Best Claude Prompts for Recruiters in 2026

Recruiters in 2026 use Claude for twelve high-leverage moves across the funnel: rewriting JDs for inclusivity with must-haves vs nice-to-haves split out, building Boolean search strings, summarizing candidates from resume plus LinkedIn, personalizing outreach off a single mutual connection, generating interview-question banks by competency, drafting scorecard rubrics with bias-mitigation built in, writing stage-appropriate rejection emails, briefing the hiring manager for handoff, justifying an offer pull with market comp data, structuring post-interview debriefs, analyzing channel mix from ATS exports, and writing the end-of-quarter recap. The winners pair tight prompts with hard guardrails: NYC Local Law 144 AEDT treats LLMs that substantially assist or replace selection as regulated tools, and the EEOC's 2023 software guidance extends Title VII disparate-impact analysis to AI-assisted decisions. Use Claude to draft, not to decide.

Two reasons dominate: long-context candidate synthesis and writing tone. Per the Anthropic model overview, Opus 4.x ships a 200K-token context window — big enough to drop a full resume, LinkedIn data, three call transcripts, and a JD into one prompt. The 2025 LinkedIn Future of Recruiting report found 62% of TA leaders ranked sourcing the workflow most improved by generative AI; Greenhouse's 2026 hiring benchmark shows AI-personalized outreach lifts response rates 1.7x on warm-source candidates.

Tone control is the second pull. Most outreach fails because it sounds like a robot. Per Anthropic's prompt engineering guide, pinning role and voice early is the strongest lever on quality — and recruiter outreach is the place where one sentence of personalization beats five paragraphs of value claims.

Anthropic's prompt engineering docs recommend a four-move skeleton: assign a role, ground with XML-tagged context, give the task, constrain output shape. Recruiters layer a guardrail line on top:

``` You are <role>. <Context in XML tags>. <Task>. <Output format>. <Guardrail: do not score, rank, or recommend hire/no-hire decisions; flag any output that would constitute a selection decision under NYC Local Law 144 or EEOC guidance>. ```

NYC Local Law 144 defines an AEDT as a computational process that issues a simplified output (score, classification, recommendation) and substantially assists hiring decisions. A prompt asking for a 1-5 fit score crosses that line. A prompt asking for structured evidence for a human to score does not. Every prompt below is built for the second pattern.

**Block:**

``` You are a senior talent acquisition partner with expertise in inclusive job description writing. Read the JD inside the <jd> tags. Produce three outputs: 1. A rewritten JD that removes coded language flagged in the Textio 2024 language report and the Harvard Business Review research on gendered job descriptions. Replace masculine-coded terms (rockstar, ninja, aggressive, dominant, competitive) with neutral equivalents. 2. A split list: MUST-HAVES (max 5, each tied to a concrete on-the-job task) vs NICE-TO-HAVES (everything else). 3. A flag list: any requirement that, per EEOC guidance, may produce disparate impact (years-of-experience floors above 7, degree requirements not validated as job-related, citizenship language broader than ITAR). <jd> [paste current JD] </jd> Output as three labeled sections. Do not invent requirements. Do not change the seniority level. ```

**Why it works:** The 5-item must-have cap forces ranking. HBR's research on gendered wording shows JDs with masculine-coded language see 26% fewer female applicants. The disparate-impact flag is the bridge to legal — recruiter surfaces, does not adjudicate.

**Sample output excerpt:**

> MUST-HAVES: 1. Has shipped at least one production payments integration. 2. Has owned an on-call rotation for a 99.9% SLA system. 3. Production-quality Go or Kotlin. 4. Cross-functional partnership. 5. US work authorization. > > FLAGS: 'minimum 10 years of experience' — experience floors above 7 years correlate with age disparate impact; recommend 5 or remove.

**Block:**

``` You are a sourcing strategist who builds LinkedIn Recruiter, GitHub, and Google X-Ray Boolean strings. Given the role context and must-have list, produce three Boolean strings: 1. A LinkedIn Recruiter string using AND, OR, NOT and parenthetical grouping. Target current title plus past title plus skills. 2. A GitHub search string targeting repos and language filters. 3. A Google X-Ray string targeting site:linkedin.com/in with the must-have skills. For each string list (a) the must-have it maps to, (b) one expected false positive, (c) one variant string that broadens by one degree if the first returns under 100 results. Role context: <paste 5 bullets> Must-haves: <paste list from Prompt 1> ```

**Why it works:** LinkedIn Talent Insights shows targeted Boolean strings cut time-to-first-qualified-candidate by 31%. Forcing one expected false positive surfaces the search's blind spot before a sourcing shift is wasted on it.

**Block:**

``` You are a recruiter writing a 200-word candidate summary for the hiring manager. Read the resume and the public LinkedIn data inside the tags. Produce exactly these sections: 1. Current role and tenure (one sentence). 2. Three most relevant experiences for the must-haves below, each quote-anchored to a line in the resume. 3. Two open questions a recruiter screen should answer. 4. One reason this candidate might not be a fit (be honest). Do NOT produce a fit score, a rank, or a hire/no-hire recommendation. This is a structured summary for human review. <resume> [paste resume] </resume> <linkedin> [paste public LinkedIn data] </linkedin> Must-haves: <paste list> ```

**Why it works:** The 200-word cap forces ranking. Quote-anchoring kills hallucinated experience. The no-score guardrail keeps the output outside NYC Local Law 144's AEDT definition — the recruiter scores, the model just organizes evidence. Per Greenhouse's structured hiring data, structured summaries cut HM screen-rejection 24%.

**Block:**

``` You are a senior recruiter writing a first-touch outreach message to a passive candidate. The mutual connection between me and the candidate is below. Write three message variants: 1. LinkedIn InMail, 85 words max, subject line under 7 words. 2. Email, 110 words max, single CTA: 15-minute exploratory call. 3. WhatsApp / SMS, 40 words max, single question. Every variant must reference the mutual connection by name and one specific thing from the candidate's profile that is not their current title. Conversational tone. Do not say 'I came across your profile.' Do not say 'rockstar.' Do not promise outcomes (compensation, promotion, equity range). Candidate profile: <paste 5 bullets> Mutual connection: <name + one shared context, e.g., 'we worked together at Datadog 2019-2021'> Role: <one-sentence pitch> ```

**Why it works:** Per LinkedIn's 2025 Future of Recruiting, InMails referencing a mutual connection see 41% higher response rates. The 'do not promise outcomes' guardrail keeps recruiters clear of FTC truth-in-advertising territory on comp claims.

**Block:**

``` You are an interview design specialist trained on Lou Adler's Performance-Based Hiring methodology and the Greenhouse structured interviewing framework. Given the role's must-haves, build a question bank with 4 questions per competency, structured as: - Behavioral past-experience question (STAR-eligible) - Hypothetical scenario question - Skills demonstration / work sample prompt - Probe question to surface depth For each question include: (a) what good looks like in 2 sentences, (b) what a red flag answer sounds like, (c) one follow-up probe. Competencies: <list 4-6> Must-haves: <paste from Prompt 1> Level: <IC4 / IC5 / Manager / Director> Do not produce a numeric score. Produce evidence rubrics for human interviewers to score. ```

**Why it works:** Greenhouse's 2026 data shows structured interviews with predefined rubrics lift quality-of-hire 40% over unstructured panels. The 'good looks like / red flag' pair makes scoring consistent without Claude doing the scoring.

**Block:**

``` You are an industrial-organizational psychologist designing a structured interview scorecard. Given the must-haves and the competencies below, produce a scorecard with: 1. 5-7 evidence dimensions tied directly to on-the-job tasks (job-relatedness validated). 2. A 1-4 anchored rating scale per dimension with behaviorally anchored definitions (1 = no evidence, 2 = some evidence, 3 = strong evidence, 4 = exceptional evidence). 3. Calibration notes: a one-line example of what each rating sounds like. 4. Two bias-mitigation reminders at the top of the scorecard: (a) score evidence not impressions, (b) reference candidates from the underrepresented half of the slate first to anchor against the strongest, not against the rep majority. 5. A required free-text field per dimension capturing the evidence quote. Competencies: <list> Must-haves: <paste> Level: <IC4 / IC5 / Manager / Director> ```

**Why it works:** BARS are the SIOP-validated standard for structured interviewing. The 'underrepresented candidates first' anchor move is from Iris Bohnet's counter-stereotype research. The evidence-quote field is what makes the scorecard defensible under EEOC technical guidance — the human scored, and the evidence is traceable.

**Block:**

``` You are a recruiter writing rejection emails that protect candidate experience and brand. Write three variants for these stages: 1. Application rejection (no recruiter screen happened) — 60 words max, warm, do not say 'we will keep your resume on file' unless we will. 2. Post-recruiter-screen rejection — 90 words max, one specific reason tied to a must-have, invite to future roles if appropriate. 3. Post-onsite rejection — 130 words max, specific feedback (one strength + one growth area), referral to similar roles or partner companies, offer 10-minute debrief call if requested. No vague language. No false hope. Comply with GDPR Article 22 disclosure if the candidate is in the EU — flag if so. Candidate stage: <stage> Reason for rejection: <evidence tied to scorecard> EU candidate: <yes/no> ```

**Why it works:** Per LinkedIn's 2025 candidate-experience data, 94% of candidates who get specific post-interview feedback report a positive brand impression even after rejection. GDPR Recital 71 gives EU candidates the right to human review of automated rejections — acknowledging this closes the loop.

**Block:**

``` You are a senior recruiter preparing a one-page hiring-manager briefing for the lead onsite interview. Output in this exact structure: 1. Candidate snapshot (40 words). 2. Three reasons we advanced this candidate, each quote-anchored to a scorecard dimension. 3. Two open questions the panel still needs to answer. 4. One bias-watch note: if this candidate is from an underrepresented group on the panel, remind the HM to anchor against the strongest candidate not the modal candidate. 5. Logistics: time, panel, format, accommodations requested. Inputs: <screen-notes> [paste recruiter screen notes] </screen-notes> <scorecards> [paste prior stage scorecards] </scorecards> <role-context> [paste must-haves + level] </role-context> ```

**Why it works:** Per Greenhouse's structured hiring report, interviewers briefed with quote-anchored evidence score 33% more consistently across the panel. The bias-watch note is the human-in-the-loop counter to the EEOC's 2023 guidance on disparate-impact monitoring.

**Block:**

``` You are a comp partner building the offer-approval justification for a candidate the panel wants to pull. Output in this exact structure: 1. Candidate summary (3 lines). 2. Panel scorecards summary (one line per interviewer, no aggregate score). 3. Proposed offer: base, equity (RSU value at trailing 30-day VWAP), sign-on, target total comp. 4. Market comp band: cite Levels.fyi median + 75th + 90th percentile for this title and level at peer-group companies. Cite Pave or Option Impact data if available. Show the gap to median. 5. Justification: one paragraph tying scorecard evidence to the percentile being offered. 6. Risk: one paragraph on what a counter-offer at OldCo would look like and how we respond. Inputs: <scorecards> [paste all scorecards] </scorecards> <comp-data> [paste Levels.fyi screenshot or Pave export] </comp-data> <candidate-comp-history> [paste self-reported comp] </candidate-comp-history> ```

**Why it works:** Most TA teams write offer-pull memos badly. Levels.fyi's 2026 comp data and Pave's benchmarks are the field-standard references. The 'no aggregate score' guardrail keeps the model out of selection-decision territory while still pulling evidence together for the human approver.

**Block:**

``` You are an interview-process designer. Given the panel scorecards below, prepare a debrief discussion agenda for a 30-minute panel meeting. Output: 1. Agenda: 5-minute calibration, 15-minute evidence review by competency, 5-minute open questions, 5-minute decision discussion. 2. For each competency, surface (a) the dimension with the widest score variance across panelists, (b) the specific evidence each panelist cited, (c) one calibration question to ask the panel. 3. Decision framing: 'What evidence would change your mind?' — listed once per panelist. 4. Do not produce a hire/no-hire recommendation. The debrief surfaces evidence and disagreement; the panel decides. Inputs: <scorecards> [paste 4-5 scorecards] </scorecards> ```

**Why it works:** Score variance is the signal — when panelists disagree by 2+ points, that is the conversation worth having. Google's re:Work guide calls variance-driven debriefs the largest single lift in interview-decision quality. The 'no recommendation' guardrail is the AEDT-safety move.

**Block:**

``` You are a TA operations analyst. Given the ATS export below (last 4 quarters of applications, screens, onsites, offers, hires by source), produce a channel mix analysis: 1. Funnel conversion table by source: application -> screen, screen -> onsite, onsite -> offer, offer -> hire. 2. Cost per hire by source (estimate where not provided, label clearly). 3. Time-to-hire median by source. 4. The three sources I should double down on and the two I should cut, each with one-line justification. 5. One disparate-impact flag: if any source shows a >20% gap in offer-rate by self-disclosed demographic group, surface it for HR/legal review — do not act on the data, surface it. <ats-export> [paste CSV] </ats-export> ```

**Why it works:** Per Lever's 2026 TA benchmark, top-quartile teams kill 30%+ of sourcing channels each year on funnel-conversion data. The flag-not-act pattern is the EEOC-recommended four-fifths-rule check, surfaced for human investigation.

**Block:**

``` You are a head of talent acquisition writing the Q-end recap to the executive team. Output in exactly this structure, no marketing language, 500 words max total: 1. Headline number: hires closed vs plan. 2. Top three wins (named hires + role criticality). 3. Top three misses (open reqs + diagnosis). 4. Channel mix shift this quarter, with data. 5. Diversity-of-slate progress: % of pipelines with 2+ underrepresented finalists, trend vs prior quarter. 6. One systemic change we are making next quarter and the expected impact. 7. One ask of the executive team. Inputs: <ats-summary> [paste Q-end ATS summary] </ats-summary> <channel-analysis> [paste output of Prompt 11] </channel-analysis> <prior-quarter-recap> [paste last quarter's recap] </prior-quarter-recap> ```

**Why it works:** Execs read recaps in two minutes — the 500-word cap forces ranking. Per LinkedIn's Future of Recruiting, teams reporting diversity-of-slate (not just diversity-of-hire) lift underrepresented-finalist rate 1.4x year-over-year because slate is upstream of the panel's decision.

Funnel-stage mapping follows Greenhouse's structured hiring playbook and Lever's 2026 TA benchmark. Compliance mapping follows NYC Local Law 144 AEDT guidance and the EEOC's 2023 technical guidance on software, algorithms, and AI.

They ask Claude to score candidates. A prompt that says 'rate this candidate 1-10 for fit' produces exactly the simplified output NYC Local Law 144 flags as an AEDT — triggering published bias audit, candidate notice, and ongoing monitoring obligations the recruiter has not set up. Same trap with 'rank these 50 resumes by fit.' That is a classifier.

The fix is structural: ask Claude to organize evidence, surface gaps, and produce rubrics. Humans score, decide, and own the call. Per Anthropic's prompt engineering guide, explicit output formats plus explicit no-go lists are the highest-impact quality lever — in recruiting, the no-go list is what keeps you out of regulator headlines.

### Which Claude model should recruiters default to?

Sonnet 4.x covers 80% of the prompts above at a fraction of Opus cost. Use Opus for long-context candidate summaries (full resume + LinkedIn + 3 calls) and channel-mix CSV analysis. Cost difference is roughly 5x per million tokens per Anthropic's pricing.

### Does using Claude for candidate summaries trigger NYC Local Law 144?

Not if the summary is structured evidence for a human reviewer with no score, rank, or recommendation. It does trigger AEDT obligations if you prompt for a numeric fit score, rank-ordered shortlist, or hire/no-hire recommendation. See NYC DCWP AEDT guidance.

### Is it safe to paste resumes and LinkedIn data into Claude?

Depends on contract and candidate location. Anthropic's commercial terms exclude API and Claude for Work data from training the public model by default; consumer Claude.ai differs. For EU candidates, GDPR applies — run a DPIA before pasting at scale.

### What is the cheapest way to add bias-mitigation to a Claude recruiting workflow?

Three moves cost nothing. Structured scorecards with behaviorally anchored ratings (SIOP-validated). Required evidence quotes per dimension. Scoring underrepresented candidates first as anchors per Iris Bohnet's research. Claude just drafts the scaffolding.

### Can Claude replace the recruiter screen?

No. The recruiter screen is where a human verifies work authorization, gauges motivation, calibrates comp, and reads tone in a way no LLM can defend in audit. Claude preps the question bank and summarizes the resume; humans run the call.

### What is the fastest way to start if our TA team has never used Claude?

Run Prompt 1 (JD rewrite) on the three highest-volume open reqs this week. Add Prompt 4 (outreach personalization) on one passive-sourcing campaign next week. One prompt per week makes the habit stick and lets you measure response-rate lift without confounders.

### How do I version-control the recruiter prompt library so it does not drift?

Store every prompt in a shared workspace, version with a date, assign one named owner, and audit quarterly. Per Anthropic's release notes, Claude models update monthly; outputs drift if prompts do not.