Why Product Managers Who Can't Use Gemini Risk Obsolescence – A Step‑by‑Step PRD Generation Guide

When Gemini 3 Pro was announced the AI community buzzed, yet many product managers, especially outside AI‑centric firms, still lack a clear method to harness large models efficiently. The author argues that even Gemini 2.5 can reduce certain product‑manager tasks by more than 90% and that product managers who cannot embed such models into their workflow risk becoming obsolete.

1. Using FeatureList to Draft Requirements

The author demonstrates that backend requirements, which are highly structured, are ideal for large‑model generation. By describing the workflow and data structures succinctly, Gemini can flesh out detailed requirement documents, handling boundary scenarios and formal language. The process relies on a "FeatureList" – a hierarchical list of features used originally in automotive projects and later in internet products. The three focus points are module hierarchy, completeness of sub‑features, and priority assessment.

A fictional scenario – building a "League of Legends item‑lookup tool" – is used to illustrate the interaction. The author prompts Gemini with a concise description of the needed modules (hero management, equipment management, talent tree, item configuration, and skill‑point configuration) and requests a FeatureList in table form. Gemini returns an initial version, asks clarifying business questions, and after the author supplies answers, produces a refined FeatureList. Two common table‑generation pitfalls are highlighted: loss of formatting when copying to Excel and Gemini outputting tab‑delimited text; both are solved by exporting to Google Sheets or re‑prompting Gemini to convert the text to a proper table.

2. Generating Logical Diagrams with Mermaid

After the FeatureList is ready, the author uses Gemini to create visual representations of the data model and workflow. Gemini cannot directly output image files but can generate Mermaid code, which can be rendered in tools like Feishu (Lark) by inserting a "/mermaid" block. The author first obtains an ER diagram code snippet, renders it, and notes that while the diagram may need verification, it usually captures the intended relationships.

For workflow visualization, the author attempts a sequence (or swim‑lane) diagram. Initial attempts produce incorrect diagrams because Gemini misinterprets the term "swim‑lane". By providing a corrected Mermaid snippet found online, Gemini finally generates the desired diagram, demonstrating the importance of precise terminology.

3. Writing the PRD

With FeatureList and diagrams in place, the author proceeds to generate the PRD. The recommended approach is to describe each page separately, keeping prompts within Gemini's competence. Detailed prompts include attaching a template file and a sample PRD, then iteratively refining Gemini's output by pointing out missing interaction details (e.g., handling empty data, filter logic). The author shows how Gemini can produce HTML code for simple backend pages, which can be directly used or further edited.

Key practices include:

Breaking complex pages into smaller states.

Providing Gemini with both a PRD guide and example documents.

Iteratively correcting mistakes (e.g., missing interaction details, mixing technical documentation with PRD content).

Using the generated HTML as a maintainable prototype that can be updated by feeding diffs back to Gemini.

The author also shares a link to a separate "0‑foundation 5‑minute AI programming guide" for further reference.

4. Reflections on AI’s Impact on Product Management

The author reflects that while Gemini can automate many documentation tasks, the future may not require traditional PRDs at all. They speculate about end‑to‑end AI‑driven workflows where product managers converse with agents to obtain IDs that drive code generation downstream. The piece concludes with a broader market insight: "Market insight remains the most scarce and vital skill for product managers; technology serves insight, not the other way around." The author maintains a balanced view, acknowledging both the accelerating capabilities of large models and the uncertainty of when a qualitative shift will occur.