Skip to content

raven-sourav/Discovery-Call-Engine

Repository files navigation

Discovery Call Intelligence Engine

Scores SDR discovery calls and AE value-selling calls against your sales methodology and marketing materials. Produces coaching cards, deal bucketing, marketing intelligence, and competitive intel capture from call transcripts.

Built as a Claude Code skill system with an orchestrator-agent architecture. No external APIs, no database dependencies, no infrastructure. Call transcripts go in, structured intelligence comes out.

What this does

For sales leadership:

  • Scores every call against your defined sales methodology
  • Buckets deals automatically: Rescue, Advance, Route to self-serve, Disqualified, Benchmark
  • Threshold-based bucketing (ICP fit + score + behavioral signals), not gut feel
  • One coaching priority per call, not fifteen notes
  • Team-wide pattern detection across weeks
  • AE value-selling scoring with cherry-pick detection (5 signals, composite 0-15)
  • Deal-level stitching: SDR score + AE score + CRM outcome per deal
  • Score-outcome reconciliation that compares AI scores against CRM reality

For PMMs:

  • Messaging alignment scored 0-10 per call with transcript evidence
  • Proof point deployment tracking by vertical
  • Competitive intelligence flowing back: which competitors, what context, exact prospect language
  • Prospect verbatim capture for battlecards, ad copy, landing page headlines
  • Bad-fit lead source identification for budget reallocation
  • Week-over-week enablement trend: is your training actually changing behavior?

Same transcript. Scored once. Two functions served.

How it works

                         CLAUDE.md (orchestrator)
                               |
              Detects mode from user request
                               |
          ┌────────────────────┼────────────────────┐
          |                    |                     |
    Mode 1: Direct       Mode 1B/2: Batch      Mode 3-5: Direct
    (single call)        (day or week)          (Q&A / calibrate / brief)
          |                    |                     |
   score-call skill    Spawn call-scorer        Load relevant
          |            agents in parallel        skill or answer
          |                    |                from references
          v                    v
   Scored property      Collect all scored
   set (1 call)         property sets
                               |
                        report-writer agent
                               |
                   ┌───────────┼───────────┐
                   |           |           |
              Daily Report  Weekly    Marketing
              (Mode 1B)     Report    Intelligence
                            + Cards   Log (Mode 2)
                    AE CALL SCORING TRACK
                         CLAUDE.md
                              |
             Detects AE mode from user request
                              |
         ┌───────────────────┼────────────────────┐
         |                   |                     |
   Mode 9: Direct      Mode 10: Batch        Mode 11: Calibrate
   (single AE call)    (week + deal review)   (AE score overrides)
         |                   |
  ae-score-call skill   Spawn ae-call-scorer
         |              agents in parallel
         v                   |
  AE scored property    Deal Synthesis
  set (1 call)          (composites per deal)
                              |
                        + CRM weekly dump
                              |
                    ae-report-writer agent
                              |
                ┌─────────────┼─────────────┐
                |             |             |
         Rejection       Loss         Funnel Health
         Audit           Autopsy      + AE Coaching

Core principle: Score once, report many times. The transcript is read and scored exactly once. Every downstream report reads from the scored property set.

Operating modes

Mode Trigger What it does
1 Single transcript submitted Score one SDR call
1B Day's transcripts Score a day's calls in parallel + daily report
2 Week's transcripts Score a week + all reports + marketing intelligence
3 Knowledge question Answer from references
4 Leader disagrees with score Calibrate scoring rules
5 Marketing intelligence request Generate marketing brief from scored data
8 CRM outcomes available Reconcile scores vs reality
9 Single AE transcript Score one AE value-selling call
10 AE week + CRM dump Score AE calls + deal review (rejection audit, loss autopsy, funnel health)
11 Leader disagrees with AE score Calibrate AE rules + cherry-pick verdicts

Getting started

1. Build your knowledge base

The engine scores calls against your marketing materials. You need to create these files in references/:

File What it contains
icp-and-pains.md Who you sell to, who you don't, buyer personas, pain mapping
positioning-and-narrative.md How you positioned the product, narrative arc
messaging-matrix.md Value props by persona, JTBD mapping
product-knowledge.md Features, pricing, capabilities, limitations
allowed-claims.md What reps can assert (with source), what they can't
evidence-vault.md Case studies, proof points indexed by pain
differentiation-and-competitive.md Competitor profiles, battle cards
discovery-scoring-protocol.md Your SDR scoring methodology and rubrics
ae-scoring-protocol.md Your AE scoring methodology and cherry-pick model

Don't have these files yet? Use messaging-gpt-builder to create them. It has reasoning scaffolds for each file type that guide you through building strategic, source-tagged knowledge bases. The scaffolds teach thinking, not form-filling.

Each placeholder file in references/ shows the expected structure and links to the relevant messaging-gpt-builder scaffold.

2. Configure file paths

Update the File Paths section in CLAUDE.md to point to where your call transcripts live.

3. Encode your sales methodology

Edit references/discovery-scoring-protocol.md with your scoring rubrics. The engine ships with example rules (Sandler-based), but the methodology is yours to define. Start with the seven rules in the file and calibrate through use.

4. Score your first call

Open Claude Code in this project directory.
Provide a transcript: "Score this call" + paste or point to transcript file.

5. Score a week

"Score all calls from this week"
Point to the folder containing the week's transcripts.

Architecture

discoveryclaude/
├── CLAUDE.md                          ← Orchestrator + router (start here)
├── skills/
│   ├── score-call/SKILL.md            ← Mode 1: single call scoring
│   ├── score-day/SKILL.md             ← Mode 1B: day batch + daily report
│   ├── score-week/SKILL.md            ← Mode 2: week batch + all reports
│   ├── calibrate/SKILL.md             ← Mode 4: score override handling
│   ├── marketing-brief/SKILL.md       ← Mode 5: marketing intelligence
│   ├── reconcile/SKILL.md             ← Mode 8: score vs CRM outcome reconciliation
│   ├── observe/SKILL.md               ← Self-improving: execution logging
│   ├── inspect-amend/SKILL.md         ← Self-improving: inspect, amend, evaluate
│   ├── ae-score-call/SKILL.md         ← Mode 9: single AE call scoring
│   ├── ae-score-week/SKILL.md         ← Mode 10: AE batch + deal review
│   └── ae-calibrate/SKILL.md          ← Mode 11: AE score override handling
├── agents/
│   ├── triage.md                      ← Haiku pre-scan, flags Tier 2/3 refs
│   ├── call-scorer.md                 ← Scores one SDR call in parallel
│   ├── report-writer.md               ← Aggregates SDR scores into reports
│   ├── ae-call-scorer.md              ← Scores one AE call in parallel
│   └── ae-report-writer.md            ← Aggregates AE scores + deal review
├── references/                        ← Your knowledge base (replace placeholders)
│   ├── discovery-scoring-protocol.md  ← SDR scoring rubrics + rules
│   ├── ae-scoring-protocol.md         ← AE scoring + cherry-pick model
│   ├── icp-and-pains.md               ← ICP, personas, pain mapping
│   ├── positioning-and-narrative.md   ← Positioning, narrative arc
│   ├── messaging-matrix.md            ← Value props by persona
│   ├── evidence-vault.md              ← Proof points, case studies
│   ├── product-knowledge.md           ← Features, pricing, capabilities
│   ├── allowed-claims.md              ← Verified + forbidden claims
│   ├── differentiation-and-competitive.md ← Competitor profiles
│   ├── vertical-example.md            ← Vertical playbook template
│   ├── call-score-schema.md           ← JSON schema for SDR scored calls
│   ├── ae-call-score-schema.md        ← JSON schema for AE scored calls
│   ├── hubspot-notes-template.md      ← SDR CRM notes formatting
│   └── ae-hubspot-notes-template.md   ← AE CRM notes formatting
├── db/                                ← Database + web app specs (optional)
└── _logs/                             ← Self-improving skill observation logs

Scoring framework

Call buckets (SDR)

Bucket Criteria Action
B1 Rescue ICP fit + Score < 65 + buying signal + SDR gap Sales leader acts within 48hr
B2 Advance ICP fit + Score >= 65 Deal moves to AE
B3-PS Route out Real use case but below ICP size Route to self-serve/PLG motion
B3-DQ Disqualified No product fit Remove + flag lead source
B4 Benchmark ICP fit + Score >= 85 Team training material

AE segments

Segment Criteria Action
A1 Rescue AE Score < 60 + SDR scored well + deal alive Sales leader review
A2 Advancing AE Score >= 65 + concrete next step Coach on gaps
A3-VR Valid Rejection Rejected + cherry-pick score 0-6 Legitimate DQ
A3-CP Cherry-Pick Rejected + cherry-pick score >= 7 Escalate with evidence
A4 Benchmark AE Score >= 85 + Value Presentation >= 8 AE training material

Cherry-pick detection

5 signals, each scored 0-3. Composite 0-15.

Signal What it measures
Effort Did the AE actually try to sell?
DQ Reasoning Is the rejection evidence-backed or vague?
SDR-AE Gap SDR scored well but AE rejected. Why?
Value Presentation Did the AE present value or just screen?
Pattern Does this AE reject similar leads repeatedly?

Verdicts: 0-3 Clear, 4-6 Watch, 7-9 Likely, 10+ Confirmed.

Related projects

  • messaging-gpt-builder creates the knowledge base files this engine scores against. Use it to build your positioning, messaging matrix, ICP, battlecards, and proof points with reasoning scaffolds.

About

Part of the [Build with Claude Code for PMMs]) series. Built with Claude Code. No code written manually.

About

Call intelligence engine for Sales & PMMs. Scores SDR discovery calls and AE value-selling calls against your marketing materials. Built with Claude Code.

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors