Interview Feedback Aggregator

What this skill does

Combines feedback notes from multiple interviewers for one candidate and produces a single packet: per-competency rollup, themed strengths, themed concerns, dissent calls, and a recommendation aligned with a configurable rubric.

Inputs

• feedback_dir: a directory of files, one per interviewer, named <interviewer>-<round>.md (or a single CSV with interviewer, round, score, notes).
• rubric_path: a YAML file defining competencies and weights.
• Optional decision_threshold: weighted score above which hire is recommended (default 3.5/5).

Steps

1. Parse each feedback file. Each interviewer should provide a per-competency rating (1-5) and qualitative notes.
2. Validate against the rubric: competency names in feedback must match those in rubric_path. Surface any mismatches as schema errors.
3. Build a per-competency table: rows are interviewers, columns are competencies, cells are scores. Compute median per competency.
4. Compute weighted overall score: sum(median(comp) * weight(comp)).
5. Extract strengths: notes containing words matching a positive-signal lexicon (great, strong, clearly, excellent, independently).
6. Extract concerns: similarly for negative-signal lexicon (struggled, unclear, had to push, gap in, concern).
7. Detect dissent: a competency where the spread between the highest and lowest score is >= 2; tag explicitly with each interviewer's score.
8. Infer recommendation: weighted_score >= decision_threshold -> lean-hire; if median across all competencies >= 4 -> strong-hire; with any 1 score -> lean-no-hire.
9. Compose the packet markdown with sections: Recommendation, Per-Competency Scores, Strengths, Concerns, Dissents, Verbatim Notes (paginated by interviewer).
10. Save the packet.

Output

packet-<candidate-id>.md with the named sections. Plus a JSON sidecar with the structured scores for ATS import. Stdout prints the recommendation and any dissents.

Verification

Re-read the per-competency table and confirm the weighted score formula matches the rubric. Spot-check one strength and one concern: each must trace back to a verbatim quote in the source feedback (avoid hallucinated synthesis). For each dissent, ensure the report lists the actual interviewers and scores, not a paraphrase.

Edge cases

• Interviewer who didn't provide ratings, only notes: include qualitative input, mark scores as "no rating", and skip them in median computation.
• Single interviewer (rare): no dissent section; flag the small sample at the top.
• Feedback referencing protected characteristics (age, family, etc.): redact and surface a top-of-packet warning so hiring managers see it before bias creeps in.
• Late-arriving feedback: the packet should be regeneratable; never lock immutably until the loop is closed.