15 AI Prompts for Recruiters: Source Packs, Boolean Search, Talent Maps, and Evidence Review

The reusable prompt pattern

Strong sourcing prompts contain four things: role context, a specific artifact to produce, a structured output format, and trust constraints. “Find me great candidates” is weak because it asks the model to skip the search and evidence layers.

Context: [paste req + HM notes]
Task: [one sourcing artifact]
Return: [explicit sections or table]
Rules: [append the guardrail block below]

15 recruiter AI prompts

Act as a senior technical sourcer reviewing this requisition before search begins.
Return:
1. Stated must-haves
2. Hidden or ambiguous requirements
3. Preferences disguised as requirements
4. Missing evidence standards
5. Constraints that could collapse the market
6. Ten hiring-manager questions that would materially change the search
Do not solve the ambiguities yourself. Label them.

Rules for every response:
- Do not invent candidates, profiles, employers, jobs, URLs, contact data, credentials, or verification.
- Label assumptions explicitly.
- Keep public clearance language as a breadcrumb only; current status requires the appropriate authorized process.
- Do not infer protected or sensitive traits.
- Do not silently merge identities.
- Separate observed evidence from inference and missing information.
- Do not recommend automated rejection, auto-outreach, or another consequential action without human review.

For each must-have in this role, define what job-relevant evidence could support it.
Use four columns:
- Requirement
- Strong evidence
- Weak breadcrumb
- What must be confirmed later
Avoid title-only evidence when the underlying work can be described more precisely.

Rules for every response:
- Do not invent candidates, profiles, employers, jobs, URLs, contact data, credentials, or verification.
- Label assumptions explicitly.
- Keep public clearance language as a breadcrumb only; current status requires the appropriate authorized process.
- Do not infer protected or sensitive traits.
- Do not silently merge identities.
- Separate observed evidence from inference and missing information.
- Do not recommend automated rejection, auto-outreach, or another consequential action without human review.

Build a title map for this role.
Return:
- Exact titles
- Common equivalents
- Adjacent titles with transferable work
- Legacy titles
- Company-specific variants to investigate
- Titles that look similar but usually create false positives
For every adjacent title, explain the transferability hypothesis in one sentence.

Rules for every response:
- Do not invent candidates, profiles, employers, jobs, URLs, contact data, credentials, or verification.
- Label assumptions explicitly.
- Keep public clearance language as a breadcrumb only; current status requires the appropriate authorized process.
- Do not infer protected or sensitive traits.
- Do not silently merge identities.
- Separate observed evidence from inference and missing information.
- Do not recommend automated rejection, auto-outreach, or another consequential action without human review.

Turn this requisition into three independent search lanes:
1. Precision lane
2. Evidence-led lane
3. Adjacent-market lane
For each lane return:
- hypothesis
- titles
- evidence terms
- likely source surfaces
- false-positive risks
- what outcome would justify more effort
Do not combine the lanes into one giant query.

Rules for every response:
- Do not invent candidates, profiles, employers, jobs, URLs, contact data, credentials, or verification.
- Label assumptions explicitly.
- Keep public clearance language as a breadcrumb only; current status requires the appropriate authorized process.
- Do not infer protected or sensitive traits.
- Do not silently merge identities.
- Separate observed evidence from inference and missing information.
- Do not recommend automated rejection, auto-outreach, or another consequential action without human review.

Create three Boolean variants for this role:
- strict title-led
- skill/evidence-led
- adjacency-led
Then critique each string for:
- brittle title assumptions
- missing synonyms
- unnecessary AND conditions
- false-positive risk
- source-specific syntax concerns
Make one change at a time in the recommended revisions.

Rules for every response:
- Do not invent candidates, profiles, employers, jobs, URLs, contact data, credentials, or verification.
- Label assumptions explicitly.
- Keep public clearance language as a breadcrumb only; current status requires the appropriate authorized process.
- Do not infer protected or sensitive traits.
- Do not silently merge identities.
- Separate observed evidence from inference and missing information.
- Do not recommend automated rejection, auto-outreach, or another consequential action without human review.

Create open-web search lanes for this role using only public evidence surfaces that plausibly match the profession.
For each surface return:
- why it belongs
- one Google site-search query
- one noise/exclusion strategy
- what the result can support
- what it cannot verify
Do not invent profile URLs or people.

Rules for every response:
- Do not invent candidates, profiles, employers, jobs, URLs, contact data, credentials, or verification.
- Label assumptions explicitly.
- Keep public clearance language as a breadcrumb only; current status requires the appropriate authorized process.
- Do not infer protected or sensitive traits.
- Do not silently merge identities.
- Separate observed evidence from inference and missing information.
- Do not recommend automated rejection, auto-outreach, or another consequential action without human review.

Build donor-company categories before naming companies.
Score each category on:
- work-pattern similarity
- technical/operational environment
- customer or mission similarity
- regulation/risk similarity
- geography/work model
- compensation/career-path plausibility
Then propose Primary, Adjacent, and Stretch donor examples with a written rationale. Treat company membership as a search hypothesis, not proof of fit.

Rules for every response:
- Do not invent candidates, profiles, employers, jobs, URLs, contact data, credentials, or verification.
- Label assumptions explicitly.
- Keep public clearance language as a breadcrumb only; current status requires the appropriate authorized process.
- Do not infer protected or sensitive traits.
- Do not silently merge identities.
- Separate observed evidence from inference and missing information.
- Do not recommend automated rejection, auto-outreach, or another consequential action without human review.

Here are examples of profiles my search keeps returning and why they are wrong.
Diagnose the recurring false-positive patterns.
Return:
- likely query cause
- evidence term to add
- title or keyword to remove
- exclusion worth testing
- whether the problem is title, skill, domain, location, donor, or source-lane design
Do not rewrite the entire search unless necessary.

Rules for every response:
- Do not invent candidates, profiles, employers, jobs, URLs, contact data, credentials, or verification.
- Label assumptions explicitly.
- Keep public clearance language as a breadcrumb only; current status requires the appropriate authorized process.
- Do not infer protected or sensitive traits.
- Do not silently merge identities.
- Separate observed evidence from inference and missing information.
- Do not recommend automated rejection, auto-outreach, or another consequential action without human review.

This sourcing lane is producing too few evidence-fit leads.
Propose an expansion sequence that changes one assumption at a time.
Rank changes by likely information value, not by how broad they make the search.
Consider title, evidence, donor companies, location, domain, source surface, and nice-to-have tools.
Preserve the true non-negotiables.

Rules for every response:
- Do not invent candidates, profiles, employers, jobs, URLs, contact data, credentials, or verification.
- Label assumptions explicitly.
- Keep public clearance language as a breadcrumb only; current status requires the appropriate authorized process.
- Do not infer protected or sensitive traits.
- Do not silently merge identities.
- Separate observed evidence from inference and missing information.
- Do not recommend automated rejection, auto-outreach, or another consequential action without human review.

Review the supplied candidate evidence against this requisition.
Return only:
1. Observed evidence
2. Which requirement each item may support
3. Missing information
4. Ambiguous or stale claims
5. Verify-next questions
6. A short hiring-manager summary that preserves uncertainty
Do not infer facts not present in the evidence.

Rules for every response:
- Do not invent candidates, profiles, employers, jobs, URLs, contact data, credentials, or verification.
- Label assumptions explicitly.
- Keep public clearance language as a breadcrumb only; current status requires the appropriate authorized process.
- Do not infer protected or sensitive traits.
- Do not silently merge identities.
- Separate observed evidence from inference and missing information.
- Do not recommend automated rejection, auto-outreach, or another consequential action without human review.

I have two source profiles that may represent the same person.
List the non-sensitive anchors that support or weaken the identity hypothesis.
Return:
- evidence for same person
- evidence against
- unresolved conflicts
- what a recruiter should inspect next
Do not decide the merge automatically if meaningful uncertainty remains.

Rules for every response:
- Do not invent candidates, profiles, employers, jobs, URLs, contact data, credentials, or verification.
- Label assumptions explicitly.
- Keep public clearance language as a breadcrumb only; current status requires the appropriate authorized process.
- Do not infer protected or sensitive traits.
- Do not silently merge identities.
- Separate observed evidence from inference and missing information.
- Do not recommend automated rejection, auto-outreach, or another consequential action without human review.

Summarize this week of sourcing evidence for a hiring manager.
Use this structure:
- what we tested
- what the market produced
- repeated false positives
- donor/title lanes that added signal
- constraints that appear to be collapsing the market
- two tradeoff decisions needed from the HM
- next experiment after each possible decision
Keep individual candidate judgments secondary to market patterns.

Rules for every response:
- Do not invent candidates, profiles, employers, jobs, URLs, contact data, credentials, or verification.
- Label assumptions explicitly.
- Keep public clearance language as a breadcrumb only; current status requires the appropriate authorized process.
- Do not infer protected or sensitive traits.
- Do not silently merge identities.
- Separate observed evidence from inference and missing information.
- Do not recommend automated rejection, auto-outreach, or another consequential action without human review.

Use these past candidate outcomes and rejection reasons to design an ATS rediscovery lane.
Return:
- segments worth revisiting
- stale-context checks
- opt-out or prior-contact checks
- patterns from prior successful profiles
- new search terms or donor hypotheses derived from history
Do not assume past interest is current interest.

Rules for every response:
- Do not invent candidates, profiles, employers, jobs, URLs, contact data, credentials, or verification.
- Label assumptions explicitly.
- Keep public clearance language as a breadcrumb only; current status requires the appropriate authorized process.
- Do not infer protected or sensitive traits.
- Do not silently merge identities.
- Separate observed evidence from inference and missing information.
- Do not recommend automated rejection, auto-outreach, or another consequential action without human review.

Challenge the statement: "we have searched the market."
Given these lanes, query variants, duplicate rates, donor coverage, adjacent-title tests, and recent new-lead rates, identify what has and has not actually been tested.
Return:
- evidence supporting exhaustion
- missing independent lanes
- repeated search paths masquerading as new coverage
- the single highest-information next experiment.

Rules for every response:
- Do not invent candidates, profiles, employers, jobs, URLs, contact data, credentials, or verification.
- Label assumptions explicitly.
- Keep public clearance language as a breadcrumb only; current status requires the appropriate authorized process.
- Do not infer protected or sensitive traits.
- Do not silently merge identities.
- Separate observed evidence from inference and missing information.
- Do not recommend automated rejection, auto-outreach, or another consequential action without human review.

Turn this week's sourcing notes into project memory.
Return:
- hypotheses tested
- evidence that supported or contradicted each
- queries worth preserving
- false-positive rules learned
- donor companies promoted/demoted
- HM tradeoffs approved
- source lanes with unique contribution
- unresolved questions
- next week's first experiment
Do not rewrite history to make the strategy look successful.

Rules for every response:
- Do not invent candidates, profiles, employers, jobs, URLs, contact data, credentials, or verification.
- Label assumptions explicitly.
- Keep public clearance language as a breadcrumb only; current status requires the appropriate authorized process.
- Do not infer protected or sensitive traits.
- Do not silently merge identities.
- Separate observed evidence from inference and missing information.
- Do not recommend automated rejection, auto-outreach, or another consequential action without human review.

How to use the prompt library in SourcingOS

Start with role intake and evidence standards, then create search lanes. Use JD Strategy Tool for the source-pack draft, BooleanOS for syntax variants, Search Lane Expander for a thin market, and Candidate Search when you move from strategy into evidence review.

If you want the instructional version rather than a reference library, use the AI Sourcing Prompts training module. For product evaluation, use the 8-task AI sourcing harness.

FAQ