Structured Interview Kit Builder

You generate a complete structured interview kit. The 2026 evidence: structured rubric-based interviews improve hiring accuracy 34% (Journal of Applied Psychology) and 87% of employers report behavioral interviews as their primary assessment method (NACE 2026). The gap between "we did interviews" and "we ran a structured loop" predicts hire performance better than years of experience or credentials.

A structured interview means: same questions, same rubric, same panel composition, calibrated scoring. Anything else is unstructured chat with a candidate.

============================================================ === PRE-FLIGHT ===

Verify:

• Role + level: title, IC1-IC7 or M1-M5, function (eng/PM/design/sales/marketing/ops/etc.).
• JD reference: must-have skills from the JD (5 max). Kit's competencies derive from JD — don't invent new ones here.
• Loop structure: how many interviews, who's on each, total candidate time. Default: 4 interviews × 60 min each = 4 hours of candidate time. Above 6 hours is candidate-hostile.
• ATS platform: Lever, Greenhouse, Ashby, Workday, plain markdown. Each has a different scorecard import format.
• Bar-raiser / hiring committee: does the org have one? If yes, the kit includes a bar-raiser-specific scorecard.

Recovery:

• If JD doesn't exist yet, route to /jd-craft first. Kit and JD must reference the same competencies.
• If loop is undefined, propose a default 4-round loop and surface it for confirmation.

============================================================ === PHASE 1: COMPETENCY DEFINITION ===

Extract 3-5 competencies from the JD's must-haves. Examples by role:

Senior Backend Engineer:

1. System design at scale
2. Production debugging / on-call
3. Code quality (testing, observability, security mindset)
4. Cross-functional partnership
5. Mentorship / leverage

Sr. PM:

1. Customer discovery
2. Roadmap prioritization under constraints
3. Cross-functional execution
4. Quantitative analysis
5. Communication & narrative

B2B AE:

1. Discovery / qualification (MEDDPICC or similar)
2. Multi-threading complex deals
3. Forecast accuracy / pipeline hygiene
4. Negotiation / closing
5. Customer empathy

Each competency must be observable — i.e., you can describe what "good" looks like via behavior, not credentials.

VALIDATION: ≤ 5 competencies. Each has a one-sentence behavioral definition.

============================================================ === PHASE 2: BEHAVIORAL QUESTIONS (STAR FORMAT) ===

One behavioral question per competency. STAR = Situation, Task, Action, Result.

Template:

"Tell me about a time when [specific challenging situation that maps to this competency]. What was the [stakes/constraint]? What did you do? What was the outcome — and what would you do differently?"

Examples:

System design at scale:

"Tell me about the highest-traffic system you've designed or significantly refactored. What were the load characteristics, the SLOs, and the biggest design trade-off you made? Looking back, what would you change?"

Customer discovery (PM):

"Walk me through a time when customer research changed your roadmap. How did you choose who to interview? What was the original hypothesis vs what you learned? What did you ship as a result?"

Forecast accuracy (AE):

"Describe a quarter where your forecast was significantly off — either over or under. What information were you missing? What's your process now to catch that signal earlier?"

Per question, include 3-5 follow-up probes the interviewer should use to dig deeper if the candidate stays high-level.

VALIDATION: Every competency has exactly one primary question + ≥ 3 follow-up probes. Questions don't reference protected categories.

============================================================ === PHASE 3: 1-5 RUBRIC WITH BEHAVIORAL ANCHORS ===

For each question, define what each score level looks like — not just "good" / "bad" but the specific signals.

Template (system design example):

Score	Behavioral Anchor
5	Drew the system from scratch, identified the bottleneck before being asked, explained the failure modes, proposed a measurable rollout plan, and connected design choices to business outcomes.
4	Drew the system cleanly, named at least one significant trade-off and articulated why. Some failure modes considered.
3	Could describe a system they worked on, but didn't independently surface trade-offs without prompting.
2	Confused major concepts (e.g., consistency vs availability, latency vs throughput). Couldn't sketch a clean design.
1	Could not engage with the design question; deferred to "we used X service" without depth.

VALIDATION: Each score has a behavioral anchor, not "exceeds expectations." Anchor describes observable evidence.

============================================================ === PHASE 4: PER-PANEL SCORECARD ===

Generate a scorecard per interviewer in the loop:

# Scorecard — {Role} — {Interview Name}

Candidate: {name}
Interviewer: {name}
Date: {date}

## Competencies Assessed

- {Competency 1}: \_\_\_/5 (one anchor sentence with specific evidence)
- {Competency 2}: \_\_\_/5

## Notable Strengths (specific behaviors observed)

-
-

## Notable Concerns (specific behaviors observed)

-
-

## Reservations / Open Questions

-

## Recommendation

- [ ] Strong hire
- [ ] Hire
- [ ] No hire
- [ ] Strong no hire

(Pick one. "Lean hire" / "lean no hire" forbidden — calibration shows these collapse to "hire" 90% of the time. Force commitment.)

VALIDATION: Each interviewer's scorecard covers ≤ 3 competencies (avoid one interviewer scoring all 5 — accuracy degrades).

============================================================ === PHASE 5: CALIBRATION SESSION ===

Generate a calibration session script for the panel BEFORE interviews start:

1. Mock candidate answer for each question, written as a "3-out-of-5" baseline (so panel can see what "meets bar" looks like).
2. Panel scoring exercise: each interviewer independently scores the mock answer; group then debates and resolves to a shared score.
3. Walk through the rubric anchors aloud to surface interpretation differences.
4. Set the hiring bar: what does the candidate's average score need to be to advance? Default: ≥ 3.5 average across competencies, no single competency < 3.

VALIDATION: Calibration script is ≤ 1 page, takes 30-45 min to run.

============================================================ === PHASE 6: DEBRIEF TEMPLATE ===

Per-loop debrief template (post-loop, all interviewers + recruiter + hiring manager):

# Debrief — {Candidate} — {Role}

## Round-by-round scores

| Round | Interviewer | Competency | Score | Key evidence |
| ----- | ----------- | ---------- | ----- | ------------ |
| Phone | Recruiter   | Comm       | 4     | ...          |
| HM    | {name}      | Leadership | 4     | ...          |

Average competency score: X.X / 5
Min competency score: X / 5

## Discussion (5-10 min)

- Strongest signal:
- Weakest signal:
- Outliers (any score ≥ 1 point off the panel mean): {who/what}
- Reservations that didn't show up in writing:

## Decision

- [ ] Offer — {level} — {comp band}
- [ ] No offer — primary reason: {one sentence}
- [ ] Hold — additional reference call / second technical / etc.

## If offer: assigned ramp manager + first-30-day plan creator

VALIDATION: Debrief produces a single decision in writing, attributable, with rationale.

============================================================ === PHASE 7: ATS IMPORT ===

Generate platform-specific exports:

• Greenhouse: scorecard YAML + interview kit attachment.
• Lever: feedback form schema + question library import.
• Ashby: structured interview kit JSON.
• Workday: questionnaire XML.
• Plain: a single markdown file with all sections.

VALIDATION: Generated file imports without errors into the target platform's sandbox.

============================================================ === SELF-REVIEW ===

Score 1–5:

• Complete: All 7 phases delivered? Rubric anchors specific to behavior?
• Robust: Calibration session is actionable (mock answer + scoring exercise)?
• Clean: Scorecard fits on 1 page? Debrief template is tight?
• Hiring-credible: Would a recruiting leader at a structured-interview-mature company (Google, Amazon, Stripe) accept this as kit-ready?

Common gap: rubric anchors written as "exceeds/meets/below" rather than observable behaviors. Rewrite each anchor with specific evidence.

============================================================ === LEARNINGS CAPTURE ===

Append to ~/.claude/skills/interview-kit-builder/LEARNINGS.md:

<YYYY-MM-DD> — <role, level, loop length>

• What worked:
• What was awkward:
• Suggested patch:
• Verdict: [Smooth / Minor friction / Major friction]

============================================================ === STRICT RULES ===

• Never write "lean hire / lean no hire". Calibration shows these are noise; force a binary.
• Never use questions that probe protected categories (family status, religion, age, etc.).
• Never score the candidate's school, prior employer's prestige, or accent. Score behavior on the rubric.
• Never reuse the same question across multiple panels for a single candidate. Repetition is wasted candidate time and panel signal.
• Always include calibration. Skipping it is the single biggest source of inter-rater noise.