Stop writing PRDs in one go with WorkBuddy: split it into alignment, generation, and review

I think product managers get scolded by developers the most often because their requirements are too vague.

In review meetings, I'd just blurt out things like "this requirement is simple." Then devs would ask: where's the entry point? How do we handle exceptions? Who grants permissions? Where does the data come from? I'd freeze on the spot. Recently, I've been reflecting on my tendency to talk too much, and I realized the key issue was that my speech lacked structure.

I came across a post about prd-writer, which said that when using AI to write PRDs, you shouldn't try to do it all in one go. It breaks the process down into aligning on direction first, then writing details, including reviews and incremental updates.

The core approach of prd-writer is to align on direction first, then write implementation details, with reviews and incremental updates included.

I think this mindset is pretty right. But I didn't directly use that Skill; I'm more familiar with WorkBuddy. WorkBuddy is an AI office efficiency tool I use for handling documents, spreadsheets, meeting notes, and other administrative chores. I've used WorkBuddy's automated weekly reports for about a month, and recently tried AI-powered PRD generation tools. In the end, I found that the truly stable approach is to have WorkBuddy break down requirements into a runnable input table first.

Step 1: Don't rush to write the PRD; establish an input contract first

PRD stands for Product Requirements Document. An input contract is the boundary I set for it: mandatory fields and required sources must be clearly defined, and unconfirmed content cannot be fabricated.

My current method is to open WorkBuddy, click "New Document Task" on the left (I call it this; names might vary by version). At the top, select "Import from File," and dump meeting notes from WeCom Docs, Tencent Meeting transcripts, and verbal requirement notes together. I've been using WeCom Docs for a week and Tencent Meeting transcripts for just a week; the benefit is that the original text is preserved, so WorkBuddy doesn't have to rely on me filling in gaps mentally.

After importing, paste fixed fields into the task rules: target users, usage scenarios, entry points, core flows, exception branches, permission roles, data fields, acceptance criteria, and pending questions. Structured fields mean breaking requirements down into discrete cells, preventing the AI from freestyling into long essays.

After clicking run, WorkBuddy generates a field extraction table on the right. Each field has a source behind it, e.g., line 12 of the meeting notes, paragraph 3 of the WeCom Doc. If no source is found, it marks it as "pending confirmation." This step is crucial. Previously, when I let AI write PRDs directly, it could turn a fresh grocery app into a "fresh food empire," with features getting grander and grander, but when R&D asked how to implement them, it fell apart.

Here's a pitfall. Initially, I had it summarize prototype descriptions from Axure RP and Modao together. As a result, it treated page names as requirements and outputted a bunch of button labels. Later, I learned my lesson. During my first few days using Axure RP, I treat prototypes only as attachments, not as sources of truth. Requirements are based on meeting conclusions.

Step 2: Generate in three stages; don't do it all at once

Three-stage generation is the configuration I recommend most now.

Create three subtasks in WorkBuddy, or split one task into three rounds. Round 1 generates only a one-page directional summary, including user problems, business goals, and what not to do. Round 2 generates flows and fields, focusing only on main processes and exceptions. Round 3 generates acceptance criteria and pending items for R&D and QA to see.

The specific path is: after generating the input table, select the table, click "Generate Document," choose the "Summary" template, and write in the prompt: "Do not expand; output only the following fields; ask questions if information is missing." After generation, copy the summary, return to WorkBuddy, use the summary as context, and have it generate a text-based flowchart and field table. Finally, have it output the initial PRD draft.

This results in a requirement version that can withstand review. I think the biggest problem with AI-written PRDs is hallucination. By breaking the problem down, each round has inputs, reducing hallucinations.

Don't mess up permissions either. I set WeCom Docs access for three groups: PMs can edit the input table, R&D and QA have read-only access, and the boss can only comment.

The initial PRD draft generated by WorkBuddy isn't sent directly to the group chat. I manually paste it into the personal draft area of WeCom Docs, and sync it only after passing review. Small teams don't need heavy systems; lightweight approval and commenting are enough.

Step 3: Daily maintenance; don't let the PRD become a new pitfall

WorkBuddy handles simple text well, but maintaining complex table formats is indeed unstable. Last week, when I used it to build a metric ledger, I encountered issues where extra fields caused formatting to disappear. So, don't generate overly complex tables in the PRD all at once.

I do three things every week: After each review, feed new conclusions back into WorkBuddy and run an incremental update so old docs don't keep misleading people. Every Friday, check pending questions; if there are more than 3, mark them red, indicating requirements aren't truly aligned yet. Before exporting to R&D, manually verify exception branches and permission fields—these two areas are where WorkBuddy is most likely to miss things.

Listen, if you're planning to use AI to write PRDs next week, don't just throw in "Help me write an App requirement doc." First, use WorkBuddy to organize meeting notes, WeCom Docs, and Tencent Meeting transcripts into an input table. Whether a requirement is simple or complex won't be known until the fields are filled.