---

name: call-analysis-workflow description: Build call analysis dashboards from CallRail data. Downloads calls, transcribes with AssemblyAI, classifies by category, and generates a professional React dashboard.

Call Analysis Workflow

Build professional call analysis dashboards for businesses using CallRail.

Overview

This workflow transforms raw CallRail data into an insightful dashboard showing:

• Call categories (Customer, Spam, Operations, Incomplete)
• Missed opportunities (customer voicemails that need follow-up)
• Sentiment analysis per call
• Transcripts with speaker diarization
• Weekly breakdowns and methodology explanations

Required Skills

This workflow uses two existing skills:

1. assemblyai-transcribe

Location: /root/clawd/skills/assemblyai-transcribe/SKILL.md Purpose: Transcribe call recordings with speaker labels and sentiment analysis Required: ASSEMBLYAI_API_KEY in environment

2. interface-design

Location: /root/clawd/skills/interface-design/SKILL.md Purpose: Design consistent, professional dashboard UI Key principles: Dark mode, borders-only depth, 4px spacing, monospace for data

Workflow Steps

Step 1: Fetch Calls from CallRail

See: references/1-fetch-calls.md

Required credentials:

• CallRail API Key (Settings → API → API Keys)
• Account ID (from CallRail URL)

API endpoint:

GET https://api.callrail.com/v3/a/{account_id}/calls.json
Authorization: Token token={api_key}

Key fields to extract:

• id, direction, duration, start_time
• customer_phone_number, customer_city, customer_state
• answered, voicemail
• recording (URL to get audio)
• transcription (CallRail's basic transcript)

Step 2: Download Recordings

See: references/2-download-recordings.md

For each call with a recording:

1. GET {recording_url}.json with auth header
2. Follow redirect URL to download MP3
3. Save to audio/{call_id}.mp3

Note: Only calls with recording field have audio. Outbound calls and missed calls without voicemail have no recording.

Step 3: Transcribe with AssemblyAI

See: references/3-transcribe.md

Use the assemblyai-transcribe skill with these options:

{
  "speaker_labels": true,
  "sentiment_analysis": true,
  "language_detection": true
}

Output per call:

• text - Full transcript
• confidence - Transcription accuracy (0-1)
• utterances - Array of {speaker, text, start, end}
• sentiment_analysis_results - Array of {text, sentiment, confidence}

Step 4: Classify Calls

See: references/4-classify.md

Categories: | Category | Description | Patterns | |----------|-------------|----------| | customer | Real customer inquiries | quote, estimate, price, schedule, project | | spam | Robocalls, solicitation | warranty, medicare, press 1, insurance |

| operations | Internal/employee calls | delivery, materials, job site | | incomplete | Can't classify | No recording, too short, unclear |

Classification logic:

1. No transcript → incomplete

2. Check spam patterns → spam if matched

3. Check customer patterns → customer if matched

4. Check operations patterns → operations if matched

5. Fallback by direction: outbound → operations, inbound → customer or incomplete

See patterns/ directory for full pattern lists.

Step 5: Build Dashboard

See: references/5-build-dashboard.md

Tech stack:

• React 18 + Vite
• Tailwind-style CSS (custom tokens)
• No external UI libraries

Dashboard structure:

├── Tabs: Call List | Missed Opportunities | Methodology
├── Stats Grid: Total, Customers, Spam, Operations, Answered, Missed
├── Filters: Category, Status, Date Range, Search
├── Call Table: Date, Status, Duration, Category, Confidence, Sentiment, Recording
└── Call Detail Modal: Full transcript, speaker labels, sentiment breakdown

Design tokens (dark mode):

--bg-base: #0a0a0b
--bg-surface-1: #111113
--text-primary: #fafafa
--text-muted: #71717a
--color-customer: #22c55e
--color-spam: #ef4444

Step 6: Deploy to Vercel

See: references/6-deploy.md

npm run build
cp -r public/data dist/
vercel --prod --yes

Data Structure

calls-merged.json

{
  "id": "CAL...",
  "direction": "inbound|outbound",
  "duration": 120,
  "recording_duration": 110,
  "start_time": "2026-01-15T10:30:00-08:00",
  "customer_phone": "+1234567890",
  "customer_city": "San Francisco",
  "answered": true,
  "voicemail": false,
  "has_recording": true,
  "has_assemblyai": true,
  "confidence": 0.95,
  "assemblyai_text": "Full transcript...",
  "utterances": [...],
  "sentiment_results": [...]
}

results.json

{
  "generated_at": "2026-02-03T...",
  "stats": {
    "total": 415,
    "inbound": 333,
    "outbound": 82,
    "inbound_answered": 228,
    "inbound_missed": 105,
    "with_recording": 249,
    "voicemail": 38,
    "by_category": { "customer": 34, "spam": 133, "operations": 12, "incomplete": 236 },
    "inbound_answered_by_category": { "customer": 29, "spam": 111, "operations": 9, "incomplete": 79 }
  },
  "calls": [
    {
      "...merged fields...",
      "category": "customer",
      "confidence_classification": 0.85,
      "reasoning": ["Customer patterns detected"],
      "sentiment_summary": { "positive": 40, "negative": 10, "neutral": 50 },
      "incomplete_reason": null
    }
  ]
}

Key Insights

Why calls have no recording

• Outbound calls: CallRail doesn't record by default
• Missed without voicemail: Nothing to record
• Softphone calls: May not route through tracking number

Duration vs Recording Duration

• Duration: Total call time (includes ringing)
• Recording Duration: Only recorded portion (starts after answer)
• Average difference: ~9 seconds (ringing time)

Missed Opportunities

Customer voicemails = leads that weren't answered. Filter by:

calls.filter(c => c.voicemail && c.category === 'customer')

Example Project

See: examples/rhino-concrete/

Live dashboard: https://rhino-call-analyzer.vercel.app

Stats (Jan 2026):

• 415 total calls (333 inbound + 82 outbound)
• 333 inbound breakdown: 228 answered + 38 voicemail + 67 missed
• 249 with recording (all inbound), 166 without

Classification of 228 Inbound Answered:

• 111 spam (49%) - cold calls, robocalls, sales pitches
• 29 customers (13%) - real inquiries about concrete work
• 9 operations (4%) - suppliers, insurance, accounting
• 79 incomplete (35%) - too short/unclear to classify

Missed Opportunities (6 customer voicemails): Real leads that went to voicemail and need follow-up.

Why 166 calls have no recording:

• 82 outbound (not recorded by CallRail)
• 70 inbound missed without voicemail (nothing to record)
• 14 inbound answered without recording (CallRail config issue)