AI-assisted job description writing without sending req details to a third party
An LLM is good at turning a hiring manager's rough req notes into a clean job description. Those notes also carry unannounced headcount plans and internal salary bands, so a self-hosted model is the safer place to draft before the role is public: the req goes to your own model, not to an AI vendor.
Not legal advice. This post describes an infrastructure setup, not employment law guidance. Job description requirements, including pay transparency and non-discrimination rules, vary by jurisdiction. Check with your own HR and legal counsel before publishing.
Why req details don't belong on a third-party API
A draft req is a leading indicator of a company's plans: which team is growing, which role is opening, and almost always a salary band that isn't public yet. Feeding it into a general-purpose AI product gives a third party a copy before the role posts, under whatever retention policy applies to that vendor's plan. It's a smaller-stakes version of the problem in private AI for internal communications: pre-announcement information is better kept out of a general-purpose AI vendor's systems.
On a dedicated instance, the hiring manager's notes, the salary band and every draft go to the company's own model and nowhere else until the company publishes the finished posting. A rented Spark is still GPUwerk's hardware, operated under its published DPA.
What it's good for
The main use is turning rough, unstructured notes into a clean draft: taking a hiring manager's bullet points and a template the company already uses, and producing a first draft with a consistent structure, tone, and length. It also checks consistency across a batch of postings: whether the same seniority level gets comparable language across different teams' reqs. That matters for candidate experience and, in some jurisdictions, for pay transparency rules.
A second use is generating variations for different channels: a longer version for the careers page, a shorter one for a job board with a character limit, without a recruiter rewriting the same content three times by hand.
Where it stops being a drop-in replacement
A drafted job description still needs review by someone who understands the legal requirements that apply to the posting: pay transparency laws that require a disclosed salary range in some jurisdictions, language that could read as discriminatory even unintentionally, and internal consistency with how the company levels and titles roles. An LLM has no accountability for getting any of that wrong, and no reliable way to know which jurisdiction's rules apply unless it's told explicitly and correctly. HR review before publishing is required; that step decides whether the posting is compliant, not the drafting step.
One known failure mode: models trained on a wide mix of internet job postings can reproduce subtly gendered or otherwise skewed phrasing. Have a reviewer read every draft for that before it goes out.
A concrete example
A company's engineering director sends a recruiter a few rough paragraphs about a senior backend role: what the team owns, the tech stack, why the seat opened up. The recruiter runs it through a prompt template that follows the company's standard job-description format and produces a clean draft in the company's usual voice. The recruiter reviews it, adjusts a few lines, checks the salary range against the internal band and the jurisdiction's disclosure rules, and posts it. The unannounced headcount plan and the salary band it started from went to the company's own model, never to an AI vendor.
A light workload for one Spark
Short prompts, short outputs, no large context. A single DGX Spark handles a company's whole hiring pipeline of drafting with room to spare, and at $0.79/hour on demand it can run through a hiring push without the cost becoming a factor. See pricing for the reserved rate if the Spark also serves other internal workflows between hiring cycles.