Interview Debrief

---
name: metaview-interview-debrief
description: >
  Run an evidence-based debrief for a candidate (or a whole final-stage slate) from Metaview.
  Trigger when the user asks to debrief, decide, or compare — "pull together a debrief for
  Amara", "what evidence do we have on Tom across the rounds", "compare our three finalists for
  the PM role", "prep the debrief meeting", "did the panel actually agree on Sofia?". Builds a
  competency-by-competency evidence matrix from every interview, surfaces where interviewers
  agree and contradict each other, checks scorecards against what transcripts actually show,
  and produces a decision-ready pack — including how to position the offer if it's a yes.
---

# Metaview Interview Debrief

Debriefs run on memory and volume; this skill makes them run on evidence. It reconstructs what every interviewer actually learned about a candidate — from transcripts, Notes, and scorecards — and lays it out so the hiring team debates the evidence, not their recollections.

**Who it's for:** hiring managers and recruiters running debriefs or hiring decisions.
**Metaview products needed:** AI Notetaker. Scorecards need a connected ATS. Works best when all rounds were recorded.

## Why this is powerful

Debriefs are decided by whoever speaks most confidently about a call nobody else attended — and scorecards get written hours later from memory. This levels the room: every claim traceable to a transcript moment, disagreements surfaced *before* the meeting so the panel debates the two conflicting quotes instead of talking past each other, and the untested competencies named before someone mistakes "not assessed" for "no concerns". Debriefs move from recall to evidence, without adding a step to your process.

## Operating rules

- Tool names may carry a connector prefix in your environment; match on the trailing name. Never surface internal UUIDs — use names and conversation URLs.
- Every cell in the evidence matrix must trace to a specific conversation. No evidence = say "no evidence", never infer.
- You are producing decision *support*, not a decision. Present the evidence for and against; do not offer a read on the candidate — lay out the evidence and let the panel form the view.
- Read-only by default. The only writes are optional follow-ups the user explicitly asks for (e.g. generating a debrief note in Metaview).

## Workflow

### Step 1 — Context and scope

Call `get_user_context` (workspace, data access — if the user can't see colleagues' conversations, the debrief may be missing rounds; flag it).

Clarify scope if ambiguous: one candidate, or a slate to compare? Which role? Against what bar — a rubric/JD the user can paste, competencies from the role intake call, or (fallback) the competencies the interviews themselves appear designed to test?

### Step 2 — Collect the candidate's full record

For each candidate in scope:

1. Resolve them: `list_field_values` with `field_id="default:candidate"`, `search_term="<full name>"` (fall back to email — it's the most reliable identifier). Disambiguate duplicates with the user.
2. All conversations: `search_conversations` filtered by `{"field_id": "default:candidate", "operation": "includes_one_of", "value": ["<participant-id>"]}`. Note dates, event titles (stage names), interviewers.
3. Depth strategy: ≤3 interviews → pull `default:transcript` for all; 4-8 → `default:summary` for all + transcripts for the rounds where the decision actually hangs (typically the most senior/most recent); 9+ → summaries plus targeted transcript reads. A 1-hour transcript is ~10-15K tokens.
4. `fetch_candidates` (participant ID or email) → profile + ATS scorecards, application stage, interviewer feedback.

Skip unrecorded/ghost sessions (no transcript) — count them in coverage but don't cite them.

### Step 3 — Establish the competency frame

Use, in order of preference: (1) the rubric/JD the user provided, (2) the role's intake call if one exists (search conversations of type "Role Intake" for the role, or by event title), (3) a frame you derive from what interviewers consistently probed — and say that's what you did.

Keep it to 4-8 competencies plus "motivation & logistics".

### Step 4 — Build the evidence matrix

For each competency × each round, extract: what was actually tested, the strongest evidence for, the strongest evidence against, and the interviewer's stated conclusion where they gave one. Then compute the three things debriefs miss:

- **Coverage gaps** — competencies nobody tested. These are unknowns, not neutral scores.
- **Divergence** — rounds that reached different conclusions on the same competency. Quote both sides.
- **Scorecard vs transcript** — where a scorecard's rating isn't supported (or is contradicted) by anything said in that round's transcript, flag it explicitly. Treat a missing scorecard as missing — note the round's signal exists only in the transcript.

### Step 5 — Produce the debrief pack

```
# Debrief — <Candidate> · <Role> · <N> interviews over <period>

**Coverage:** <N> of <M> rounds recorded · competencies tested <X>/<Y> · scorecards submitted <A>/<B>

## The shape of the signal
3-5 bullets: where the panel converges, where it splits, and the biggest untested area.

## Evidence by competency
Per competency: verdict-neutral summary, evidence for (round, quote/paraphrase), evidence
against (round, quote/paraphrase), who assessed it. Link each round's Metaview URL.

## Contradictions to resolve in the room
Each one: the two conflicting reads, verbatim-ish, and a suggested question that would resolve it.

## Unknowns
What no round tested + a suggested probe if there's another touchpoint before decision.

## Candidate's position
Their motivation, concerns, other processes, comp expectations, timeline — as stated, with rounds.

## If it's a yes: offer positioning
What they're moving toward / away from (weighted by how often they repeated it), what to
emphasise in the offer conversation, risks to closing.

## If it's a no
The evidence the team can point to when giving feedback — review and edit before it reaches the candidate.
```

For a **slate comparison**, add a final section: candidates side-by-side per competency (evidence-based, no scores invented), plus "different strengths" framing — never collapse this into a single ranking; the panel decides how to weigh the evidence.

### Step 6 — Optional follow-ups (only on request)

- **Debrief note in Metaview** for the whole hiring team: `generate_notes` on the candidate's most recent interview with `template="generic_debrief"` (or `team_role_debrief`/`technical_role_debrief`), then `manage_notes_sources` with `action="add"` and the candidate's other conversation IDs so the note covers every round. Adding sources triggers regeneration — wait ~30s, then confirm via `search_conversations` on that conversation with `fields=["default:summary"]`. You can also add the JD or rubric as a document source.
- Export the pack as a document, or draft the debrief-meeting agenda from the contradictions list.

## Edge cases

- **Only one interview exists:** say so — this is a screen review, not a debrief; offer the prep-skill instead for the next round.
- **Panel interviewed together (one conversation, many interviewers):** attribute evidence to the conversation, and to named interviewers only when the transcript speaker labels make it unambiguous.
- **Interviews in multiple languages:** work in the language of each transcript; write the pack in the user's language.
- **Sensitive content:** comp, health, visa, protected characteristics — include only what's decision-relevant (comp, visa/right-to-work, notice) and never speculate beyond what was said.

## Example invocations

- "Build the debrief pack for Jai Singhal — panel meets tomorrow at 10."
- "Compare our three Engagement Manager finalists against the intake call."
- "The scorecards say strong yes but I'm nervous — what does the transcript evidence actually show for Dana?"