Interviewer Coaching

---
name: metaview-interviewer-coaching
description: >
  Turn interview recordings into interviewer coaching — for yourself, or aggregate quality
  signal for a team. Trigger on "how am I doing as an interviewer", "coach me on my screens",
  "give me feedback on yesterday's interview", "review my call against the screening guide",
  "how consistent are our interviewers", "talk-time by interviewer", "build interviewer
  training from our best interviewer's calls", or any ask about interview quality, coverage,
  question technique, or candidate experience. Reviews transcripts against a coaching
  framework (the user's own guide, or the bundled Screening Call Guide), quotes the actual
  moments, and delivers private coaching packs, team-level patterns, and training guides.
---

# Metaview Interviewer Coaching

Interviewing is a skill almost nobody gets feedback on — the only people watching are candidates. This skill puts the tape to work: it reviews real transcripts against a concrete benchmark, quotes the exact moments that worked and the follow-ups that got missed, and turns patterns across calls into two or three things actually worth changing.

## Why this is powerful

Every interview is already recorded and transcribed — the coaching evidence exists, it's just unread. A human coach would need to shadow ten calls to see a pattern in your follow-ups; this surfaces the ten transcript moments in minutes, privately. It's the difference between "be a better interviewer" and "here are the three follow-ups you didn't ask last week, and what to say next time."

## Ethics and governance — read first

- **Individual coaching is private.** Output goes to the person being coached, framed constructively, marked "for your eyes only". Lead with strengths.
- **Team views aggregate.** Patterns and distributions, anonymised examples. Never rank named interviewers, never build a league table; describe outliers abstractly and let the leader follow up humanely.
- **Assess the interviewer, not the candidate.** A weak candidate doesn't make a thorough interviewer bad; a strong candidate doesn't make an unstructured one good. Only the interviewer's questions, framing, listening, and selling count.
- Some organisations restrict interviewer-performance analysis (works councils, legal). If the user hints at this, suggest confirming internally before team-level analysis.

## The framework

Coaching needs a benchmark, so establish it first — ask the user, in order:

1. **Their own guide** — "Do you have an interview guide, rubric, or scorecard this round is supposed to follow? Paste or attach it." Use it as the primary frame if given.
2. **The bundled default** — otherwise coach against `references/screening-call-guide.md` (read it before reviewing any transcript). Six areas: **Rapport, Probing, Talk Time, Candidate Motivations, Selling the Opportunity** — plus **Housekeeping as a completeness check only** (one covered/missed line; never a headline development area unless wholesale skipped). Tell the user which frame you're using.

The evaluation rules travel with the framework regardless of source: be specific (quote the moment), be constructive (strengths first), benchmark against good/poor examples, adjust for context (a junior-role quick screen won't hit everything), patterns > one-offs, flag interviewer talk share above ~40%, and check motivations were *used* in the sell, not just collected.

For non-screening rounds (technical, HM, panel), keep Probing/Talk Time/Rapport, swap Motivations+Selling for the round's assessment goals (from the user's guide or the round's scorecard), and say you've adapted the frame.

## Operating rules

- Tool names may carry a connector prefix; match on the trailing name. No internal UUIDs in output; link conversations by URL.
- Every piece of feedback cites a specific moment from a specific call. Feedback without evidence is vibes.
- Scale: 1-5 calls → read transcripts in full; 6-20 → summaries + targeted transcripts; 20+ → AI fields (borrow the calibration discipline from metaview-interview-intelligence).

## Workflow A — Coach one interviewer (usually the requester)

1. **Setup** (ask, don't assume): whose calls (resolve via `list_field_values` on `default:interviewer` — internal people, one record each); which calls (default: their screens from the last 30 days — filter `default:interviewer` `includes_one_of`, `default:call_duration` greater_than ~10 min); and the framework (above). 5-15 calls is the sweet spot; with 1-2, coach per-call and say the sample is thin for patterns.
2. **Objective layer** — `group_conversations`/`get_chart_data` metrics: talk-time (`aggregation:talk_time`, `metric_arguments {"scope":"internal"}`; flag >40%), questions per call (`aggregation:question_count`, scope internal), duration vs expected, scorecard submitted + submission lag. Pull the team-wide median in one extra query so every number has context ("you vs team median" — never vs named colleagues).
3. **Craft layer** — read the calls against the framework. For each area, describe what happened with 1-2 verbatim moments as evidence, and note where the framework's guidance wasn't followed. Hunt specifically for: the 2-3 moments a great follow-up was missed (write the follow-up they should have asked), whether motivations were referenced later in the sell, and one thing they do well that colleagues could copy.
4. **Deliver the pack:**

```
# Interview coaching — <name> · <n> calls, <period> · for your eyes only

## What's working (specific, quoted)
2-3 strengths with the transcript moment and why it works.

## The numbers
Talk-time, questions/call, scorecard completion — vs team median, one line each.

## Areas to develop (max 3 — highest leverage only)
Each: the pattern → 1-2 quoted examples → the follow-up/phrasing to use instead →
one concrete thing to try in the next call.

## Missed opportunities
Moments to dig deeper or sell harder, with the line that was left on the table.

## Area check
Rapport · Probing · Talk time · Motivations · Selling — one line each.
Housekeeping: covered ✅/missed ❌ item list, single line, no lecture.

## Keep this
One sentence to pin above the desk.
```

5. **Offer the recurring version**: after each interview (or weekly), review new calls against the same framework and DM the pack **to the person being coached, and only them** — confirm the recipient before scheduling. Track "since last time": which development area improved, what's new. Keep each recurring note under ~300 words.

## Workflow B — Team interview quality (for a lead)

1. Scope the population (team/stage/window) and the framework with the user, plus the aggregation ethics reminder.
2. Objective spread via `group_conversations` grouped by interviewer: report distributions, not rankings ("talk-time ranges 35-75%, median 48%; three of nine screens run >40%").
3. At 20+ calls, build interviewer-behaviour AI fields (confirm first; calibrate on 8-12 transcripts before trusting — the field's prompt must carry the "assess the interviewer, not the candidate" inversion explicitly). Proven shapes: framework areas covered / not covered (a list, not a score) · behavioural-question count (framing and follow-up probes don't count) · missed-probing flag · motivations-used-in-sell boolean · next-steps-set boolean.
4. Report: patterns, consistency findings ("the same role is screened three different ways"), 2-3 anonymised strong/weak moments, and recommended interventions (guide update, calibration session, one shared template). Dashboards render as artifacts/inline — never code files. Offer `create_report` (confirm first) if they want it live in Metaview.

## Workflow C — Training guide from an interviewer your team learns from

1. The user names the exemplar. Confirm the exemplar knows and agrees before pulling their calls — celebration, not surveillance. Then pull their last 8-12 relevant calls.
2. Extract the playbook: how they open, question archetypes (verbatim), how they probe and push back, how they use silence, how they weave motivations into the sell, how they close.
3. Deliver a training guide: their call structure as a repeatable template, 10-15 actual questions with what each tests, 3-5 annotated "listen to how this was handled" moments with conversation links, and a do/don't list grounded in the calls — mapped to the same framework areas so it doubles as the team's benchmark.

## Edge cases

- **Panel interviews:** attribute behaviour per speaker only when labels are unambiguous; otherwise coach the panel collectively.
- **"Who's our worst interviewer?"** — decline the ranking; offer the distribution plus private coaching for anyone who wants it.
- **Non-admins** may only see their own calls — perfect for Workflow A; Workflow B needs broader access (check `data_access` from `get_user_context` and say so).
- **The user disputes a finding:** show the transcript moment and let the tape settle it; if the context genuinely justifies it, drop the finding.

## Example invocations

- "Coach me on my last five screens — use our interview guide, attached."
- "Review yesterday's call with Maria against the screening framework."
- "How consistent are our HM interviews for the AE role this quarter?"
- "Build a training guide from Priya's last ten finals — she's our best closer."
- "Every Friday, send me a private coaching note on that week's screens."