cc-call-summariser

name	cc-call-summariser
description	Generates structured post-call summaries from call transcripts or CSM notes, extracting key discussion points, decisions, action items with owners and deadlines, sentiment signals, and risk or opportunity flags. Routes extracted data to downstream skills for tracking. Use when asked to summarise a call, extract action items from a meeting, generate call notes, create a post-meeting summary, or when any workflow needs structured output from a customer conversation. Also triggers for questions about call documentation, meeting notes, action item extraction, or conversation capture.
license	MIT
metadata	{"author":"Stephen Rogan","version":"1.0.0","pillar":"communication-content","category":"content-engine"}

Call Summariser

Generates structured post-call summaries from transcripts or CSM notes. Extracts discussion points, decisions, action items, sentiment signals, and flags -- then routes extracted data to downstream skills for tracking and action. Part of the Communication & Content Production pillar.

This is a content extraction and structuring skill. It transforms unstructured conversation data into structured, actionable records. It does not interpret relationship dynamics or make strategic assessments -- those are human reads the CSM applies after reviewing the summary.

When to Run

• Triggered: After a customer call when a transcript is available (via Gong, Chorus, or equivalent)
• On-demand: When a CSM provides notes and requests a structured summary
• Batch: End-of-day processing of all calls from that day

Core Execution Logic

Step 1: Ingest Source Material

Accept input from any available source:

Source	Quality	Processing Notes
Full transcript (Gong, Chorus)	High	Best extraction quality. Speaker labels enable per-person analysis
Partial transcript (auto-generated)	Medium	May miss segments. Speaker labels may be inaccurate. Flag sections with low confidence
CSM-provided notes	Variable	Depends on note quality. Summary is only as good as the input
Meeting recording metadata	Low (supplementary)	Attendees, duration, date -- context only, not content

Step 2: Extract Discussion Points

Identify the 3-7 main topics covered during the call:

• Detect topic transitions in the transcript (shift in subject, new questions, agenda item changes)
• Summarise each topic in 1-2 sentences capturing what was discussed and any conclusion reached
• Preserve conversation order -- the sequence of discussion sometimes matters (what was raised first indicates priority)

Step 3: Extract Decisions and Commitments

Scan for explicit and implicit commitments:

Type	Signal Language	Example	Confidence
Explicit commitment	"I will...", "We will...", "Let me..."	"I will send the API docs by Friday"	High
Customer commitment	"We can...", "I'll have...", "Our team will..."	"We'll have the data ready by next week"	High
Agreed action	"Let's do...", "We should...", "Agreed"	"Let's schedule a follow-up for April"	Medium -- "let's" without a named owner is ambiguous
Implicit commitment	Offered without a deadline	"I can look into that"	Low -- may be polite hedging. Flag for CSM confirmation
Decision	"We've decided...", "The plan is...", "Going with..."	"Going with Option B for the integration"	High

For each action item captured:

• Action: What needs to happen (stated clearly, not verbatim)
• Owner: Who committed. If "let's" without a name, flag as "Owner TBD -- confirm"
• Deadline: Explicit date if stated. Inferred date if implied ("next week" = [specific date]). "No deadline stated" if neither
• Priority: High (affects customer directly or has a hard deadline), Medium (important but flexible), Low (internal housekeeping)
• Source quote: The specific passage where the commitment was made (for CSM verification)

Step 4: Assess Sentiment

Analyse the emotional undercurrent of the conversation:

Signal Category	Indicators	Confidence
Positive	Praise, enthusiasm, forward-looking language, expansion interest, laughter (if detectable), "love" or "great" language	High for explicit praise. Medium for tone inference
Negative	Frustration, concern, hedging, multiple caveats, "disappointed" or "concerned" language, repeated questions (indicating the answer was not satisfactory)	High for explicit frustration. Medium for hedging detection
Competitive	Mentions of alternatives, comparison language ("Competitor X does..."), pricing pressure framed as "other options"	High for explicit competitor naming. Medium for implied evaluation
Disengagement	Short answers, low energy, delegation to a junior person, "we'll get back to you" with no timeline	Low-Medium -- may indicate disengagement or simply a busy day

Produce an overall sentiment classification: positive / neutral / negative / mixed. Include specific evidence for the classification.

Step 5: Route Extracted Data

Structured data from the summary feeds downstream skills:

Extracted Element	Routes To	Purpose
Action items with owners and deadlines	lo-sla-monitor	Commitment tracking
Risk signals (frustration, competitive mentions, disengagement)	bi-risk-detector	Risk signal input
Expansion signals (growth interest, new team mentions)	bi-expansion-detector	Expansion signal input
Product feedback (feature requests, product complaints)	pa-feedback-aggregator	Product feedback pipeline
Competitive mentions	bi-competitive-intel	Competitive signal input
Full summary	CRM activity record (via cc-crm-updater)	Account history
Summary + action items	cc-email-drafter	Post-call follow-up email generation

Output Format

{
  "call_id": "string",
  "account_id": "string",
  "date": "2026-03-10",
  "duration_minutes": 32,
  "source": "gong_transcript",
  "attendees": {
    "customer": [
      { "name": "Tom Chen", "role": "VP Engineering", "speaking_share": 0.45 },
      { "name": "Lisa Park", "role": "Engineering Manager", "speaking_share": 0.20 }
    ],
    "internal": 
             
    
  
   
   
    
       
       
       
    
    
       
       
       
    
    
       
       
       
    
  
   
    
  
   
    
       
       
       
       
       
       
    
    
       
       
       
       
       
       
    
    
       
       
       
       
       
       
    
    
       
       
       
       
       
       
    
  
   
     
     
             
             
    
  
   
     
     
             
             
    
     
     
             
    
  
   
     
     
     
     
     
  
   

Handoff to Human

Decision Point	Context Provided	What the Human Decides
Summary accuracy	Full structured summary with source quotes	Whether the agent captured the tone and intent correctly. Did it miss nuance, misinterpret a polite deflection as agreement, or overweight a casual mention?
Commitment validation	Action items with confidence ratings	Whether low-confidence items are real commitments or polite conversation. "I can look into that" may or may not be an action item -- the CSM was in the room and knows
Sentiment interpretation	Sentiment signals with evidence	Whether the classification matches the CSM's read. A customer can say positive things while being fundamentally dissatisfied -- the CSM detects this from tone, body language, and context
Routing overrides	Suggested routing for flags and signals	Whether the expansion signal is worth routing (or just casual curiosity), whether the product feedback is formal enough to log
Follow-up content	Summary and action items ready for cc-email-drafter	What to include in the follow-up email. Some discussion points are internal-only. Some action items do not need to appear in the customer-facing recap

Confidence and Limitations

• High confidence for action item extraction from clear transcripts with explicit commitment language ("I will", "by Friday")
• High confidence for topic identification -- major subject transitions are detectable in transcript structure
• Medium confidence for sentiment assessment -- explicit praise and frustration are clear, but subtle signals (hedging, low energy, polite dissatisfaction) require human interpretation
• Medium confidence for implicit commitments -- "let's circle back on that" may be a real commitment, a polite deferral, or an empty phrase. The skill flags these for confirmation rather than asserting
• Low confidence for disengagement detection from transcript alone. Short answers and delegation may indicate disengagement, a busy day, or simply a reserved communication style
• Cannot extract information from sidebar conversations, body language, facial expressions, or unrecorded portions of the meeting. The most important moment in a call may happen in the 30 seconds after recording stops
• Transcript quality varies by platform, audio quality, speaker count, and accent. Poor transcripts degrade all extraction
• Multi-person calls with overlapping speakers produce lower-quality transcripts and action item attribution may be incorrect

Dependencies

Required:

• Call transcript integration (Gong, Chorus, or equivalent) OR CSM notes input

Strongly recommended:

• CRM API (for activity logging via cc-crm-updater)
• lo-sla-monitor (action item routing for commitment tracking)
• bi-risk-detector (risk signal routing)
• bi-expansion-detector (expansion signal routing)
• pa-feedback-aggregator (product feedback routing)
• bi-competitive-intel (competitive mention routing)
• cc-email-drafter (follow-up email generation)

Downstream consumers:

• lo-sla-monitor (action items)
• bi-risk-detector (risk signals)
• bi-expansion-detector (expansion signals)
• bi-competitive-intel (competitive mentions)
• pa-feedback-aggregator (product feedback)
• cc-crm-updater (activity record)
• cc-email-drafter (follow-up email content)

References

• references/extraction-patterns.md -- Signal language patterns, confidence calibration, and extraction rules