Tip 1Freeze literal requirements before you expand the search
Separate what the hiring manager actually requires from synonyms, adjacent titles, donor-company ideas, and discovery-only concepts before you run searches.
Use it when: At intake and any time a search starts drifting away from the original role.
- Write strict requirements in plain language.
- Normalize synonyms without changing meaning.
- Put discovery expansions in a separate list.
- Require approval before loosening a true must-have.
Watch out: Do not let an AI-generated expansion silently become a hiring requirement.
intake · requirements · calibration · reviewed 2026-09-06Tip 2Build search lanes instead of one giant Boolean string
A lane should test one sourcing hypothesis at a time so you can tell why a search works or fails.
Use it when: For hard-to-fill, technical, cleared, healthcare, or executive searches.
- Create an exact-fit lane.
- Create adjacent-title and donor-company lanes.
- Add public-evidence lanes where useful.
- Measure yield separately by lane.
Watch out: One giant Boolean query hides which assumption is collapsing the market.
search lanes · boolean · strategy · reviewed 2026-09-06Tip 3Write false-positive exclusions before the search gets noisy
Exclusion logic is part of the search strategy, not cleanup work after 300 bad results.
Use it when: Whenever a role has predictable title or keyword ambiguity.
- List common false-positive personas.
- Add title and context exclusions.
- Track what is excluded by each lane.
- Review exclusions after calibration.
Watch out: Over-aggressive exclusions can hide adjacent talent that the hiring manager would accept.
precision · boolean · false positives · reviewed 2026-09-06Tip 4Calibrate on the first five candidates before scaling the search
Early recruiter and hiring-manager feedback is more valuable than expanding volume before the search hypothesis is validated.
Use it when: At the start of a difficult or ambiguous role.
- Review five deliberately varied profiles.
- Capture exact approve/reject reasons.
- Translate feedback into explicit rules.
- Re-run the search plan before adding volume.
Watch out: Do not encode vague feedback like “not senior enough” without defining the evidence that would change the decision.
calibration · feedback · learning · reviewed 2026-09-06Tip 5Turn rejection reasons into explicit search-plan deltas
Useful feedback should change a requirement, lane, exclusion, weighting, or verification question—not disappear into notes.
Use it when: After candidate review or hiring-manager calibration.
- Record the rejection reason.
- Classify it as requirement, preference, false positive, or evidence gap.
- Apply a specific search-plan change.
- Keep the change reversible.
Watch out: Do not learn protected-trait proxies or subjective preferences that are unrelated to job requirements.
learning · feedback · search plan · reviewed 2026-09-06Tip 6Measure unique source contribution, not just raw result volume
A source is valuable when it finds relevant people the rest of the stack did not already surface.
Use it when: When evaluating sourcing channels, providers, or search lanes.
- Track discoveries by source.
- Deduplicate identities before comparing yield.
- Measure unique retained candidates.
- Compare quality and cost together.
Watch out: High result counts can look impressive while contributing almost no net-new talent.
source diversity · metrics · providers · reviewed 2026-09-06Tip 7Use duplicate pressure as a search-exhaustion signal
When new lanes mostly return people you have already seen, the market may be converging and the next move should be deliberate.
Use it when: Late in a search or when expanding across multiple providers.
- Track duplicate rate by lane.
- Track net-new retained candidates.
- Compare recent yield with earlier yield.
- Decide whether to expand geography, titles, donors, or requirements.
Watch out: Duplicates do not prove the entire market is exhausted; they only show your current source hypotheses are converging.
search exhaustion · duplicates · metrics · reviewed 2026-09-06Tip 8Keep discovery evidence separate from verification
A public breadcrumb can justify investigation without proving a verification-sensitive requirement.
Use it when: For clearance, licenses, credentials, employment dates, contact data, and identity-sensitive claims.
- Record the public evidence.
- Label the assertion as observed or inferred.
- Create a verification question.
- Upgrade status only after an authoritative verification step.
Watch out: Never convert a mention of Secret clearance, RN licensure, or certification into current verified status automatically.
verification · evidence · trust · reviewed 2026-09-06Tip 9Show one best contact first and collapse alternatives
Recruiters need a usable primary contact path, not a wall of fifteen possible emails and phone numbers.
Use it when: During contact enrichment and Candidate 360 review.
- Score contact candidates by verification, source, freshness, and type.
- Promote one best work email, personal email, and phone where appropriate.
- Collapse lower-confidence alternatives.
- Retain provenance and verification status.
Watch out: More contact records are not the same thing as better contactability.
contact data · enrichment · candidate 360 · reviewed 2026-09-06Tip 10Verify contact quality before outreach
Contact discovery and outreach readiness are separate states.
Use it when: Before starting email or phone outreach from enriched data.
- Prefer recently verified professional contact paths.
- Check bounce or verification history where available.
- Respect opt-outs and employer policy.
- Keep outreach approval human-controlled.
Watch out: An enriched email address is not consent and should not bypass privacy or outreach policy.
outreach · contact verification · privacy · reviewed 2026-09-06Tip 11Require corroborating evidence before merging candidate identities
Two similar names or profiles are not enough to create one candidate record.
Use it when: When combining GitHub, resume, ATS, provider, publication, patent, or social evidence.
- Compare employer and role history.
- Compare location and timeline.
- Use stable public identifiers where available.
- Preserve conflicting evidence instead of forcing a merge.
Watch out: A false identity merge contaminates every downstream ranking, contact, and outreach decision.
identity resolution · candidate graph · evidence · reviewed 2026-09-06Tip 12Search for role evidence, not only job titles
Technical and emerging roles often have inconsistent titles, so public artifacts and capability signals can outperform title-only search.
Use it when: For AI/ML, platform, DevSecOps, cybersecurity, research, and specialist engineering roles.
- Define capability clusters.
- Search tools, artifacts, repos, publications, or project evidence.
- Use titles as one signal, not the whole query.
- Validate role relevance in context.
Watch out: Evidence of using a technology does not automatically prove depth, recency, or job-level responsibility.
technical sourcing · evidence · adjacent titles · reviewed 2026-09-06Tip 13Map donor companies by capability, not logo prestige
The best donor company is the one that repeatedly produces the work pattern your role needs.
Use it when: For market mapping, executive search, cleared programs, and specialized technical roles.
- Define the capability you need.
- Identify companies and teams that perform that work.
- Map adjacent competitors and subcontractors.
- Track which donors actually yield qualified candidates.
Watch out: A famous company name does not guarantee the candidate performed the relevant work.
donor companies · market mapping · talent intelligence · reviewed 2026-09-06Tip 14Treat GitHub as an evidence surface, not a resume database
GitHub can reveal public technical work, but the absence or presence of activity should not be treated as a complete career record.
Use it when: For software, infrastructure, AI/ML, security, and open-source-heavy searches.
- Search role-specific technologies and project evidence.
- Inspect context rather than contribution count alone.
- Pair GitHub with another identity or career source.
- Respect public-profile and contact boundaries.
Watch out: Do not penalize candidates for lacking public GitHub activity; many strong engineers work primarily in private repositories.
github · technical sourcing · public evidence · reviewed 2026-09-06Tip 15For AI/ML talent, search model and research artifacts in addition to titles
AI/ML titles are noisy; repositories, Hugging Face models, papers, benchmarks, inference tooling, and evaluation work can provide stronger discovery signals.
Use it when: For ML engineering, research engineering, MLOps, LLM infrastructure, and applied AI roles.
- Build separate research, model, infrastructure, and application lanes.
- Search Hugging Face, GitHub, and OpenAlex where relevant.
- Look for model-serving, eval, data, and deployment evidence.
- Keep academic and production evidence distinct.
Watch out: A paper author or model uploader is not automatically a production ML engineer.
ai ml · hugging face · openalex · reviewed 2026-09-06Tip 16Source data engineers by architecture clusters
A modern data role is easier to understand when you separate orchestration, transformation, streaming, warehouse/lakehouse, and platform signals.
Use it when: For data engineering, analytics engineering, streaming, and platform data roles.
- Map the target architecture.
- Create tool clusters rather than keyword soup.
- Separate batch from streaming needs.
- Search for migration and scale context where it matters.
Watch out: Do not require every tool in a modern stack when the underlying architectural experience is transferable.
data engineering · architecture · technical sourcing · reviewed 2026-09-06Tip 17Use authoritative sources for professional-license verification
Discovery can happen anywhere, but verification should come from the appropriate authoritative licensing source or approved process.
Use it when: For nursing, healthcare, licensed professional, or regulated roles.
- Use public sources for discovery.
- Identify the governing licensing authority.
- Verify status through the authoritative service.
- Record verification date and source.
Watch out: A resume or profile saying “RN” is not a substitute for current license verification.
healthcare · licenses · verification · reviewed 2026-09-06Tip 18Track marginal yield so you know when to stop searching the same way
The question is not how many profiles you reviewed; it is how many net-new relevant candidates the next unit of search effort produces.
Use it when: During long-running or aging reqs.
- Track new retained candidates per search batch.
- Compare lane yield over time.
- Watch duplicate and false-positive pressure.
- Change the search hypothesis when marginal yield collapses.
Watch out: Do not declare a market exhausted simply because one source or Boolean string stopped producing results.
search exhaustion · yield · metrics · reviewed 2026-09-06Tip 19Always be able to explain why a candidate appeared
Discovery provenance should be visible enough that a recruiter can understand which lane, source, or evidence caused the system to evaluate the person.
Use it when: Any time AI or multi-source search produces candidate results.
- Show source lane and discovery reason.
- Show requirement evidence separately.
- Expose missing evidence.
- Keep ranking rationale inspectable.
Watch out: A score without evidence is not an explanation.
ranking · explainability · candidate review · reviewed 2026-09-06Tip 20Budget expensive provider actions separately from search
Contact enrichment, paid people-data lookups, deep research, and live AI search can create variable cost even when ordinary product usage is cheap.
Use it when: When adding external APIs, people-data providers, AI research, or contact enrichment.
- Define per-action costs.
- Set project and provider budgets.
- Add rate limits and failure circuits.
- Measure cost per retained candidate or useful outcome.
Watch out: Unlimited public endpoints connected to paid providers can become both a security risk and a cost leak.
providers · cost control · security · reviewed 2026-09-06Tip 21Treat external text as untrusted data in AI workflows
A resume, webpage, profile, or job description can contain instructions, but those instructions should never redefine the AI system’s permissions.
Use it when: Any AI workflow that reads resumes, websites, job descriptions, provider data, or user-supplied documents.
- Separate system instructions from retrieved text.
- Use tool allowlists.
- Validate structured outputs.
- Log consequential tool actions.
- Test prompt-injection cases.
Watch out: Do not let retrieved content instruct the model to reveal secrets, call new tools, or bypass recruiter approval.
ai security · prompt injection · tool safety · reviewed 2026-09-06Tip 22Use AI to propose; keep consequential recruiting decisions human-controlled
AI is strongest when it structures, expands, summarizes, and explains evidence while recruiters retain authority over consequential actions.
Use it when: Across intake, sourcing, ranking, contact enrichment, outreach, and learning workflows.
- Let AI propose search plans.
- Require evidence for candidate claims.
- Keep outreach approval explicit.
- Make learned rules visible and reversible.
Watch out: Do not turn convenience automation into silent hiring judgment.
human in the loop · responsible ai · agentic recruiting · reviewed 2026-09-06Tip 23Use live search for changing external facts, not for undocumented product truth
Current pricing, integrations, product launches, provider policies, and market developments should be checked live; SourcingOS behavior should come from canonical SourcingOS documentation.
Use it when: When answering time-sensitive recruiting-tech questions.
- Check canonical SourcingOS knowledge first.
- Use live web research for changing external facts.
- Prefer first-party sources.
- Show source provenance and dates.
- Call out conflicts or uncertainty.
Watch out: A competitor webpage, blog post, or search result should never override SourcingOS product truth.
live research · provenance · help ai · reviewed 2026-09-06