cold-call-coach

name	cold-call-coach
description	Pull real cold-call recordings from your SmrtPhone account, transcribe them with real tonality notes (an audio model hears the calls), and grade every conversation against the DataSift cold-calling rubric - opener, motivation probing (4 pillars), objection handling, tonality, close quality. Produces per-call coaching reports, per-caller scorecards, and one clean Excel workbook. Use when asked to review cold calls, grade callers, run call coaching, or audit call quality.
version	2.0.0

Cold Call Coach

Grades real first-touch cold calls against the DataSift cold-calling playbook. The pipeline pulls recordings straight from your SmrtPhone web session (no API exists for this), transcribes them with an audio-capable model so tonality observations are heard rather than guessed, then Claude grades each gradeable conversation against references/rubric.md and exports everything to Excel.

Requirements and setup (once)

1. Python 3.10+ with: pip install playwright openpyxl then playwright install chromium
2. SmrtPhone login: python scripts/smrtphone_login.py (headed window opens, log in once; the session is saved to smrtphone_state.json and lasts weeks)
3. OpenRouter key for transcription: copy scripts/env.example to .env and set OPENROUTER_API_KEY (get one at openrouter.ai)
4. Optional team roster: copy scripts/roster.example.json to roster.json to pin caller roles (a dedicated cold caller's calls never route to another rubric) and exclude departed callers

Costs (transparent)

• Transcription: about $0.002 per audio minute (Gemini 2.5 Flash via OpenRouter). A week of coaching for a 3-caller team (say 60 gradeable calls averaging 2 minutes) costs roughly $0.25 in transcription.
• Recording downloads and the call log are free (your existing SmrtPhone plan).
• Grading runs inside your Claude session.

Pipeline

Run from your project folder:

python scripts/pull_calls.py --min-seconds 60 --days 7   # call log + MP3s
python scripts/transcribe.py                              # audio -> transcripts + triage

Outputs land in output/call_coaching/:

• call_log.json / calls_to_review.json - every call, then the review set
• transcripts/{call_id}.md - diarized transcript with inline delivery notes ([long pause 4s], [rushed], [warm tone]) plus a DELIVERY SUMMARY block (pace, agent energy, talk balance, notable audio moments) and a label check line
• review_queue.json - calls grouped: cold_call / lead_management / closing / not_gradeable

Only grade calls in the cold_call group with worth_grading: true. Voicemails, wrong numbers, and dead air are logged, never scored.

Grading procedure

1. Read references/rubric.md fully (applicability gate, short-call rule, auto-fails, 5 categories weighted 15/30/20/15/20, score anchors, both report templates, the required SCORES JSON footer).
2. Read the worked examples in references/calibration/ before your first session and stay consistent with them.
3. For each call: read the whole transcript once; verify AGENT/SELLER labels by content (our agent is whoever asks about buying the property); route full vs short (short = a DECLINE inside ~30 seconds; doubt resolves to full); score each reachable criterion against its 0/3/5 anchors (1, 2, 4 interpolate).
4. Write one report per call to output/call_coaching/reports/cold_call/{call_id}_{caller}.md using the rubric's templates exactly, including the CRITERION SCORES table and the SCORES JSON footer.
5. Roll up a per-caller scorecard, then export the workbook:

python scripts/export_excel.py

One styled Excel file: Summary (full calls, /100, band colored), Short Calls (opener + conversion on their own scale), Coaching Detail, Caller Scorecard, Criterion Scores.

Grading quality protocol (mandatory verify pass)

• Every quote must appear verbatim in the transcript (check by substring).
• Every criterion score cites a quote or an explicit absence.
• N/A only where the rubric allows; list every N/A and redistributed weight.
• Recompute the math: tables must foot exactly and the JSON footer must match the tables number for number.
• Grade ONLY what is audible on the call. CRM state and dispositions are never scored.
• Full calls and short calls are different scales; never rank them together.

Measured reliability

Independent graders reproduce full-call totals within about +/-3 points and agreed on the grade band in every test comparison; see references/reliability.md. Treat totals within 5 points as equivalent, and get a second grade before any band-boundary personnel decision.

Failure modes

• pull_calls.py exits 2 or reports an HTML response: session expired; re-run python scripts/smrtphone_login.py.
• transcribe.py key error: set OPENROUTER_API_KEY in .env.
• Recording URL 404: the recording was deleted in SmrtPhone; log and skip.
• Transcripts can truncate on very long calls; N/A the unreachable categories and note it.

Related

• Follow-up and qualification calls: the lead-manager-coach skill.
• Offer and negotiation calls: the closer-coach skill.
• This skill grades call QUALITY, not dial volume KPIs.