recruiter-prompt-bank · 239 records
Labels: hiring manager 3 of 239, vendor 173.
The note
173 of the 239 records in my bank of hiring-side AI prompts carry the vendor label. The rest: recruiter 36, unknown 27, and 3 labelled hiring-manager, all three ranking prompts from third-party articles.
Every record carries a role, so the field is complete. What it shows is a dataset in the vendor's voice. Seventy-two percent of it is content a company published, and only 39 of those 173 are screening rubrics describing how a product reads a resume. The vendor label also covers 54 ChatGPT and 19 Claude tool prompts, 59 sourcing queries and 42 records from article pages. My reading is that these are the idealised version: what the tool is meant to ask, not what a recruiter typed at 6 pm with a requisition open.
The 36 recruiter-labelled records came almost entirely from articles too. The one place I expected to find recruiters talking about their own prompts, public forums, gave nothing: searches there surfaced no thread where a recruiter shared an actual prompt.
The 27 unknowns could include recruiters. That makes 72 percent a floor for vendors and 15 percent a floor for recruiters, and the space between them is a mix I cannot fully attribute.
Limits. The role field is checked only against a fixed list of allowed values; no file says how each label was assigned. "Vendor" covers any company's content. It counts records rather than authors, and 58 of 79 source pages are vendor pages. Three hiring-manager records support no rate at all.
If nearly all the public record of recruiter prompting was published by vendors, what would a benchmark built from it actually be measuring?
Where the numbers live
Roles assigned by the collector per record. Vendor covers any company's content, including 42 article records.