How to write search prompts for AI sourcing

An AI search doesn't guess what's in your head. You give it a description in plain language and it hands back candidates; how good they are depends almost entirely on how you wrote that description. Here's how to build it so the first list is already useful.

The most common mistake: asking for the title, not the person

"I need a Backend Developer" is a job title, not a profile. The search will bring back everyone with those two words on their LinkedIn, including the person who put them there after a course five years ago.

Describe the person instead: what they built, with which stack, in what kind of company and at what scale. "Backend engineer in Python or Go who has taken a service to production with real traffic, ideally at a fintech or a marketplace, not an agency" filters far better.

The five parts of a good prompt

1. The concrete work. Not the title: the activity. "Migrated a monolith to services", "built the data function from scratch", "closed enterprise contracts above 50k".
2. Company context. Size, industry, stage. "Series A or B startup", "traditional bank", "consultancy, not product". Context predicts how someone works.
3. Level signals. Years matter little. Say what they should already have done: led others, been first in their role, reported to leadership.
4. What rules people out. Be explicit: "no consulting-only backgrounds", "I want people who built, not people who only maintained".
5. Location and setup. City, country, whether they need to overlap hours with a team in another time zone.

Example, weak to good

Weak: "Senior Product Manager with SaaS experience."

Good: "PM who has owned a B2B self-serve product end to end: set the roadmap, worked closely with design and engineering, and moved an activation or retention metric with numbers they can talk through. Preferably at a 50 to 300 person company. No agency or consulting PMs. Remote with US or LATAM hours overlap."

The second gives you a shortlist you can send for review. The first gives you 400 profiles and three hours of filtering.

After the first list

Don't rewrite the whole prompt. Adjust in short steps: "from these, keep the ones who stayed at a single company for the last 3 years" or "weight marketplace experience higher". AI search performs better when you treat it as a conversation, not a form you fill once.

How Klyver does it

Klyver's Sourcing Agent runs this kind of search every night against your open roles and leaves the shortlist ready in the morning, inside the same ATS. You define the profile once when you open the role; the agent repeats it without you rewriting anything.