hr — independently scanned and version-tracked by SaferSkills.
SaferSkills independently audited hr (Agent Skill) and scored it 96/100 (green). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 0 high-severity and 1 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.
Findings & checks · 1 flagged
The text {match} tells the agent to skip the normal "ask the user first" gate. Used adversarially it removes the human-in-the-loop check before destructive or sensitive actions, turning a normally-gated agent into a fire-and-forget executor.
Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
Umbrella skill for all people operations: recruiting, performance management, compensation, offer drafting, interview design, onboarding, org planning, people analytics, and policy lookup. Each mode loads its own references on demand.
Scope: Decisions and artifacts involving people's careers, compensation, and organizational structure. Use csuite for strategic business decisions, data-analysis for general analytics, professional-communication for non-HR business writing.
Classify the request into exactly one mode. If the request spans modes, choose the primary and note the secondary.
| Mode | Signal Phrases | References |
|---|---|---|
| RECRUITING | Pipeline, candidates, sourcing, screening, hiring status, time to fill | references/recruiting.md |
| PERFORMANCE | Review, self-assessment, calibration, feedback, rating, promotion case | references/performance-management.md |
| COMPENSATION | Pay, salary, equity, comp bands, benchmarking, offer competitive, retention risk | references/compensation.md |
| OFFER | Draft offer, offer letter, comp package, signing bonus, start date | references/compensation.md |
| INTERVIEW | Interview plan, questions, scorecard, evaluation rubric, debrief | references/recruiting.md |
| ONBOARDING | New hire, first week, 30/60/90, onboarding checklist, buddy | references/recruiting.md |
| ORG-PLANNING | Headcount, reorg, team structure, span of control, org design | references/org-planning.md |
| PEOPLE-ANALYTICS | Attrition, headcount report, diversity metrics, org health, flight risk | references/org-planning.md |
| POLICY | PTO, benefits, leave, expenses, handbook, remote work policy | (no reference — use user-provided policy docs) |
Always load: references/llm-hr-failure-modes.md — applies to every mode.
HR content touches people's careers, livelihoods, and legal rights. These guardrails are non-negotiable.
| Rule | Rationale |
|---|---|
| Source compensation data from user-provided or public databases | Invented market rates cause real pay decisions. State "I don't have current market data" when you don't. |
| Focus recommendations on skills, behaviors, outcomes | "Hire more [group]" or "this candidate fits the culture" introduces bias. Focus on skills, behaviors, outcomes. |
| Include legal review disclaimer on all binding language | Offer letters, policy interpretations, and termination language need legal review. Always state this. |
| Ask for jurisdiction before advising on compliance | Employment law varies by country, state, city. Ask for jurisdiction before advising on compliance, leave, or termination. |
| Minimize PII retention; use role/level identifiers | Names, salaries, SSNs, demographics — minimize retention. Use role/level when names aren't needed. |
| Flag when output needs legal review | Offer letters, PIPs, termination docs, policy changes, accommodation decisions — always flag. |
Framework: DEFINE → PIPELINE → EVALUATE
Phase 1: DEFINE — Clarify role requirements and pipeline structure.
references/llm-hr-failure-modes.md for biased language (because gendered/exclusionary language reduces qualified applicant pools by 10-40%)Gate: Role defined. Pipeline stages agreed. Posting reviewed for bias.
Phase 2: PIPELINE — Track and manage candidates through stages.
Gate: Pipeline metrics current. Bottlenecks identified.
Phase 3: EVALUATE — Structure interviews and decisions.
Gate: Interview kit complete. Debrief structured. Decision evidence-based.
Onboarding sub-mode (after offer acceptance):
Framework: STRUCTURE → WRITE → CALIBRATE
Phase 1: STRUCTURE — Select review type and load template.
| Type | Use When |
|---|---|
| Self-assessment | Employee writing their own review |
| Manager review | Manager writing review for direct report |
| Calibration prep | Preparing rating distributions for calibration meeting |
Gate: Review type selected. Template loaded.
Phase 2: WRITE — Generate review content with behavioral specificity.
Self-assessment:
Manager review:
Constraint: Describe observable behavior with specific examples ("documentation was incomplete on 3 of 5 deliverables") — because personality feedback triggers defensiveness and has no actionable path. See references/llm-hr-failure-modes.md.
Gate: Review content complete. All feedback behavior-based. Development areas actionable.
Phase 3: CALIBRATE — Prepare rating distribution and promotion cases.
Constraint: Present rating targets as guidelines, with flexibility for team context (because forced ranking creates perverse incentives and has been abandoned by most organizations).
Gate: Distribution documented. Promotion cases evidence-based. Compensation actions justified.
Framework: BENCHMARK → ANALYZE → RECOMMEND
Phase 1: BENCHMARK — Establish market data context.
Gate: Components identified. Variables specified. Data sources declared.
Phase 2: ANALYZE — Score against market and internal equity.
Constraint: Always state data vintage and source limitations. "Based on 2024 Levels.fyi data" not "the market rate is $X" — because stale data presented as current causes underpayment or overpayment.
Gate: Analysis complete. Sources cited. Limitations stated.
Phase 3: RECOMMEND — Deliver actionable compensation recommendations.
Gate: Recommendations prioritized. Budget impact calculated. Sources documented.
Offer drafting sub-mode:
Framework: ASSESS → MODEL → PLAN
Phase 1: ASSESS — Map current organizational state.
| Metric | Healthy Range | Warning Sign |
|---|---|---|
| Span of control | 5-8 direct reports | <3 (too narrow) or >12 (too wide) |
| Management layers | 4-6 per 500 people | Excess layers = slow decisions |
| IC-to-manager ratio | 6:1 to 10:1 | <4:1 = top-heavy |
| Team size | 5-9 people | <4 = fragile, >12 = unmanageable |
| Single points of failure | 0 | Any = structural risk |
Gate: Current state mapped. Structural issues identified.
Phase 2: MODEL — Design target state and transition.
Constraint: Never recommend org changes based on individuals ("move Alice because she's difficult") — structure around roles and capabilities (because person-dependent org design creates fragility and masks management problems).
Gate: Target state modeled. Sequencing justified. Budget estimated.
Phase 3: PLAN — Convert to executable hiring roadmap.
Gate: Roadmap executable. Risks mitigated. Success metrics defined.
People analytics sub-mode:
Framework: FIND → EXPLAIN → CAVEAT
Constraint: Answer only from user-provided policy documents. State when no source is available — do not guess (because fabricated policy guidance creates liability).
| Error | Cause | Solution |
|---|---|---|
| No compensation data | User asks "what should we pay" without data | State limitation explicitly. Offer to analyze user-provided data or recommend public sources (Levels.fyi, Glassdoor, Radford). |
| Biased language in output | LLM generates gendered, ageist, or exclusionary phrasing | Run output against references/llm-hr-failure-modes.md bias checklist. Rewrite flagged phrases. |
| Jurisdiction unknown | Legal advice requested without location | Ask for jurisdiction before proceeding. Never default to US employment law. |
| Confidentiality scope unclear | User shares individual comp/performance data | Confirm intended audience. Remind that HR data is need-to-know. Minimize PII in outputs. |
| Template vs. real data confusion | User treats template placeholders as recommendations | Label all templates explicitly: "[PLACEHOLDER — replace with actual data]". |
| Policy not found | User asks about policy with no handbook provided | State clearly: no policy source available. Do not fabricate. Suggest uploading handbook. |
| Reference | When to Load | Content |
|---|---|---|
references/recruiting.md | RECRUITING, INTERVIEW, ONBOARDING modes | Pipeline stages, velocity metrics, interview frameworks, evaluation rubrics, onboarding checklists |
references/performance-management.md | PERFORMANCE mode | Review structure, calibration methodology, feedback patterns, development planning |
references/compensation.md | COMPENSATION, OFFER modes | Market benchmarking, internal equity analysis, offer structuring, equity modeling |
references/org-planning.md | ORG-PLANNING, PEOPLE-ANALYTICS modes | Headcount modeling, org design principles, capacity planning, people analytics |
references/llm-hr-failure-modes.md | Every mode | Bias detection, fabrication risks, compliance gaps, inappropriate language patterns |
~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.