Score every new Greenhouse application against your role scorecard with Claude, so recruiters start each day with a ranked shortlist.
An open role can pull in 200 applications in a week, and reading each one properly takes three to five minutes. Recruiters end up skimming, strong candidates sit unread for days, and screening quality depends on who had time that afternoon.
Copy the role’s scorecard from Greenhouse into a plain-text block: 4–6 must-haves and 2–3 nice-to-haves, each with what a 1, 3 and 5 looks like. Leave out anything that can act as a proxy for age, gender or ethnicity, such as graduation year.
Start a scenario with Webhooks > Custom webhook and copy its URL. In Greenhouse, open Configure > Dev Center > Web Hooks and create a hook for the “Candidate has submitted application” event that points at that URL.
Add HTTP > Get a file to download the attachment where `type` is `resume` from the webhook payload. Claude accepts PDF documents directly, so most resumes need no separate text-extraction step.
Add the Anthropic Claude module with the screening prompt below, the scorecard text and the resume. Set temperature to 0 and request JSON so every candidate is scored the same way.
Call the Harvest API from an HTTP module (`POST /v1/candidates/{id}/activity_feed/notes`) to add the scores and evidence as a private note, then apply a tag such as `ai-score-4-plus` or `ai-score-below-3`. Recruiters sort the review queue by tag.
Never auto-reject. The score orders the queue, and a recruiter still reads every candidate before advancing or declining. In week one, spot-check 20 scored applications against your own judgement and tighten scorecard wording where you disagree.
You are screening applicants for the role of {{role_title}}. Score the candidate only against the scorecard below. Do not consider name, age, gender, nationality, photos, school prestige or employment gaps.
Scorecard:
{{scorecard}}
For each criterion, give a score from 1 to 5 and quote the line from the resume or cover letter that supports it. If there is no evidence, score 1 and write "no evidence".
Return JSON:
{"criteria": [{"name": "", "score": 0, "evidence": ""}], "overall": 0, "summary": "two sentences a recruiter can read in 10 seconds", "questions_for_phone_screen": ["", ""]}
Resume:
{{resume_text}}
Cover letter:
{{cover_letter}}An hour before each discovery call, get a one-page brief on the account, the attendee and recent news, saved to Salesforce.
Turn one campaign brief into 20 Meta and Google ad variants, score each against your brand rules and limits, and queue the best for approval.
Make checks QuickBooks every morning for overdue invoices, and Claude drafts Gmail reminders that get firmer at 3, 14 and 30 days late.