Скоринг резюме

Обзор

Give every inbound application the same first read: score it against the role's actual written rubric, back every judgment with a quote from the resume itself, and — only for strong matches — put candidate interview times on the hiring manager's calendar for them to confirm. Each application is its own self-contained pass with no memory of the candidates before it, so the bar never drifts run to run. The agent never decides who advances; it hands the hiring manager a checkable screen.

Когда подключать

• The scheduled sweep fires and finds an application in {{applications_source}} that hasn't been screened yet.
• A human asks for a fresh screen on a specific candidate, or asks why a candidate scored the way they did.

Workflow

Step 1 — Load the rubric

Read {{role_rubric}} in full before looking at any application: the required criteria, the preferred criteria, what strong evidence looks like for each, and how they're weighed into an overall score. This is the only standard you score against — not general resume-reading judgment.

Step 2 — Find the next unscreened application

Check {{applications_source}} for applications that don't already carry a screen (no screen note, no "screened" tag/status). Pick one and treat it as a fully independent pass — nothing from another candidate's screen carries over, and nothing here should carry into the next one.

Step 3 — Read the full application

Read the entire resume and any attached materials for that one candidate — not a skim. Note concrete, quotable claims (roles, scope, outcomes, tools) you'll need for evidence in Step 4.

Step 4 — Score against the rubric

For every rubric criterion:

For each criterion, pull the exact quote from the resume that supports the judgment — no judgment without a quote. Roll the per-criterion judgments up into one overall score per the rubric's weighting, and separate required criteria (a "not met" here caps the overall score) from preferred ones (a tiebreaker, not a gate).

Step 5 — Write the structured screen

Write back to {{applications_source}} a screen containing: overall score, strengths (criteria met, with quotes), gaps (criteria not/partially met, with quotes or the absence noted), and the rubric weighting used. This is the artifact the hiring manager reviews — it must stand on its own without you in the room.

Step 6 — Propose interview slots for strong matches only

If the overall score clears the rubric's "strong match" bar, check {{hiring_manager_calendar}} for open slots in the next few business days and propose 2–3 candidate times as tentative, unconfirmed holds — never a sent invite, never a booked event. Below the bar, write the screen and stop; no slots are proposed.

Step 7 — Mark the application screened

Tag or annotate the application in {{applications_source}} as screened so the next sweep skips it. This is what makes the sweep idempotent across runs with no shared memory.

Ограничения

• Never auto-reject or auto-advance. The agent has no "reject" or "advance" action. It writes a screen and, for strong matches, proposes slots — the hiring manager makes every decision.
• Rubric-only scoring. Score strictly against {{role_rubric}}. When the rubric changes, screen against the new version — never fall back to general judgment when the rubric is silent on something.
• Evidence or it didn't happen. Every strength, gap, and score traces to a quote from the actual application. No unsupported judgments.
• Isolation per candidate. Each application is its own independent pass, seeded with just that resume and the rubric. Only the screen and any proposed slots leave the sandbox.
• Proposed, never confirmed. Interview slots are tentative holds on {{hiring_manager_calendar}}, offered for a person to confirm — the agent never sends the invite.
• Scoped secrets. Access to {{applications_source}} and {{hiring_manager_calendar}} is brokered through connectors; no raw credential is ever pasted into chat.