choosing the wrong level is not a small politeness mistake. In the wrong context it signals contempt, false intimacy, or a school-principal distance the audience did not consent to.

Default AI behaviour on the honorific question is to reach for तपाईं — formally safe, often awkwardly so. A direct-to-consumer brand on Instagram aimed at twenty-somethings sounds, in pure तपाईं, like the bank manager wrote it. A condolence message that drops into तिमी reads as cold. A youth-skewing fintech that uses यहाँ sounds like it is begging.

The honorific level is something the marketer briefs in explicitly, every single time. Tell the model who the audience is, what register fits this audience and this moment, and what level you have chosen. Tell it which moments — condolences, regulatory notices, founder letters — escalate to a higher register, and which moments — short captions, casual stories, friendly comments — stay at the chosen level. Then re-check the draft. The model will still drift on the honorific level inside a long piece; the marketer’s job is to catch the drift before it ships.

Idiom and tone — the place the model rings false

The clearest mark of an AI-translated Nepali post is the translated idiom. The English original said “let’s get the ball rolling” and the model produced a Nepali sentence that means “let the ball roll” — grammatically intact, semantically empty, idiomatically dead. Or the English said “we’ve got your back” and the model rendered it as something about a back. The phrase ports across; the meaning does not.

The fix is two things together. First, in the brief, explicitly ask for native Nepali idiom rather than translation. Tell the model that any idiom in the source can be replaced with an equivalent Nepali idiom or rewritten plainly — never literally translated. The model will do better than you expect when given permission to deviate.

Second, the human in the loop is non-negotiable here. Idiom is the layer of language where the model is reliably weakest — it has read enough Nepali to imitate the surface, not enough to feel which phrase rings and which does not. A native reader catches the dead idioms in seconds. Without that reader, the brand ships content that is technically correct and emotionally inert.

The same applies to tone. The model can hit a target tone if you describe it precisely — warm, slightly informal, hopeful but grounded, respectful of elders. It cannot infer the right tone from a vague brief, and it tends to overshoot toward either formal-corporate or casual-American. Naming the tone explicitly, and revising the draft against a tone reference the model has seen, closes most of the gap.

The mixed-script reality

A surprising number of marketers, when briefing AI on Nepali content, ask for pure Devanagari with no English mixed in. The output then sounds artificial, because the way Nepalis actually write in 2026 is mixed.

A typical Nepali Facebook post for an urban brand looks like this in practice: Devanagari for the body, English for the brand name and product names, English for technical terms the category has not translated (cashback, instalment, EMI in some contexts), Roman-script Nepali for some casual phrases, hashtags in both scripts. This is not corruption of the language. This is the actual register educated, urban Nepali audiences read and write in. A brand that refuses to use it sounds like a textbook.

Tell the model this is acceptable. In the prompt, name which terms stay in English — the brand name, specific product names, the legal-mandated English of the category — and which terms get the Nepali rendering. Let the model use the mixed register where the audience uses it; ask for pure Devanagari only where the moment requires it (formal notices, religious moments, copy aimed at older audiences or rural readers).

Outside the urban frame, the calibration changes. Madhesh audiences read a different mix — Devanagari with Hindi-language drift, less English, sometimes Bhojpuri or Maithili register. Older audiences read more formal Nepali with less code-mixing. Rural audiences in the hills sometimes read Devanagari but speak a regional language at home, and the copy lands better when it does not assume English fluency. The mixed-script default is for urban Kathmandu, Pokhara, and Biratnagar audiences. The marketer chooses the mix for each audience.

The personal terminology glossary

The habit that holds Nepali content production together across a year of campaigns is a personal terminology glossary — a short document, kept in a note or a shared doc, that lists how your brand renders each of its recurring terms in Nepali. The brand name in Devanagari, the product names, the category terms that have a Nepali rendering you have chosen, the honorific level you have settled on, the words you have decided to keep in English and the words you have decided to translate.

The glossary gets fed into every prompt. The model is told — these are the terms; use them exactly; do not paraphrase. The result is that the brand name is rendered the same way across every post for the year. The product names do not drift. The category terminology stays consistent. The audience reading three posts across the year sees one coherent brand voice rather than three competing renderings of the same words.

Check your understanding

Quick check

—

A Nepali e-commerce brand uses AI to produce all its Facebook copy in Devanagari. Which single workflow change raises the quality of the output the most over the next campaign?

1. AAsking the model to translate sentence by sentence rather than paragraph by paragraph.
2. BBriefing in English but asking for a Nepali draft tailored to Nepali conventions, plus a personal terminology glossary fed into every prompt and a Nepali-native reviewer on every draft.
3. CSwitching to a model with a Devanagari-aware tokeniser.
4. DBanning all English brand names and writing everything in pure Devanagari.

Check answer→

What comes next

Native-feeling Nepali content is one half of the audience question. The other half is who the content is for — because the brand on a single Facebook page is often serving two very different audiences at once. The next section is about the diaspora-versus-domestic split, why most Nepali brands underuse it, and how to make AI produce two variants from one brief without collapsing the difference.