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)
- Python 3.10+ with:
pip install playwright openpyxl then playwright install chromium
- SmrtPhone login:
python scripts/smrtphone_login.py (headed window opens,
log in once; the session is saved to smrtphone_state.json and lasts weeks)
- OpenRouter key for transcription: copy
scripts/env.example to .env and
set OPENROUTER_API_KEY (get one at openrouter.ai)
- 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 settranscripts/{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 linereview_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
- 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).
- Read the worked examples in
references/calibration/ before your first
session and stay consistent with them.
- 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).
- 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.
- 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.
1---2name: cold-call-coach3description: 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.4---56# Cold Call Coach78Grades real first-touch cold calls against the DataSift cold-calling playbook.9The pipeline pulls recordings straight from your SmrtPhone web session (no API10exists for this), transcribes them with an audio-capable model so tonality11observations are heard rather than guessed, then Claude grades each gradeable12conversation against `references/rubric.md` and exports everything to Excel.1314## Requirements and setup (once)15161. Python 3.10+ with: `pip install playwright openpyxl` then `playwright install chromium`172. SmrtPhone login: `python scripts/smrtphone_login.py` (headed window opens,18 log in once; the session is saved to `smrtphone_state.json` and lasts weeks)193. OpenRouter key for transcription: copy `scripts/env.example` to `.env` and20 set `OPENROUTER_API_KEY` (get one at openrouter.ai)214. Optional team roster: copy `scripts/roster.example.json` to `roster.json` to22 pin caller roles (a dedicated cold caller's calls never route to another23 rubric) and exclude departed callers2425## Costs (transparent)2627- Transcription: about $0.002 per audio minute (Gemini 2.5 Flash via OpenRouter).28 A week of coaching for a 3-caller team (say 60 gradeable calls averaging 229 minutes) costs roughly $0.25 in transcription.30- Recording downloads and the call log are free (your existing SmrtPhone plan).31- Grading runs inside your Claude session.3233## Pipeline3435Run from your project folder:3637```bash38python scripts/pull_calls.py --min-seconds 60 --days 7 # call log + MP3s39python scripts/transcribe.py # audio -> transcripts + triage40```4142Outputs land in `output/call_coaching/`:43- `call_log.json` / `calls_to_review.json` - every call, then the review set44- `transcripts/{call_id}.md` - diarized transcript with inline delivery notes45 ([long pause 4s], [rushed], [warm tone]) plus a DELIVERY SUMMARY block (pace,46 agent energy, talk balance, notable audio moments) and a label check line47- `review_queue.json` - calls grouped: cold_call / lead_management / closing /48 not_gradeable4950Only grade calls in the `cold_call` group with `worth_grading: true`.51Voicemails, wrong numbers, and dead air are logged, never scored.5253## Grading procedure54551. Read `references/rubric.md` fully (applicability gate, short-call rule,56 auto-fails, 5 categories weighted 15/30/20/15/20, score anchors, both report57 templates, the required SCORES JSON footer).582. Read the worked examples in `references/calibration/` before your first59 session and stay consistent with them.603. For each call: read the whole transcript once; verify AGENT/SELLER labels by61 content (our agent is whoever asks about buying the property); route full vs62 short (short = a DECLINE inside ~30 seconds; doubt resolves to full); score63 each reachable criterion against its 0/3/5 anchors (1, 2, 4 interpolate).644. Write one report per call to65 `output/call_coaching/reports/cold_call/{call_id}_{caller}.md` using the66 rubric's templates exactly, including the CRITERION SCORES table and the67 SCORES JSON footer.685. Roll up a per-caller scorecard, then export the workbook:6970```bash71python scripts/export_excel.py72```7374One styled Excel file: Summary (full calls, /100, band colored), Short Calls75(opener + conversion on their own scale), Coaching Detail, Caller Scorecard,76Criterion Scores.7778## Grading quality protocol (mandatory verify pass)7980- Every quote must appear verbatim in the transcript (check by substring).81- Every criterion score cites a quote or an explicit absence.82- N/A only where the rubric allows; list every N/A and redistributed weight.83- Recompute the math: tables must foot exactly and the JSON footer must match84 the tables number for number.85- Grade ONLY what is audible on the call. CRM state and dispositions are never86 scored.87- Full calls and short calls are different scales; never rank them together.8889## Measured reliability9091Independent graders reproduce full-call totals within about +/-3 points and92agreed on the grade band in every test comparison; see93`references/reliability.md`. Treat totals within 5 points as equivalent, and94get a second grade before any band-boundary personnel decision.9596## Failure modes9798- `pull_calls.py` exits 2 or reports an HTML response: session expired; re-run99 `python scripts/smrtphone_login.py`.100- `transcribe.py` key error: set OPENROUTER_API_KEY in `.env`.101- Recording URL 404: the recording was deleted in SmrtPhone; log and skip.102- Transcripts can truncate on very long calls; N/A the unreachable categories103 and note it.104105## Related106107- Follow-up and qualification calls: the `lead-manager-coach` skill.108- Offer and negotiation calls: the `closer-coach` skill.109- This skill grades call QUALITY, not dial volume KPIs.
DataSift-Ty-Personal