The prompt

You are a product manager planning an Indian-language version of a product.

Product: [PRODUCT]
Language(s): [LANGUAGES]
Where the users are: [STATES]
Surfaces to translate: [SURFACES]

Write a PRD with:
1. Who this is for and what they cannot do today
2. Scope: text, voice, or both; what stays in English
3. How we will translate (people, machine, or both) and who checks quality
4. Text-to-speech and speech-to-text: when they are worth it and what they cost us
5. Design changes (font, text length, numerals, mixed Hindi-English input)
6. How we measure success, by state
7. Risks and what we do if a translation is wrong

Rules:
- Use the India figures listed below and name the source and month next to each one you use.
- If you need a figure that is not listed, write "Assumption:" and give your reasoning instead of a number you cannot source.
- Plain words. Short sections with headings. No filler.

India figures to use (from productgrowth.in/numbers):
- India had 958 million active internet users in 2025 (IAMAI-Kantar). Source: https://productgrowth.in/numbers/internet-users-india/
- Uttar Pradesh has the most internet subscriptions of any Indian state, 155.39 million at the end of June 2026, followed by Maharashtra with 130.11 million, out of 1,114.17 million in India (TRAI). Source: https://productgrowth.in/numbers/internet-subscribers-by-state-india/
- India's text-to-speech market was $108.57 million to $237.8 million in 2024 (MarketsandMarkets and Data Bridge). Source: https://productgrowth.in/numbers/text-to-speech-market-india/

Run this in Kriyā → Kriyā opens with this prompt in its box; nothing is sent until you press send. Replace the [BRACKETS] first.

What to fill in

Replace	With	Example (used for the sample answer)
`[PRODUCT]`	Your product	A crop-advice app for farmers
`[LANGUAGES]`	Which languages	Hindi and Marathi
`[STATES]`	Where your users are	Maharashtra, Madhya Pradesh and Uttar Pradesh
`[SURFACES]`	What to translate	Onboarding, the advice cards and push notifications

The numbers this prompt uses

The prompt carries these figures so the answer starts from dated, sourced numbers instead of guesses. Each page shows the source, the month and the history.

How many internet users are there in India? India had 958 million active internet users in 2025 (IAMAI-Kantar).
Which Indian states have the most internet users? Uttar Pradesh has the most internet subscriptions of any Indian state, 155.39 million at the end of June 2026, followed by Maharashtra with 130.11 million, out of 1,114.17 million in India (TRAI).
How big is the text-to-speech market in India? India's text-to-speech market was $108.57 million to $237.8 million in 2024 (MarketsandMarkets and Data Bridge).

Sample answer

Sample answer from Kriyā, 26 October 2026, for the example inputs above. Kept as it came back; links added beside each figure.

PRD: Hindi and Marathi crop-advice app

1. Who this is for, and what they cannot do today

Users: smallholder farmers in Maharashtra (Marathi), Madhya Pradesh and Uttar Pradesh (Hindi). Many are first-time smartphone users, on entry Android, on patchy data.
Today: advice arrives in English or in stiff, translated Hindi. Farmers cannot read dosage and timing reliably, so they trust the input dealer instead.
Scale: India had 958M active internet users in 2025 (IAMAI-Kantar) [3]. UP has the most subscriptions at 155.39M and Maharashtra has 130.11M, of 1,114.17M nationally, end of June 2026 (TRAI). MP's figure was not in the data given. Assumption: it is large enough to justify Hindi-first, since it shares the language with UP.

2. Scope

Launch: text plus voice playback (listen-only) for advice cards. Onboarding is text plus a voice prompt.
Later: voice input, after we see usage.
Stays in English: brand name, product codes, chemical names on labels (shown beside the local name), legal and licence text.
Hindi and Marathi separately: Marathi is not a Hindi copy. Both use Devanagari, but vocabulary, crop and pest names, and units differ by region.

3. How we translate and check quality

Both. Machine draft first, then human editing.
Editors: one agronomy-literate native editor per language, ideally from the target state. A second review by a field officer or farmer panel.
Glossary first: a locked list of crop, pest, pesticide, fertiliser and unit terms, in local usage, not textbook usage.
Gate: no advice card ships without human sign-off on dosage, timing and safety lines.
Assumption: regional dialect differences (Vidarbha vs Western Maharashtra, Bundelkhand vs Awadh) matter. Test with 10-15 farmers per region before launch.

4. Text-to-speech and speech-to-text

TTS (worth it at launch): many users read slowly or not at all. Listen-only advice is the highest-value voice feature and is cheap relative to input.
STT (defer): accents, noisy fields and mixed Hindi-English make accuracy risky. Add after we have real query logs.
Cost: India's TTS market was $108.57M to $237.8M in 2024 (MarketsandMarkets and Data Bridge). That range is market size, not our price. Assumption: we cost per character via a vendor API, and cache audio for repeat cards so each card is generated once. source
Offline: pre-download the day's audio on Wi-Fi to cut data cost.

5. Design changes

Font: Noto Sans Devanagari or similar. Minimum 16sp body, generous line height for matras.
Text length: Hindi and Marathi run longer than English. Design cards to flex, and keep one instruction per line.
Numerals: Devanagari numerals vs Western digits. Assumption: default to Western digits for dosage and prices (matches packaging and UPI screens), with a setting to switch.
Mixed input: accept Hinglish and Marathi typed in Roman script. Make search tolerant of spelling variants.
Units: show local units (acre, bigha, quintal) with a converter. Dosage in both local and label units.

6. Success measures, by state

Track separately for Maharashtra, MP, UP. Set baselines before launch, as no figures were provided.

Activation: onboarding completion in the local language.
Comprehension: share of users who correctly repeat a dosage after reading or listening, tested on a panel.
Engagement: advice cards opened, audio plays, and push opt-out rate.
Outcome: advice followed, and repeat use across a crop season.
Quality: translation error reports per 1,000 cards. Kriyā's estimate

7. Risks and what we do if a translation is wrong

Wrong dosage or pesticide advice is the severe risk, as it can damage crops or health. Mitigation: human sign-off gate, glossary lock, and numbers kept out of free machine translation.
If an error ships: pull the card within hours, push a correction to everyone who saw it, and log the root cause in the glossary.
Report path: "Is this wrong?" button on every card, routed to the editor.
Push fatigue: cap unsolicited pushes at about 3 per week [2].
Data and consent: collect location and voice data under the DPDP Act 2023 and Rules (notified Nov 2025), with core obligations phasing in through May 2027 [1]. Use clear consent in the user's language.
Dialect mismatch: mitigate with regional pilots before wider rollout.

Next action: run a 2-week pilot of 20 advice cards in Marathi (Maharashtra) and Hindi (UP), with human-edited text plus TTS audio, and measure comprehension on a 30-farmer panel.

Beside each figure: source links the number page it comes from, your input means it came from the example inputs, and Kriyā's estimate marks a figure Kriyā worked out or assumed with no source given; check those before you use them. Your own run will differ.

Prompt to write a PRD for a regional-language feature