Using WorkBuddy to Turn One Business Request into 5 Approved JDs

Yesterday, the business side threw a line in the group chat: "Hire an AI Product Manager who understands large models." I replied "Okay," but internally thought, dear, if this sentence could be written into a JD, then "responsible for" in resumes could count as evidence. I scrolled past an article saying HR can use AI to generate 5 recruitment versions of a JD in 1 minute. My first reaction was speed; my second was five ways to do it wrong.

I've used WorkBuddy for about a month. It's fairly stable for breaking down documents and organizing tables. Last week, I tried single-agent mode, where one AI does everything from start to finish. After running it for a week, I found it easily wrote JDs that looked like resume templates—complete job titles, but full of clichés. Later, I turned WorkBuddy into a JD production line. A JD is a Job Description; it needs to show candidates the role and force the business side to clarify their demands.

Open WorkBuddy, go to "Document Space" on the left, click "New Project," and name it "JD Production Line." Drag a blank "Business Requirements Collection Form" into it. Don't throw raw requirements in directly; first select Expert Team mode. An Expert Team consists of multiple AI roles dividing labor by steps, which is more stable than single-agent one-shot generation. I fixed four roles: HRBP, Business Interviewer, Compliance Reviewer, and Channel Copywriter.

Paste the original business request into the input box, then add a hard instruction.

Do not write the JD yet. Please list 5 questions based on the principles of facts, verifiability, and the business side being able to answer in 1 minute. Add an example answer after each question. Do not use empty words like "excellent," "responsible," or "stress-resistant."

The expected result is a "Requirements Gap List." Last week, I used it to ask about an AI Product Manager role. It first asked: What type of large model products do you expect the candidate to have worked on? Calling APIs to build workflows, or training models themselves? This question is a bit sharp, but sharply correct. Previously, I thought instructions should be gentle, but I later found gentleness useless; the business side would just reply, "You decide."

I stepped on a pitfall here. Initially, I wrote "Ask me nicely," resulting in it asking, "What excellent qualities do you hope the candidate possesses?" This is no different from a candidate writing "Strong communication skills"—no evidence. Later, I changed it to "Only ask for verifiable facts," and it started resembling interview assessment.

Generating 5 versions can't rely on inspiration; it relies on field tables. My environment uses WorkBuddy's table module, which may not apply to everyone. Open "Tables," click "New Table," and name it "JD Field Table." Columns include Position, Department, Reporting To, Location, Hard Skills, Soft Skill Scenarios, Q1 Results, Banned Words, and Channel Version. Hard skills must be judgeable by Yes/No, e.g., "Can independently decompose RAG retrieval pipelines." Soft skills must come with scenarios, e.g., "Can pull the business side back to acceptable standards when requirements change repeatedly."

I fix key parameters. The official website standard version is kept under 600 characters, with 4 to 6 job responsibilities, and requirements split into hard conditions and bonus points. The internal referral short version is around 120 characters, mentioning only the position, key abilities, first result, and referral bonus. The technical role evidence version lists real tech stacks and problems, avoiding "proficient." The sales role result version lists customer types, sales cycles, deal amounts, and CRM tools. The overseas role compliance version omits age/marital status, uses placeholders for salary ranges, and specifies language requirements for work scenarios.

When generating, pass both the "Requirements Gap List" and the field table to WorkBuddy, letting it run by role. The HRBP produces the draft, the interviewer deletes fluff, the compliance reviewer checks banned words, and the channel copywriter tweaks titles and opening paragraphs. I generally don't let it revise five versions at once; I produce one version, confirm it's okay, then batch derive others. This habit was learned through stumbling blocks.

I also set permissions strictly. In the WorkBuddy project, the business side gets "Comment Only," HR gets "Edit," and the hiring lead gets "Publish." If the field table syncs to a shared spreadsheet, columns for hard skills, salary bands, and banned words are locked, editable only by HR. I've also used our ATS for about a month, trying to keep field names consistent with WorkBuddy tables, otherwise imported positions get misclassified. Collaboration fears most turning a one-sentence requirement into a one-sentence promise via telephone game.

Daily maintenance isn't complex. Every Friday, I have WorkBuddy merge new positions from the week into the position library and check for duplicate JDs. Monthly reviews look at which versions have high open rates but poor resumes, and which hard conditions deter applicants. Not all positions suit AI drafting; non-technical roles especially need manual touch. AI can verify if the link from requirement to position is complete, but details still need human polishing.

Looking forward, HR writing JDs will likely focus more on who can turn vague requirements into verifiable recruitment entry points. What tools like WorkBuddy truly save is the whole day spent wrestling with the business side afterwards.