Guides
Setting up KI screening correctly - write the prompt, validate, calibrate
KI pre-sorting stands or falls with the prompt. Here's how to write a good one, test it in 20 applications, and tune over time - without overriding your HR judgement.
Key takeaways
1. Reduce profile to three mandatory skills
Look at the job ad. Which three skills can you not drop without the role becoming a different one? Those three go into the prompt. Everything else is plus or irrelevant.
2. Write the prompt in three-sentence structure
Example: 'Look for 3+ years backend experience with Python or Go, sufficient DB experience (Postgres or MySQL), and EU work permit. Pluses: Kubernetes experience, open-source contributions with own code. Frontend experience is not relevant for this role.'
3. Validate against 20 talent-pool applications
Pick 20 applications from your pool, ideally mixed 'top back then' and 'didn't fit back then'. Have the prompt score. Compare top-5: does it match your memory? If yes, continue. If not, back to step 2.
4. Go live and observe
Flip the switch in role editing. Next 30 applications come with score + reasoning. Spot-check every 5-10 applications: does the reasoning match the CV? If reasoning hallucinates ('person has 8 years of experience' for a 3-year CV), that signals a prompt problem.
5. Re-calibrate after 50 applications
After 50 applications with the prompt: look at score distribution. Are 80% of your invitees in the top 20%? If yes, continue. If 'top invitees' land in mid-score, adjust - usually a plus component is missing or a mandatory component is too hard.
What a good screening prompt contains
Three components in this exact order. One - mandatory skills (max 3). 'Look for Python + backend experience + EU work permit'. Two - strong pluses (max 2). 'Pluses: Kubernetes experience and open-source contributions'. Three - what's not relevant. 'Frontend experience is not relevant for this role'.
Why exactly three. A longer prompt isn't evenly weighted - the seventh criterion has statistically half the impact of the third. Listing 12 requirements gives you a score primarily reflecting the first three. Focus on the three that genuinely decide.
Anti-patterns we see often
One - 'look for a top application impression'. Meaningless to the KI; it doesn't know what 'top impression' means. Stay concrete: 'look for structured CV language and concrete project descriptions'.
Two - 'assess cultural fit'. Forbidden under AGG, risky under AI Act, technically unreliable. Ask 'how does this person work with others?' in the structured interview - not in KI screening.
Three - 'reject applications under score 40'. Auto-rejection without human review is forbidden (Art. 22 GDPR). The system may at most 'auto-sort into a lane', never 'send auto-rejection'.
Validate before trusting the KI
Before going live, have the KI score 20 historical applications from your talent pool. Compare scores with your own past assessment: does the top-5 selection match? If yes, prompt works. If not, adjust - usually a rewording of mandatory skills suffices.
The 80% match rule of thumb: in our pilot data, top-20% KI selection matches HR judgement ~80% of the time. If your validation is clearly below that (~50%), the prompt has a problem - usually too unspecific or too creative.
FAQ
Frequently asked
Try KI BMS
Free plan, no credit card. We host in Germany. You can export and delete everything self-serve.
Read next
Recruiting with KI - a practical guide
A practical guide, not hype - with clear legal limits and concrete step-by-step instructions.
Read
AI Act 2026 for recruiting - high-risk checklist for German HR teams
Eight concrete duties + three myths + the question 'does this even apply to our small HR team?'
Read
Publishing your first job with KI BMS - online in 8 minutes
Step-by-step from empty account to public role with application form and KI pre-sorting. No magic, no setup day.
Read