IllustrationWrite the PRD for an LLM-connected device as nine sections: product, model and unit cost, latency budget, data flow and retention, security baseline, labeling, BOM and power, offline behavior, and updates. Fill each from a dated public source or your own measurement, and mark every unmeasured number as an assumption.
A hardware PRD for a device that calls a language model has requirements a normal PRD lacks: a per-call cost that recurs for the life of the unit, a retention policy owned by a third party, and a latency path that crosses your radio, your network and someone else's servers. This page is a fill-in template. It contains no Duobay price, no measured result and no legal advice.
What sections should the PRD have?
Nine sections cover the decisions that are expensive to reverse after tooling. The third column says where the requirement comes from, so a reviewer can check it.
| Section | What to write down | Basis |
|---|---|---|
| 1. Product and user | Who wears or uses the device, in what place, doing what. What the device must do with no network at all. | Your product decision. |
| 2. Model and unit cost | Provider, exact model name, pricing tier, date the price was read, per-device monthly cost formula (below), and the date any scheduled price change takes effect. | Provider pricing pages (table below). |
| 3. Latency budget | Target time from end of user input to first useful output, split by stage: capture, uplink, model, downlink, playback or display. Each stage gets a target and an owner. | Vendor published no figure that applies to your device. Measure on your board (see the measurement table). |
| 4. Privacy and data flow | A diagram of every place audio, text and transcripts travel and rest. Retention period per hop. Which provider tier you ship on. Whether you need Zero Data Retention and who applies for it. | OpenAI and Google data pages (below). |
| 5. Security baseline | One requirement row per capability: device identification, device configuration, data protection, logical access to interfaces, software update, cybersecurity state awareness. | NIST IR 8259A. |
| 6. Compliance and labeling | Which FCC rules apply to each radio, where the compliance statement appears on the product or its manual, and who owns lab testing. | 47 CFR 15.19 (see the note on scope below). |
| 7. BOM and power | Target bill-of-materials cost at your volume, battery capacity, runtime promise, heat limit at the skin or enclosure surface. | Your quotes and your measurements. |
| 8. Offline and degraded modes | What the device does when the network is down, the provider is slow, the model is retired, or the account is out of quota. | Your product decision. |
| 9. Updates and ownership | How firmware and prompts are updated, who signs them, who holds keys and accounts, and what is handed over at each gate. | NIST IR 8259A Software Update capability. |
How do you write the model and unit-cost section?
List the model, the price and the date. Prices below are official list prices per million tokens, read on October 31, 2026. The last column is an arithmetic unit: the cost of 1,000 input tokens plus 1,000 output tokens, recomputed from the two published prices. It is not a usage estimate.
| Model | Input, per million tokens | Output, per million tokens | Cost of 1,000 in plus 1,000 out | Source |
|---|---|---|---|---|
| Claude Sonnet 5.5 | $2 | $10 | $0.012 | Anthropic pricing page, as of October 31, 2026 |
| Claude Haiku 4.5 | $1 | $5 | $0.006 | Anthropic pricing page, as of October 31, 2026 |
| Claude Haiku 5.5, prompts up to 100,000 tokens | $0.10 | $0.50 | $0.0006 | Anthropic pricing page, as of October 31, 2026 |
| Claude Haiku 5.5, prompts over 100,000 tokens | $0.50 | $2.50 | $0.003 | Anthropic pricing page, as of October 31, 2026 |
| Gemini 3.5 Flash-Lite | $0.30 | $2.50 | $0.0028 | Google pricing page, as of October 31, 2026 |
| Gemini 2.5 Flash-Lite | $0.10 | $0.40 | $0.0005 | Google pricing page, as of October 31, 2026 |
| Gemini 3.6 Flash | $0.75 through December 31, 2026; $1.50 from January 1, 2027 | Read it from the page before you use it | Not computed here | Google pricing page, as of October 31, 2026 |
Formula for the PRD:
monthly_cost_per_device = N × 30 × (T_in × p_in + T_out × p_out) / 1,000,000
N model calls per device per day (count)
T_in input tokens per call (tokens)
T_out output tokens per call (tokens)
p_in input price per million tokens (dollars)
p_out output price per million tokens (dollars)Worked arithmetic with Claude Sonnet 5.5 at $2 and $10: one call with T_in = 1,000 and T_out = 1,000 costs 1,000 × 2 / 1,000,000 + 1,000 × 10 / 1,000,000 = $0.002 + $0.010 = $0.012. Each additional call per day multiplies the monthly figure by 30, so one such call per day is $0.36 a month per device. The same call on Claude Haiku 5.5 at $0.10 and $0.50 is $0.0001 + $0.0005 = $0.0006, or $0.018 a month. The 1,000-token sizes are chosen for easy arithmetic and are not a typical prompt.
Your assumptions to replace: N, T_in and T_out. Log them on a prototype, including system prompt and resent history, and use the mean and the 95th percentile. Only p_in and p_out come from a source.
How do you write the privacy and data-flow section?
Draw every hop that audio, text or a transcript takes, and write the retention for each. Two provider facts shape this section.
- OpenAI: the data controls page says "abuse monitoring logs are generated for all API feature usage and retained for up to 30 days." Its default is that API data is not used for training. Zero Data Retention requires prior approval and covers listed endpoints, so check the current list for the endpoint you call (OpenAI data controls, as of October 31, 2026).
- Google: the pricing page marks paid-tier content "Content not used to improve our products" and free-tier content as used to improve them (Gemini API pricing, as of October 31, 2026). The PRD should state which tier production uses.
We did not read Anthropic retention terms for this page. If you use that provider, add the retention row from its own terms and date it.
How do you write the security section?
NIST IR 8259A defines six core cybersecurity capabilities for IoT devices: device identification, device configuration, data protection, logical access to interfaces, software update and cybersecurity state awareness. Make each a row with an owner and a test. Two have direct effects on an LLM device. Software Update means the device software "can be updated by authorized entities only." Data Protection includes "the ability to use demonstrably secure cryptographic modules for standardized cryptographic algorithms" (NIST IR 8259A, as of October 31, 2026).
What about labeling?
For devices under FCC Part 15, 47 CFR 15.19(a)(3) sets the compliance statement, which begins "This device complies with part 15 of the FCC Rules. Operation is subject to the following two conditions:" (eCFR, section 15.19, as of October 31, 2026). This page does not cover electronic labeling or any other provision of that section. Give the labeling requirement a line item in the PRD and have your lab confirm the exact text and placement. This is an engineering summary and not legal advice. Have a lawyer or accredited lab confirm.
What are the common traps?
These follow from the sources above. They are engineering reasoning about how the published terms behave. None describes an incident at any company.
- Free tier on a shipping device. The Gemini pricing page lists "Content used to improve our products" for the free tier and "Content not used to improve our products" for the paid tier. A prototype built on a free key and shipped unchanged sends user content under the free-tier terms. Put the tier in the PRD.
- Assuming no retention. The OpenAI data page says abuse monitoring logs are generated for all API feature usage and retained for up to 30 days. Zero Data Retention needs prior approval and applies to listed endpoints. A privacy line that says "nothing is stored" contradicts the page unless you hold that approval.
- A price with an end date. Gemini 3.6 Flash input is $0.75 through December 31, 2026 and $1.50 from January 1, 2027. A cost model built on the launch price understates year two.
- A long-prompt tier. Claude Haiku 5.5 is $0.10 input and $0.50 output for prompts up to 100,000 tokens, and $0.50 and $2.50 above that. A device that keeps appending history can cross the line without a product change.
- Skipping the update path. NIST IR 8259A lists Software Update as a core capability, in which software is updated by authorized entities only. A PRD with no update section leaves no place to decide signing, rollback or who holds the keys.
- Unencrypted hops. The Data Protection capability asks for demonstrably secure cryptographic modules for standardized algorithms. Audio between device and cloud is user data, so every hop in your data-flow diagram needs a named mechanism.
What should you measure on your own board?
We have not measured latency, power or heat for you, and no result appears here. The vendor pages do not publish figures that apply to your device. Run these on your own hardware and fill the PRD from your data.
| What to measure | How | Statistic and pass line |
|---|---|---|
| End-to-end latency, per stage | Add a timestamp at each stage boundary in firmware (input end, request sent, first byte back, first output). Log to flash or a debug UART. Run a fixed script of requests on the real hardware over your real network types. | Report median and 95th percentile per stage. Set the pass line from the longest wait your user will accept, and write it down before you run. |
| Behavior when the network drops | Cut the connection at each stage (before send, mid-stream, before reply) and record what the device does and how long it takes to recover. | Pass line is the offline behavior you wrote in section 8 of the PRD. |
| Calls and tokens per real interaction | Log prompts and replies from prototype sessions, including system prompt and resent history. Count tokens with the provider tokenizer. | Mean and 95th percentile replace T_in, T_out and N in the cost formula. |
| Power during a model interaction | Put a current meter on the battery rail. Record idle, one interaction and a repeated-use session, with the radio active. | Convert to battery hours for your usage profile and compare with the runtime in section 7. |
| Heat during repeated use | Run repeated interactions in the real housing. Log surface temperature at the contact point and the chip temperature. | Choose the surface limit from the safety standard for your category before testing. |
| Answer quality | Build a request set from prototype logs. Score outputs with a rubric written before scoring, on the easy and the hard cases. | Set the minimum score before you see any results. |
The vendor published no figure for these on your hardware, so the answer is to measure it on your board. Fix every pass line before you run the test.
What should you ask before you submit the PRD?
- Which exact model name and pricing tier does the firmware call, and on what date was the price read?
- Do we ship on a paid tier whose terms say content is not used to improve the provider products?
- What retention applies to our requests, and do we need an approval such as Zero Data Retention?
- Which hop in the data-flow diagram has no named encryption mechanism?
- Which latency stages have an owner and a target, and which are still blank?
- What does the device do with no network, and is that written as a testable requirement?
- Who signs updates, who holds the keys, and what is the rollback path?
- Which radio rules apply, and who is responsible for the compliance statement and lab testing?
- Which numbers in the PRD are measured on our own board and which are still guesses?
- What happens to unit cost on the date a published price changes?
Questions about the PRD template
Is this a full compliance checklist?
No. It is an engineering template for the product requirements document. It covers the sections a PM must fill in and points to the public sources behind them. Compliance scope depends on your radios, your market and your product category, and needs a qualified lab or counsel.
Why does the template leave latency numbers blank?
No vendor page we read publishes a figure that applies to your device, your network and your prompts. Any number we wrote would look like a fact. The measurement table shows how to produce your own.
Which provider tier should a shipping device use?
The Gemini pricing page states that paid-tier content is not used to improve Google products and free-tier content is. The OpenAI data page describes default retention and the approval needed for Zero Data Retention. Read both pages on the day you decide, and record the date in the PRD.
Does NIST IR 8259A apply to my product?
It is a public baseline from NIST for IoT device cybersecurity capabilities, and it is useful as a skeleton for the security section. Whether a customer, retailer or regulator requires it depends on your sales channel. Ask them.
Can Duobay fill in this template for me?
The Gate 0 Feasibility study turns open rows into decisions and measured numbers for your product. Tell us about your device and we will explain how it works. You can also see our templates and checklists.
If you want help turning the open rows into decisions for your device, tell us about your device. You can also read how the engineering gates run, and compare cloud and edge cost in the total-cost guide.
Sources
All pages accessed October 31, 2026. Prices and terms change, so re-read them before you decide.
- NIST, IR 8259A, IoT Device Cybersecurity Capability Core Baseline: https://nvlpubs.nist.gov/nistpubs/ir/2020/NIST.IR.8259A.pdf (accessed 2026-10-31)
- eCFR, 47 CFR 15.19: https://www.ecfr.gov/current/title-47/chapter-I/subchapter-A/part-15/subpart-A/section-15.19 (accessed 2026-10-31)
- OpenAI, data controls in the API platform: https://developers.openai.com/api/docs/guides/your-data (accessed 2026-10-31)
- Google, Gemini API pricing: https://ai.google.dev/gemini-api/docs/pricing (accessed 2026-10-31)
- Anthropic, Claude API pricing: https://platform.claude.com/docs/en/about-claude/pricing (accessed 2026-10-31)
Planning a device like this? Tell us what it needs to do. A Duobay engineer replies within 1 business day. Send a project brief →