My Product Management Toolkit (67): Using AI to write a PRD

Image Credit: ChatGPT

Writing Product Requirements Documents is one of the areas where ChatGPT or Claude can become your friend. You can save yourself a lot of precious time by using AI to turn rough product ideas into detailed PRDs within no time.

There’s a variety of short prompts that you can use to generate a PRD. I have used these prompts, but prefer an approach that requires a bit more upfront work to get a more detailed PRD that matches your preferred style. These are the steps I take:

1. Document my preferred PRD template with examples. See my PRD template here.
2. Create a project in Claude or ChatGPT. Having a project dedicated to your draft PRDs will make it easy to keep relevant chats and documents in a single place.
3. Set detailed system instructions. In “Project knowledge” (Claude) or “Instructions” (ChatGPT) I create detailed system instructions that you want the AI assistant to follow each time you ask it to turn your crude notes into a detailed PRD. I add my PRD template as an important reference for the assistant. To have the most complete system instructions I use the following format:

• Role
• Context
• Instructions
• Instructions
• Criteria
• Examples

Let’s look at these different system instruction elements and illustrate with my draft PRD system instructions in my Claude project (using simple XML tags to delineate the different sections of the system instructions).

Role — example:

<role>
Your task is to turn product ideas and a draft set of requirements into a detailed Product Requirements Doc (PRD).
</role>

I will at times use more traditional role prompting where I assign a persona or role to an LLM to guide its tone, style and behaviour. Prompts like “you are a world class requirements gatherer” or “you’re an expert copywriter specialising in product management content” are good examples of role prompting. Learning from research and experience, I am less convinced of the added value of role prompting. Instead, I tend to focus on the task that I want the AI assistant to help me with.

Context — example:

<context>
I am a product manager and I often need to write PRDs to provide clarity and direction to my team of full-stack engineers and a product designers. It is important that I create a PRD that provide a comprehensive overview of the required solution and is easy to follow.
</context>

In the “context” section I provide more background information to the AI assistant to take into account when drafting a PRD. Depending on the use case, you can make the “context” as detailed as you like to help the AI understand the situation it’s operating in (e.g. company information, ideal customer profile, market information, etc.).

Instructions — example:

<instructions>

Please follow these steps carefully in order to generate the PRD:

1. Analyse to understand my preferred PRD format and the level of detail. In tags, briefly summarise the key characteristics of my PRD format.
2. Ask me to share my notes about the product idea next.
3. Ask me clarifying questions to better understand the requirements. Ask me no more than 10 questions and present them in a numbered list.
4. Then structure the PRD as follows:

• Problem: Clearly describe in only a few sentences the problem, the customer affected by the problem and the impact of the problem on the outcome the customer is trying to achieve.
• Objective and key results: Include a business objective, the desired outcome for the customer and 2–3 leading indicators to show that this product or feature is having the desired impact on business and customer objectives.
• Solution requirements: Provide a clear outline of the solution and the different implementation milestones. For each milestone, write concise user stories (“I can …”, “I see …”) with nested bullet points about how the feature will work.
• Assumptions and hypothesis: Include between 5–10 customer and business assumptions (“I believe my customers have a need to …”, “We can solve this need by …” and “the number 1 benefit the customer will get …”).
• Product risks and dependencies: Identify at least 3 product risks for us to consider and include between 5–10 technical or cross-functional dependencies (e.g. other parts of the product journey, product marketing for rollout planning, etc.).
• User flows: Include the main user flows that the solution needs to cater for and ask me to include relevant design assets (e.g. Figma files).
• Delivery plan: Outline a high level delivery plan with the milestones, target delivery dates and a brief description of each milestone.

</instructions>

I learned from Dave Killeen to include a step that instructs the AI assistant to ask me clarifying questions. I’ve found this a very helpful step, an effective way to make sure that the AI is clear about the requirements and prevent it from hallucinating.

Criteria — example:

<criteria>
- The sentences in the PRD should concise and actionable.
- Information provided must be accurate and up to date.
- The PRD should be no longer than 2 pages max.
</criteria>

I use the “criteria” section the specific standards that I want the AI assistant to adhere to in its responses.

Examples — PRD for Digital Insurance Product — Claims Tracker Feature:

<example>

1. Problem to solve

Customers want visibility into the status of their insurance claims after submission. Today, once a claim is filed, users receive minimal updates and have no easy way to check progress, leading to frustration and repeated support calls (support tickets related to claim status account for 35% of incoming volume — see appendix). We want to test a claims tracker that provides real-time status updates and expected timelines directly within the app.

2. Objective and key results

Objective — Improve customer trust and satisfaction by providing greater transparency into the insurance claims process.
Customer Outcome — Customers can easily view the real-time status of their submitted claims without needing to contact support.
Key Results:
- Reduce support tickets related to claims status by 40% within the first 3 months post-launch.
- Achieve 70+% usage rate of the claims tracker among users who have an active claim.

3. Solution requirements

We want to test the claims in two milestones:

Milestone 1: Launch a basic claims tracker showing claim status and last update

Milestone 2: Add real-time status updates and estimated resolution timelines

M1: Launch a basic claims tracker showing claim status and last update [design link]. See claim status and last update:
As a policyholder who has submitted a claim, I can view my claim’s current status (e.g., “Under Review”, “Approved”, “Pending Documents”) in the Claims section of the app.
On the claim detail screen, I see the date of the last update, the name of the adjuster assigned (if applicable),and any outstanding actions I need to take (e.g., upload documents).

As a policyholder who has submitted a claim, I can view my claim’s current status (e.g., “Under Review”, “Approved”, “Pending Documents”) in the Claims section of the app.
On the claim detail screen, I see the date of the last update, complete with a timestamp, and the estimated resolution timeline.

[and you add examples for the remaining PRD sections]

</example>

Like with all effective prompts you want to include several examples to show the AI assistant what good responses look like (“few shot learning”). Breaking the PRD down into specific milestones is very helpful when you use the PRD to generate a prototype in a tool like V0 or Lovable. Instead of the AI building the entire prototype in one go, you can tell it to apply a phased approach. This will help you review the quality of the AI’s output and iterate.

I took an idea for a healthy recipe assistant and used my project instructions in Claude to generate a PRD:

Image Credit: screenshot

Image Credit: screenshot

The AI assistant is following my instructions and is asking me to share more notes about my product idea even before we get to the clarifying questions that I’ve instructed it to ask me!

Image Credit: screenshot

You can then use voice or words to respond to the AI assistant, providing the additional information and answers it’s looking for.

Image Credit: screenshot

The AI assistant then generates a PRD that I can feed into tools like Bolt, Replit and Lovable to generate a first prototype. By publishing the PRD, I can easily share it with my team or stakeholders (see my Healthy Recipe Coach PRD here). The PRD generated has the structure and sections that I’ve outlined in the system instructions. Naturally, I need to review and iterate on the information in there, but I’ve got good starting point to work from.

Image Credit: screenshot

My main learning point: If you want AI to help write effective prompts, you’ll need to invest upfront effort to make sure the AI returns the structure and level of detail that you’re looking for. Doing this will remove the cold start problem we as PMs often face when writing PRDs.

Related links for further learning:

1. Is Role Prompting Effective? by Sander Schulhoff
2. The PRD That Actually Works by Aakash Gupta
3. Please don’t damage your career and credibility with poor PRDs by Alena Panshina
4. I Wrote a PRD with AI and It Worked Surprisingly Well by Peter Yang
5. How I use ChatGPT to write PRDs as a Product Manager? by Sushant Kumar