Writing a JD with AI in 2026 is not one prompt. It is a four-step loop: an agent drafts from calibrated inputs, the hiring manager edits, the recruiter normalizes language, and the final lands somewhere both the agent and the team can revisit. Skip any step and you get a generic post.
The four-step loop
1. Calibrate inputs. A draft from ChatGPT or Claude is only as good as what you feed it. Pull the leveling rubric, the comp band from your HRIS, and the last three accepted offers at the target level. Paste those into the prompt as context. Skip this and the model invents responsibilities and underprices the role.
2. Generate a structured draft. Ask for sections in a fixed order: role summary, the problem the role solves, 90-day scope, required experience, nice-to-haves, comp range, location and visa policy. Claude tends to produce stronger structured prose, ChatGPT is faster for iterative rewrites.
3. Hiring-manager edit pass. The manager strikes anything that does not match the headcount plan, names the tools the person will actually use, and lists the two or three projects in flight. Twenty minutes. The difference between a JD that pulls qualified candidates and one that pulls everyone.
4. Recruiter normalization. The recruiter runs the draft through Textio or Datapeople to flag gendered language, jargon, and inflated requirements. Both score readability and inclusion (Textio on why job postings matter). The recruiter also strips legal landmines: unbounded availability clauses, missing comp ranges where pay transparency laws apply.
The JD then posts to Greenhouse, Workday, or Lever and syncs to LinkedIn. Clean inputs at the top produce clean signals downstream, which is what makes good AI candidate sourcing and AI candidate screening possible.
Worked example: a senior backend engineer at a Series B company
The hiring manager opens a Claude session with the leveling rubric, a $190k-$240k band from Carta, and a note that the role owns the payments service. Claude returns a structured draft in 30 seconds. The manager edits: names Stripe, Temporal, and Postgres as the stack, swaps "10+ years of experience" for "experience leading a distributed systems team," adds the on-call rotation. The recruiter runs it through Datapeople, replaces "rockstar" and "ninja" with plain nouns, confirms the comp range matches NY pay transparency law. The JD posts to Greenhouse, LinkedIn syndication kicks off automatically. Total: 90 minutes.
Where the workflow breaks: the draft lives in five places
The Claude draft lives in a chat that resets. The manager's edits live in a Google Doc. The recruiter's normalized version lives in Textio. The final lives in Greenhouse. The comp rationale lives in a Slack thread. Six months later, when the role reopens, nobody can reconstruct why the JD says what it says. The agent that helped draft it has no memory of the conversation. One way to solve this is a workspace like Dock that holds the JD draft, the comp rationale, edit history, and normalization notes as persistent rows, with a greenhouse_job_id pointer back to the canonical posting. The ATS stays the system of record. The interpretation lives somewhere the agent can re-read next quarter. That is also how agent identity works in practice: the agent does not need to remember, it needs a workspace it can read.
Why this matters
JDs set the top of the funnel. A bad JD pulls a worse pipeline, and no amount of downstream screening fixes that. Teams shipping good hires in 2026 are not the ones with the best prompt. They are the ones whose JD process has structure, named tools, and a place where the reasoning persists.
The full pillar covers the rest of the loop: how to do hiring with AI in 2026. For the persistent workspace alongside the recruiting stack, see Dock for HR.
FAQ
Q: Should I use ChatGPT or Claude to draft a JD? A: Either works. Claude tends to produce stronger long-form structured drafts on the first pass. ChatGPT is faster for short iterative rewrites. The bigger lever is what you feed the model: a leveling rubric, a real comp band, and a named stack beat any model choice.
Q: Do I still need Textio or Datapeople if I am already using a foundation model? A: Yes. They score bias and readability against benchmarked data, which a chat model does not. They also flag compliance issues (pay transparency, ADA language) that a generic model will miss.
Q: Where should the final JD actually live? A: The posted version belongs in your ATS (Greenhouse, Workday, Lever, Ashby). The draft, the comp rationale, and reviewer notes belong in a persistent workspace the agent can re-read. Do not let either one live in a chat window.
Q: How long should the workflow take for one role? A: For a standard IC role, 60 to 90 minutes end-to-end. The hiring-manager edit pass is the longest single step at roughly 20 minutes, and it is the step you cannot skip.