/interview-guide

Purpose

/interview-guide

Design thesis (one sentence). Structured interviews raise predictive validity AND reduce bias simultaneously; unstructured interviews have near-zero predictive validity and actively produce bias. This skill operationalizes structured interviewing, which is the single highest-leverage bias reducer in hiring that a product organization can install without inventing new methodology (Schmidt & Hunter 1998 meta-analysis; Levashina et al. 2014 structured interview review; EEOC structured-interview technical assistance — all public findings).

What it IS: a structured first pass that produces an interview guide a hiring manager, recruiter, and panel can review, calibrate on, and use, with every evaluation dimension grounded in the specific JD's essential functions, and every legally risky question flagged before the panel ever sees it.

What it is NOT: an Automated Employment Decision Tool (AEDT) under NYC Local Law 144 (it generates a guide, not a candidate score), a candidate scorecard, a resume screener, a reference-check template, a compensation discussion script, or a substitute for employment counsel on jurisdiction-specific forbidden-question edge cases.

This skill is governed by

hr-ai-governance

When to Use

Invoke

/interview-guide

• Open a brand-new role's interview process and need a JD-derived guide that starts from structured questions rather than a lifted template
• Calibrate an interview panel before a loop (distribute questions so three panel members don't all ask the same behavioral question three times)
• Audit an existing interview process for bias (proxy-laden rapport questions, illegal questions, missing rubrics, hypothetical-heavy questioning)
• Expand an established process into a new jurisdiction (salary-history ban, Ban-the-Box, Illinois Video Interview Act) and need the forbidden-question register re-checked
• Refresh a guide because the JD materially changed (e.g.,

  /job-description-generator

  updated the essential functions)

When NOT to Use

Do NOT use

/interview-guide

• You need to score, rank, or compare candidates — out of scope; scoring is individual-reviewer discretion at the structured debrief, not a skill output
• You need a technical skill assessment (coding exercise, case study, take-home work sample) — out of scope; see a dedicated technical-assessment skill (future)
• You need a reference check template — different workflow; reference checks have their own legal gates
• You need candidate communication templates (rejection, offer, interview scheduling) — out of scope
• You need a behavioral assessment of existing employees for promotion, discipline, or PIP — different workflow with different legal exposure; route to @performance-specialist
• You need compensation conversation framing — see

  /comp-benchmark

• You need a legal opinion on whether a specific question is enforceable under a specific state's employment law — out of scope; engage

  @employment-counsel

The skill deliberately sits DOWNSTREAM of

/job-description-generator

Modes

Create (default)

Generate a fresh interview guide from a JD, jurisdiction, and panel roles.

/interview-guide --jd Legionis/Product/jd-customer-success-lead.md --jurisdiction US-DE --panel-roles "hiring-manager,peer,cross-functional,hr"
/interview-guide --jd jds/senior-backend-engineer.md --jurisdiction US-IL --panel-roles "hiring-manager,peer,skip-level,hr" --video

Update

Refresh an existing guide. Pass the path to the current guide file.

/interview-guide update Legionis/Product/interview-guide-customer-success-lead.md --jurisdiction US-NYC

Update mode re-runs the forbidden-question register (jurisdiction-specific), re-runs the proxy-laden rapport scanner, re-checks panel-role distribution balance, and produces a diff-aware findings list (new findings marked

[NEW]

~~strikethrough~~

Check

Scan an externally-drafted interview guide (one a hiring manager wrote on their own) without regenerating it. Produces findings only, no rewrite.

/interview-guide check path/to/manager-draft-guide.md --jurisdiction US-CA

Required Inputs

The skill MUST collect the following before producing output. If any required input is missing, it asks the user rather than guessing.

Legionis/Product/jd-customer-success-lead.md

/job-description-generator

US-DE

US-NYC

US-CA

US-CO

US-WA

US-IL

US-MA

EU-DE

UK

IL

hiring-manager

peer

cross-functional

skip-level

hr

--video

US-IL

Legionis/Product/interview-guide-customer-success-lead.md

Jurisdiction is never defaulted. If the user does not supply it, the skill asks. The jurisdiction determines which forbidden-question rules apply (salary history ban, Ban-the-Box, Illinois Video Interview Act, Colorado AI Act, EU AI Act, etc.), the applicable framework under

hr-ai-governance

The JD is never summarized or inferred. The skill reads the JD and derives competencies from the essential functions. If the JD has no essential functions, the skill refuses to proceed and asks the user to run

/job-description-generator

Output Structure

Every

/interview-guide

sensitive-skill-guardrails.md

1. Disclaimer + UPL Guardrail Block (top)

> ⚠️ **Not HR or employment-law advice.** This output is a drafting and triage aid generated by a product-organization skill, not HR counsel or employment counsel. No attorney-client relationship is created by its production or use. Jurisdiction-specific questions on forbidden-question rules, pre-conditional-offer limitations, video-interview consent and analysis, pregnancy / disability / national-origin protections, or enforceability require review by a licensed employment attorney in the relevant jurisdiction and by a qualified HR professional. Do not rely on this output as the sole basis for interviewing, evaluating, or hiring any candidate.
>
> **Jurisdiction Assumed:** {jurisdiction from the required input}. If your jurisdiction differs, treat every finding below as a hypothesis to verify with local counsel.

2. Guide Metadata block

A small labeled block under the disclaimer with:

• Role: role name (from the JD)
• Source JD: path to the JD file the guide was derived from
• Jurisdiction: as input (no default)
• Panel Roles: as input (canonical set)
• Video Interview: yes/no (if yes, Illinois Video Interview Act addendum is generated when IL applies)
• Pack Inheritance:

  hr-ai-governance v{version}

  (read at runtime)
• Generation Timestamp: ISO 8601 UTC
• Governance Pack Version:

  hr-ai-governance v1.0.1+

• Behavioral : Hypothetical Ratio: ratio as generated (target ≥ 3:1)

3. Competency List (from the JD)

A numbered list of the competencies the guide evaluates, each mapped back to a specific essential function in the JD. The skill does NOT invent competencies. If the JD has 6 essential functions, the guide has 4-8 competencies (some essential functions collapse into one competency; some single functions produce two competencies — e.g., "Run QBRs" produces both "executive-communication" and "account-strategy"). The mapping is explicit.

4. Question Bank (per competency)

For each competency, a block containing:

• Competency name and the JD essential function(s) it maps to
• 2-4 behavioral questions — each phrased as "Tell me about a time when you..." grounded in a specific observable activity
• 0-1 hypothetical question — only if the behavioral questions alone cannot probe a specific situation the role requires and the candidate would not plausibly have encountered. If present, the rationale for the hypothetical is stated inline.
• Rubric for each question with three anchors:
  • Strong answer: specific situation, specific candidate action, specific outcome with numbers or named artifacts
  • Adequate answer: specific situation and candidate action but vague outcome, OR specific action+outcome but situation is normalized
  • Weak answer: abstract, second-person ("we did X"), no specific situation, or visibly rehearsed
• Standardized follow-up pattern: the probe sequence the interviewer uses regardless of the candidate's answer (e.g., "What was your specific contribution?" → "How did you measure the outcome?" → "What would you do differently?")
• Anti-pattern list: answers that sound good but don't indicate the competency (e.g., "I led a team transformation" without specifics is the classic anti-pattern; it pattern-matches as strong but contains zero evaluable signal)

5. Panel Assignment Matrix

A table showing which questions belong to which panel role, with a distribution check. Panel role boundaries follow a standard division of labor:

The distribution check: no single panel role asks more than 40% of the total questions in the guide. If one role exceeds 40%, the skill flags a P1

panel-distribution-imbalance

6. Calibration Protocol

A three-stage protocol, non-negotiable:

1. Pre-interview rubric share — all panel members receive the rubric BEFORE the interview. Not after. Not during. Reviewers who see the rubric only after they interview are cargo-cult structured interviewing; their scores are retrospectively-rationalized first impressions.
2. Independent scoring — panel members score their own questions individually and record their scores in a private field BEFORE any debrief discussion. Groupthink is the single largest threat to structured-interview validity; the first panel member to speak anchors everyone else. Independent scoring breaks the anchor.
3. Structured debrief — a standard template for aggregating scores, flagging divergence (cases where two panel members scored the same candidate >2 anchor steps apart on the same competency, indicating either a rubric ambiguity or a bias signal), and recording the hire / no-hire / additional-interview decision with explicit rationale.

The protocol also records the debrief template fields the skill emits: per-competency aggregated score, divergence flag, overall hire/no-hire, rationale, dissenting opinions (named).

7. Forbidden Question Register (standard baseline + jurisdiction-specific)

A two-part list. The standard federal baseline applies everywhere in the US. The jurisdiction-specific additions are added on top, per Section 10 below.

Standard federal baseline (never legal in any US jurisdiction):

• Age / date of birth / year of school graduation
• National origin / citizenship (beyond work authorization confirmation)
• Religion / religious practices / dietary preferences
• Marital status / spouse / family plans / pregnancy / intent to have children
• Disability (any form) pre-conditional-offer — ADA
• Sexual orientation / gender identity
• Political affiliation / union membership — NLRA Section 7 (strict liability; any surfacing is direct evidence of discrimination)
• Housing / homeownership / commute details that proxy class
• Genetic information / family medical history — GINA (strict liability)

Each entry has a brief one-line rationale and a correct alternative where one exists. For example: "age / date of birth" → no correct alternative; do not ask. "Citizenship" → the only legal question is "Are you authorized to work in {country} without employer sponsorship, now and in the foreseeable future?"

8. Proxy-Laden Rapport Scanner Findings

Rapport-building questions are the single most common vector for well-intentioned bias leakage. A hiring manager who asks "where did you grow up" is not trying to discriminate — but the question elicits national-origin, age, and socioeconomic-class signals the hiring manager cannot un-see once given.

The scanner runs over every question in the generated guide (including any examples in the calibration protocol) and flags proxy-laden patterns. Output table:

Categories the scanner catches:

• "where did you grow up"

  /

  "where are you from"

  /

  "what's your background"

  → national origin, age, socioeconomic class

• "tell me about your family"

  /

  "are you married"

  /

  "do you have kids"

  → family status, gender

• "what do you do on weekends"

  /

  "what are your hobbies"

  (open-ended) → religion, disability, family status (this one is subtle; it's flagged P2 with a suggestion to scope to work-adjacent)

• "where did you go to school"

  as rapport (not as qualification) → age, socioeconomic class, national origin

• "what do your parents do"

  /

  "how did you get into this field"

  (if phrased to elicit family background) → socioeconomic class

• "tell me about yourself"

  (without structure) → open door to protected-class disclosure the candidate may make; the skill does not ban this but flags it P2 with a suggested structure ("Walk me through your last three years professionally")

9. Video-Interview Addendum (if

--video

flag or jurisdiction requires)

If the video-interview flag is set OR the jurisdiction is

US-IL

• Candidate disclosure: the exact notice text the candidate must receive before the interview, per Illinois Video Interview Act (820 ILCS 42/). Disclosure must cover: the use of video, any AI-assisted analysis, the candidate's right to refuse video.
• Candidate consent: how consent is recorded. Verbal is not sufficient under IL; written (electronic) is required.
• Analysis limits: what AI-assisted analysis is and is NOT permitted. Facial expression analysis is prohibited under some readings of IL law and is categorically flagged as P0 regardless of jurisdiction.
• Deletion timelines: IL requires deletion within 30 days of candidate request. The addendum records the deletion-request channel.
• Accessibility: video interviews must have an accessible alternative (phone or in-person) available on request, per ADA and EU AI Act Annex III accessibility obligations.

If video is flagged but jurisdiction is NOT

US-IL

10. Jurisdiction-Specific Additions

Per the jurisdiction loaded from

hr-ai-governance

• US-NYC / US-CA / US-MA / US-CO / US-WA / US-CT / US-HI / US-RI / US-MD / US-DC and others: salary history ban — the skill adds "What is your current/prior compensation?" to the forbidden list and suggests the legal alternative: "What compensation range are you targeting for this role?"
• Ban-the-Box states (37+ US states + NYC Fair Chance Act + numerous municipal ordinances): criminal history questions are forbidden pre-conditional-offer. The skill adds a conditional-offer gate: criminal history questions may only appear in a post-offer flow, which is a different skill entirely.
• US-IL (Illinois): Illinois Video Interview Act requirements (see Section 9). Also Illinois HB 3773 IHRA amendment (effective Jan 1, 2026) — if the interview process includes any AI-assisted scoring, analysis, or decision support, the AI tool itself is subject to HB 3773 obligations; this skill generates the guide, not the scoring system, so it notes the requirement and routes AI-scoring questions to @employment-counsel.
• US-CO (Colorado): Colorado AI Act (SB 24-205, effective Feb 1, 2026) — if AI scoring is used downstream, pre-use consequential-decision notice is required. This skill notes the obligation and routes to @employment-counsel.
• EU / UK: GDPR Article 22 (automated decision-making) and EU AI Act Annex III point 4 (recruitment / selection) — high-risk system obligations attach to any scoring automation. Also, questions about pregnancy and family plans are forbidden under the Pregnant Workers Directive and UK Equality Act 2010 regardless of the federal US baseline.
• IL (Israel): national-origin-adjacent questions (country of origin, religion, IDF service details) are flagged per Israeli Equal Employment Opportunity Law; IDF service may only be asked as "military service: yes/no" without country, unit, or duration, mirroring

  hr-ai-governance

  Section 3.2 proxy register.
• All US: PDA (Pregnancy Discrimination Act) and ADA forbid pregnancy, family, and disability questions. These are already in the standard baseline but the jurisdiction-additions section restates them for completeness.

11.

## Findings

(numbered)

Every flagged item from Steps 5-10 above gets a numbered finding in a unified findings list:

• Finding # / Category (proxy-laden-rapport, forbidden-question, hypothetical-without-justification, panel-distribution-imbalance, missing-rubric, missing-calibration, video-interview-compliance-gap, jurisdiction-specific)
• What — specific question, section, or structural gap
• Why it matters — the risk, implication, or bias mechanism
• Severity — P0 / P1 / P2
• Suggested next step — address / accept-with-risk / reject-as-hypothetical
• Citation — pointer to

  hr-ai-governance

  section, EEOC guidance, Schmidt & Hunter 1998, Gaucher et al., or jurisdiction statute

12.

## Reviewer Checklist

Ten mandatory sign-off items before the guide is acted upon:

• Jurisdiction confirmed against role location and candidate residency
• JD verified as the source and reviewed for competency derivation accuracy
• Every question mapped to a competency and a rubric
• Behavioral-to-hypothetical ratio ≥ 3:1 (justify exceptions inline)
• Panel distribution reviewed; no single role > 40% of questions
• Forbidden question register reviewed against jurisdiction and candidate profile
• Proxy-laden rapport scanner findings addressed or explicitly accepted
• Calibration protocol confirmed: pre-interview rubric share scheduled, independent scoring mechanism in place, debrief template loaded
• Video-interview addendum reviewed (if applicable) including consent + analysis-limits + deletion timeline
• Employment counsel engaged for items in "Cannot Assess Without" that apply to this role or jurisdiction

13.

## Cannot Assess Without Licensed Counsel or Specialist

Minimum 5 items:

• Actual candidate interaction — the guide pre-calibrates the rubric, but individual candidate answers require human judgment the skill cannot substitute for
• Technical skill validation — coding exercises, case studies, take-home work samples, architecture reviews — all out of scope; the guide evaluates behaviors, not technical proficiency
• Reference checks — different workflow with different legal exposure (consent gates, jurisdiction-specific defamation law)
• Culture-add judgment — the guide surfaces observable values-fit questions, but the hiring panel's collective judgment on culture-add is individual and not a skill output
• Candidate-specific follow-ups — deeper probes triggered by something the candidate said are context-dependent and require interviewer discretion
• Jurisdictional edge cases — any forbidden-question call that turns on a specific statute interpretation (e.g., whether a particular "acceptable use of criminal history" exception applies in a post-conditional-offer context in NYC) requires employment counsel
• AI-assisted candidate scoring — if the deployer is using any AI tool to score, rank, or recommend candidates, that tool is subject to its own governance (

  hr-ai-governance

  Section 6 AEDT wall); this skill does not generate scoring systems

14. Audit Log Entry

Emitted per

hr-ai-governance

• run_id

  ,

  timestamp

  ,

  skill_name

  (

  interview-guide

  ),

  skill_version

  ,

  governance_pack_version

• jurisdiction

• inputs_hash

  — SHA-256 of the canonicalized input object (JD path + hash, jurisdiction, panel roles, video flag). NEVER raw candidate data — this skill doesn't see candidate data at guide-generation time.

• input_type: interview_notes

  (closest available enum value; the real type is

  interview_guide

  , recorded in

  notes

  )

• proxies_detected

  — from the proxy-laden rapport scanner, each flagged question becomes a proxies_detected entry

• redactions_applied

  — N/A for this skill (this is a generation skill, not an extraction skill)

• hitl_reviewer

  — authenticated identity of the hiring manager or recruiter who will review the guide before the panel sees it

• hitl_decision

  ,

  hitl_timestamp

  ,

  hitl_rationale

  — populated at HITL gate

• downstream_action

  — e.g.,

  published-to-panel

  ,

  held-for-revision

  ,

  routed-to-counsel

• signoff_path: standard

• retention_expiry

  — computed from jurisdiction per Section 4.3

The skill refuses to run if the deployer's auth system cannot supply an authenticated

hitl_reviewer

Method

The skill's generation flow is a 7-step pipeline. Order matters because later steps depend on earlier ones.

Step 1 — Derive Competency List from JD Essential Functions

Read the JD (path required as input). Extract the essential functions list. For each essential function, derive one or more competencies. A competency is a named observable capability, not an abstract trait:

• Good competency:

  executive-stakeholder-communication

  (derived from "Run quarterly business reviews with top-decile accounts")
• Bad competency:

  strong-communication

  (abstract; does not pass the rubric-anchorable test)

The skill produces 4-8 competencies total. If the JD has fewer than 3 essential functions, the skill refuses and routes to

/job-description-generator

Step 2 — Generate Behavioral + Hypothetical Questions per Competency

For each competency, the skill generates:

• 2-4 behavioral questions — phrased as "Tell me about a time when you..." followed by a specific observable activity the competency implies. Behavioral questions have higher predictive validity than hypothetical per the structured-interview literature (Schmidt & Hunter 1998 meta-analysis put structured behavioral interviews at r ≈ 0.51 vs. r ≈ 0.14 for unstructured interviews).
• 0-1 hypothetical question — only if the competency requires probing a scenario the candidate is statistically unlikely to have encountered (e.g., "This role is the first of its kind — how would you approach the first 90 days?"). If no such scenario exists, the competency gets zero hypotheticals. A hypothetical without a justification line is automatically downgraded or removed.

The generated behavioral-to-hypothetical ratio must be ≥ 3:1 overall. If it falls below, the skill flags it P1 and asks the user to justify the lower ratio (typical justification: the role has no direct precedent in any candidate's background).

Step 3 — Build Rubric per Question (Strong / Adequate / Weak Anchors)

For each question, the skill writes three rubric anchors grounded in the expected answer shape:

• Strong answer: specific situation, specific candidate action, specific outcome with numbers or named artifacts
• Adequate answer: specific situation and candidate action but vague outcome, OR specific action+outcome but situation is normalized or hypothetical
• Weak answer: abstract, second-person ("we did X"), no specific situation, or visibly rehearsed (uses phrases like "in general I tend to...")

The rubric does NOT include the interviewer's "gut feel" — that is precisely what structured interviewing is designed to eliminate. If a rubric anchor cannot be stated without hedging ("this feels like a strong answer if..."), the question is not anchorable and the skill either rewrites the question or removes it.

Step 4 — Assign Questions to Panel Roles

Using the panel roles the user supplied, the skill assigns each question to the role most appropriate for owning it:

• Hiring manager owns role-critical competencies (the 1-2 most central to the JD)
• Peer owns collaboration and day-to-day working behaviors
• Cross-functional owns stakeholder management across domains
• Skip-level owns strategic alignment and judgment under ambiguity (only if the role warrants a skip-level slot)
• HR owns values fit (observable), candidate questions, and compliance-neutral items

The skill runs the 40% distribution check. If any single role exceeds 40% of the total questions, the skill flags the imbalance and suggests redistribution. If the user's supplied panel role list is small (e.g., just hiring-manager + peer), the 40% check is relaxed proportionally — with two roles, the threshold rises to 60% for either role, because a two-person panel cannot distribute evenly below that.

Step 5 — Apply Forbidden-Question Register per Jurisdiction

Pull the standard federal baseline. Layer the jurisdiction-specific additions from

hr-ai-governance

1. Scans the generated questions for any forbidden pattern (e.g., "Are you a US citizen?" → citizenship question → forbidden baseline)
2. Flags any found forbidden questions as P0

   forbidden-question

   findings
3. Surfaces the legal alternative where one exists (e.g., "Are you authorized to work in {country} without employer sponsorship?")
4. For jurisdictions with salary-history ban: scans for "What is your current compensation?" and equivalents; surfaces "What compensation range are you targeting?" as the legal alternative
5. For Ban-the-Box jurisdictions: scans for criminal-history questions; if found, moves them to a separate post-conditional-offer flow and notes the gate
6. Emits the full forbidden-question register in the output so the panel sees both the forbidden list AND the jurisdiction layer

The register is informational output — its value is forcing the panel to see, in writing, which questions cannot be asked. Panels that don't see the register in advance tend to ask the forbidden questions anyway, usually during the rapport-building minutes.

Step 6 — Scan Questions for Proxy-Laden Language (inherit

hr-ai-governance

Section 3.2)

Pull the proxy register from

hr-ai-governance

• Name of place ("where did you grow up") → national origin + age + SES
• Family terms ("tell me about your family") → family status + gender
• Open-ended hobbies → religion + disability + family status
• School-as-identity ("where did you go to school" used as rapport not qualification) → age + SES + national origin
• Parents' professions → SES + national origin
• Open self-description ("tell me about yourself" without structure) → open door to protected-class disclosure

Each flagged question gets a finding row (Section 8 output). The scanner is deliberately over-inclusive: it flags well-intentioned questions because the whole point of the scanner is to catch questions the panel would never self-censor. A P2 flag for "tell me about yourself" is not an accusation of bias — it is a nudge to structure the opener.

Step 7 — Generate Calibration + Scoring + Debrief Structure

Assemble the protocol (Section 6 output). The skill:

1. Generates the pre-interview rubric share package (a one-page summary per panel member showing their questions + rubrics + anti-patterns)
2. Specifies the independent-scoring mechanism (scoring field per question, recorded before any panel discussion)
3. Generates the debrief template (per-competency aggregated score, divergence flag, overall decision, rationale, dissenting opinions field)
4. Adds the HITL gate hook — the guide cannot move from "draft" to "published-to-panel" without a recruiter or hiring manager sign-off per

   hr-ai-governance

   Section 10

Quality Gates

The skill performs a 10-item self-check BEFORE emitting output. If any check fails, the output does not publish — it produces a structural finding instead, asking the user to fix the gap.

1. Every question has a named competency mapping back to a specific JD essential function
2. Every question has a strong/adequate/weak rubric
3. Behavioral-to-hypothetical ratio ≥ 3:1 (or justified exception inline)
4. Panel distribution: no single role asks > 40% of total questions (or > 60% in a 2-role panel, proportional for ≥ 3 roles)
5. Forbidden-question register applied per jurisdiction — zero forbidden questions in the final draft
6. Proxy-laden rapport scanner run — findings emitted
7. Calibration protocol includes pre-interview rubric share (not just a post-interview debrief)
8. Independent-scoring step exists in the calibration protocol
9. Structured debrief template present with divergence flag
10. Video-interview addendum present if

    US-IL

    jurisdiction or

    --video

    flag

These 10 checks are the quality gates. They run every invocation. There is no "express mode" that skips them.

Pack Inheritance

This skill inherits the following from

hr-ai-governance

inputs_hash

jurisdiction

hitl_reviewer

signoff_path

retention_expiry

/resume-summarizer

This skill is NOT an AEDT. It produces a guide, not a candidate decision. The AEDT wall from

hr-ai-governance

Delegation Patterns Available

Default: Pattern 1 Consultation

When a specific element of the guide needs specialist input, the skill spawns a consultation per

delegation-protocol.md

Consultations are attributed in the findings section: "I consulted 👔 Employment Counsel, who noted that in {jurisdiction}, pre-conditional-offer criminal history questions are prohibited by {statute} and the safe alternative is to move the question to the post-offer flow." Ownership of the interview guide stays with Recruiter.

Adversarial Review

NOT applicable at this version. Adversarial Review (Pattern 5) is reserved for near-final deliverables with high-stakes, uncapped exposure (enterprise contracts, M&A documents, pricing commitments). An interview guide draft going to a human hiring panel is neither near-final nor uncapped-exposure — it is a draft that a human reviewer revises and the panel then uses with individual discretion.

If a specific guide becomes high-stakes (e.g., a senior executive search with public press attention, a role in a regulated industry with additional forbidden-question statutes, or a role in a jurisdiction with active litigation risk), the hiring manager can request

/interview-guide --escalate-adversarial

@employment-counsel

ROI Framing

ROI for

/interview-guide

HR blended rate: $150/hr per

feedback_roi_rates.md

Time-saved baseline: a careful, structured, bias-scanned interview guide that starts from the JD, generates rubric-anchored behavioral questions, assigns them across the panel, runs the forbidden-question register, and produces a calibration protocol is ~3-4 hours of manual drafting and triage time for a recruiter working with a hiring manager. Simpler refreshes (Update mode on an existing guide after a JD tweak) are ~1-1.5 hours baseline. Complex cases (multi-jurisdictional posting, regulated industry, video-interview requirements, executive search) are 5+ hours baseline.

The ROI tracks ONLY the time the skill saves on drafting the structured artifact, not the substantive HR or legal review time. The panel calibration, the actual interviews, and the debrief all happen in full and are owned by humans.

Example ROI line for a standard guide generation:

⏱️ ~3 hrs saved on drafting and triage in 55s, 17k tkns ~$1.0 cost, Value ~$450

Attribution and Maintenance

Owner: 🎯 Recruiter. The skill's drafting, scanner, and panel-assignment logic is Recruiter's accountability.

Consumers (skills / gateways that invoke this skill):

• ext-hr

  — HR team gateway, primary user

New consumers require a frontmatter update and a one-line note in the consuming gateway's or skill's dependency list.

Authoring: First-principles. This skill was authored from scratch during Phase 4A as the third HR skill under the

hr-ai-governance

/comp-benchmark

/job-description-generator

• Schmidt, F. L., & Hunter, J. E. (1998). "The Validity and Utility of Selection Methods in Personnel Psychology: Practical and Theoretical Implications of 85 Years of Research Findings." Psychological Bulletin, 124(2), 262-274. Meta-analysis establishing structured-interview validity at r ≈ 0.51 vs. unstructured at r ≈ 0.14.
• Levashina, J., Hartwell, C. J., Morgeson, F. P., & Campion, M. A. (2014). "The Structured Employment Interview: Narrative and Quantitative Review of the Research Literature." Personnel Psychology, 67(1), 241-293. Confirmation and extension of Schmidt & Hunter with 2000s-era meta-analytic data.
• EEOC structured-interview technical assistance — public guidance on structured interviewing as a bias-reduction mechanism under Title VII and ADEA. Last verified: 2026-04-11.
• Gaucher, Friesen, and Kay (2011) — gendered-language research previously cited by

  /job-description-generator

  ; informs the rubric-language neutrality check.
• Illinois Artificial Intelligence Video Interview Act (820 ILCS 42/) — statute referenced in the video-interview addendum.
• EEOC Uniform Guidelines on Employee Selection Procedures (1978) — adverse-impact baseline; applies to any selection procedure including structured interviews.

No content was lifted from vendor interview-guide tools (Gem, Greenhouse, Lever, Ashby, SmartRecruiters, Textio, etc.). The question banks, rubric anchors, and panel-assignment logic are derived first-principles from the JD input plus the structured-interview literature.

Dependency on the pack: The skill reads from

hr-ai-governance

Updates: Via the two-pass publication gate defined in

sensitive-skill-guardrails.md

/resume-summarizer

/job-description-generator

Minor edits (typos, formatting, proxy-list additions sourced from the pack's Section 3.2 updates) can bypass Pass 2. Any edit touching: severity thresholds, forbidden-question register structure, rubric-anchor taxonomy, panel-distribution threshold (40% / 60%), calibration protocol requirements, or the HITL gate placement — requires a full Pass 2 substantive review by Employment Counsel.

Changelog: Maintained at the bottom of this file.

Example Invocation

User: /interview-guide --jd Legionis/Product/jd-customer-success-lead.md --jurisdiction US-DE --panel-roles "hiring-manager,peer,cross-functional,hr"

/interview-guide v1.0.0 — loading:
  - JD: Legionis/Product/jd-customer-success-lead.md (6 essential functions; source: /job-description-generator output)
  - Jurisdiction: US-DE (no salary-history ban, no Ban-the-Box complication, standard federal baseline)
  - Panel roles: hiring-manager, peer, cross-functional, hr
  - Video flag: no
  - Governance pack: hr-ai-governance v1.0.1

Running pipeline:
  Step 1: Competency derivation (6 essential functions → 5 competencies)
  Step 2: Question generation (14 behavioral + 2 hypothetical, ratio 7:1)
  Step 3: Rubric construction (16 questions × 3 anchors = 48 rubric anchors)
  Step 4: Panel assignment (hiring-manager 5, peer 4, cross-functional 4, hr 3 — distribution within 40% threshold)
  Step 5: Forbidden-question register applied (standard federal baseline only for US-DE)
  Step 6: Proxy-laden rapport scanner (3 planted proxy questions in test input; all 3 caught)
  Step 7: Calibration + scoring + debrief structure generated

Producing output at:
  Legionis/Product/interview-guide-birth-test-2026-04-11.md

10/10 quality gates passed.

⏱️ ~3 hrs saved on drafting and triage in 55s, 17k tkns ~$1.0 cost, Value ~$450

Changelog

• 1.0.0 (2026-04-11) — Initial authoring. First-principles during Phase 4A as the third HR skill under the

  hr-ai-governance

  pack (after

  /comp-benchmark

  and

  /job-description-generator

  ). Authored by 🎯 Recruiter. Methodology grounded in Schmidt & Hunter 1998 meta-analysis, Levashina et al. 2014 structured-interview review, and EEOC structured-interview technical assistance (all public). Question-bank taxonomy, rubric-anchor pattern, panel-assignment boundaries, forbidden-question register, and proxy-laden rapport scanner derived first-principles. Subsequent-similar 72-hour Pass 2 review under Employment Counsel SLA. Scaffolding review by 📋 Director of HR. Birth-tested against a Legionis Customer Success Lead role (see

  Legionis/Product/interview-guide-birth-test-2026-04-11.md

  ).