Call Scorer

Interview Call Scorer

What This Skill Does

Reads candidate interview/screening calls (transcripts + notes from Granola via MCP) and produces a structured score and written feedback per candidate, mapped to the hiring rubric. The output tells the user: how the candidate scored on each dimension, with evidence quoted or paraphrased from the actual call, and a clear advance / hold / pass recommendation.

This turns a pile of recorded interviews into a ranked, evidence-backed shortlist.

---

Prerequisites & Limitations (Read First)

This skill depends on Granola MCP. Know these constraints before relying on it:

1. Transcript access requires a Granola Business plan or higher. On the free/Basic tier, the MCP

only exposes the last 30 days of notes and no transcripts. For real scoring you want transcripts — flag this to the user if scoring seems thin.

1. **Granola MCP only sees the authorizing user's private "My notes" space — never Team-space

folders.** If interviews live in a shared Team space, each interviewer must connect their own Granola, or scoring will come back empty. Surface this if a search returns nothing.

1. A rubric must exist. Score against an explicit rubric (use the rubric-builder skill). Never

invent scoring criteria silently — that just launders the scorer's bias.

If Granola isn't connected, say so and suggest enabling it. Don't fabricate call content.

---

Process

1. Confirm the rubric. Ask which rubric to score against, or build one with rubric-builder. The

5 default dimensions: Quality of Questions Asked, Evidence of Slope, Self-Awareness, Culture Fit & Values, Real Work Output.

1. Pull the call(s) from Granola. Use the MCP to find the relevant meeting(s) by candidate name,

date, or title. Prefer the natural-language query tool for open-ended retrieval; pull the full transcript when available for direct evidence.

1. Score each dimension 1–5, anchored to the rubric. For every score, cite specific evidence —

what the candidate actually said or did. No evidence = lower confidence, say so.

1. Compute the weighted total using the rubric's weights.
2. Check hard gates — any dimension where a 1 is an automatic no.
3. Write the recommendation: advance / hold / pass, with the reasoning in one tight paragraph.
4. If scoring multiple candidates, produce a ranked comparison table at the end.

---

How to Score Well

• Evidence over impression. Every score points at something concrete from the call. "Scored 5 on

Quality of Questions — asked how the company thinks about retention vs. acquisition, then pushed on the growth model" beats "great questions."

• Watch the question quality signal closely. What the candidate asked is often the cleanest

read on how they think. Pull their questions out explicitly.

• Distinguish slope from pedigree. Reward evidence of fast self-driven learning over a glossy

résumé. Flag candidates who only replay a prior company's playbook.

• Score culture fit separately and against the company's real values — never fold it into general

competence.

• Be honest about missing signal. If the call never surfaced work output, don't guess — note the

gap and recommend a work trial or take-home.

• Don't inflate. A 3 is "solid, would do the job." Reserve 5 for genuinely exceptional. If

everyone's a 5, the rubric is useless.

---

Output Format

ALWAYS use this structure (per candidate):

# Candidate Scorecard: [Name]
**Role:** [title] · **Call:** [date / type] · **Source:** Granola

| Dimension | Weight | Score (1–5) | Evidence |
|---|---|---|---|
| Quality of Questions | 15% | X | [what they asked / didn't] |
| Evidence of Slope | 20% | X | [learning signal from the call] |
| Self-Awareness | 15% | X | [how they handled pushback] |
| Culture Fit & Values | 15% | X | [values evidence] |
| Real Work Output | 35% | X | [work signal, or note the gap] |
| **Weighted Total** | 100% | **X.X / 5** | |

**Hard gates:** [pass / triggered — which]
**Recommendation:** [Advance / Hold / Pass]
**Why:** [one tight paragraph — the decisive factors]
**What to probe next round:** [gaps the call didn't resolve]

For multiple candidates, end with:

## Ranked Shortlist
| Rank | Candidate | Score | Recommendation | One-line read |
|---|---|---|---|---|

---

Scorer Failure Checklist

A scorecard FAILS if it:

• ❌ Scores without citing evidence from the actual call
• ❌ Invents criteria instead of scoring against the explicit rubric
• ❌ Inflates scores (everyone's a 4–5, no discrimination)
• ❌ Folds culture fit into general competence
• ❌ Rewards pedigree over demonstrated slope
• ❌ Guesses at dimensions the call never surfaced instead of flagging the gap
• ❌ Gives a recommendation that doesn't follow from the scores and gates

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: getclera
• Source: getclera/claude-skills
• License: MIT

Install and usage instructions live in the source repository linked above.