# Job Description Screening Pack

Turn any job description into a structured screening rubric, then score resumes against it with quoted evidence and bias guardrails. Works with any capable AI assistant — ChatGPT, Claude, and Gemini all handle it.

## What you need

- An AI assistant that accepts pasted text or file uploads (ChatGPT, Claude, Gemini)
- The job description for the role
- Resumes as PDFs or pasted text
- About 20 minutes to build the rubric

## Setup

1. Run Prompt 1 on your job description in a fresh chat.
2. Edit the rubric it returns by hand: cut criteria you would never reject someone over, fix the weights, tighten the must-haves. The model drafts the rubric; you decide it.
3. Open a new chat for scoring. Paste your final rubric inside Prompt 2 with up to five resumes; small batches keep scoring careful.
4. Run Prompt 3 in the same chat before you act on any score.

## Prompt 1 — Build the rubric

```
Here is a job description. Turn it into a screening rubric I can score resumes against.

[PASTE JOB DESCRIPTION]

Structure the rubric as:

1. MUST-HAVES: 3-5 requirements that are genuine deal-breakers. Only include items the job description states as required. Phrase each as a yes/no question, with a note on what counts as evidence on a resume.
2. WEIGHTED CRITERIA: 4-6 scoring criteria. For each, give a weight (all weights sum to 100) and describe what strong (3), adequate (2), and weak (1) evidence looks like. No evidence scores 0.
3. IGNORE LIST: restate the list below so I can confirm none of it leaked into the criteria.

Rules:
- Do not invent requirements that are not in the job description.
- If the job description is vague on something important (like seniority), ask me instead of guessing.
- The rubric must not use as criteria: candidate name, gender, age or graduation year, nationality, address or commute distance, photos, school prestige (unless the role legally requires a specific credential), employment gaps, or current employment status.
- Do not score anything yet.
```

## Prompt 2 — Score resumes

```
Here is my final screening rubric, followed by candidate resume(s). Score each resume against the rubric.

[PASTE YOUR EDITED RUBRIC]

Rules:
- For every must-have, answer PASS, FAIL, or UNCLEAR, with a word-for-word quote from the resume as evidence. UNCLEAR is not a fail; it means a human should look.
- For every weighted criterion, give a score of 0-3 with a word-for-word quote as evidence. If there is no evidence, score 0 and write NOT FOUND. Never infer skills the resume does not state.
- Ignore completely: name, gender, age or graduation years, nationality, address, photos, school prestige, employment gaps, current employment status. If you catch any of these influencing a judgment, flag it.
- Score each candidate independently against the rubric. Do not compare candidates to each other.
- Output per candidate: must-have results, criterion scores with quotes, the weighted total with arithmetic shown, and two sentences naming the strongest evidence and the biggest gap.
```

## Prompt 3 — Audit pass

Run this in the same chat, right after scoring:

```
Audit your own scoring. For each candidate:

1. Verify every quote you cited appears in the resume word for word. Flag any quote you cannot find.
2. Recalculate each weighted total and show the arithmetic.
3. Re-read your reasoning for anything from the ignore list (name, age, gaps, school prestige, address). Flag anything that may have influenced a score.
4. List every UNCLEAR and every 0 under "NEEDS HUMAN REVIEW" with a reason.

If everything checks out, say so explicitly.
```

## Bias guardrails

The prompts carry the guardrails, but they only hold if the rest of your process does:

- A human makes every decision. The output is a reading aid that gets every resume the same first pass, never a verdict.
- Never let a FAIL trigger an automatic rejection. A resume can fail a must-have because it is written badly, not because the candidate lacks the skill.
- Some jurisdictions regulate automated tools in hiring. Check the rules where you operate, and keep records of your rubric and who made each call.
- If you edit the prompts, keep the ignore list and the quote-as-evidence rule. They are what make the scores checkable.

## Calibrate before you trust it

For your first role, screen five resumes yourself before running the prompts, then compare. Where you disagree with the scores, fix the rubric: a weight is wrong, a must-have is really a nice-to-have, or a criterion is too vague to score. Then re-run.

## When this outgrows DIY

This recipe works when you hire for a role or two at a time. It stops being the right tool when applications arrive faster than anyone can paste them into a chat, when scores need to land in your ATS automatically, or when you need a consistent audit trail across every hiring manager. At that point a built system earns its keep: it reads each application as it arrives, scores it against your rubric, queues interviews, and keeps a human making every call.

That is the kind of system agentclaw builds. Start with a free AI opportunity audit at https://agentclawhq.com/book

---

Free resource from agentclaw — AI agents and workflow automation for growing businesses. https://agentclawhq.com