Chapter 8 — Workflow: Merged Call Summarizer

Category: Workflows Reading time: 7 minutes

---

Purpose

The core workflow. Merges a Zoom chat log and Fathom transcript, runs a 4-step LLM prompt chain, and produces five output files including a weekly community invite. Replaces the original chat-only summarizer.

Workflow file: workflows/merged-call-summarizer.json n8n Workflow ID: 5

Node Flow

Local File Trigger (/home/node/watch)
    → Code: Validate and Check Partner
    → Code: Merge Content
    → Code: Create Output Folder
    → Code: Save transcript.txt
    → LLM: Extract Signal          ← OpenRouter Chat Model 1
    → Code: Save extracted-signal.md
    → LLM: Community Post          ← OpenRouter Chat Model 2
    → Code: Save community-post.md
    → LLM: Compress Post           ← OpenRouter Chat Model 3
    → Code: Save community-post-compressed.md
    → Code: Calculate Next Tuesday
    → LLM: Weekly Invite           ← OpenRouter Chat Model 4
    → Code: Save weekly-invite.md

18 nodes total: 8 Code nodes, 4 LLM Chain nodes, 4 OpenRouter Chat Model sub-nodes, 1 Local File Trigger, 1 implicit entry via Execute Workflow.

The Rendezvous Pattern

The “Validate and Check Partner” node is the heart of the design:

// Determine what type of file arrived
if (filename.includes('-zoom-chat')) {
  partnerFile = `${watchDir}/${datePrefix}-transcript.txt`;
} else if (filename.includes('-transcript')) {
  partnerFile = `${watchDir}/${datePrefix}-zoom-chat.txt`;
}

// If partner isn't here yet, stop and wait
if (!fs.existsSync(partnerFile)) {
  return [];  // Empty array = workflow stops
}

// Both present — read both and proceed

Returning an empty array ([]) from a Code node stops the workflow without error. This is how the “wait for partner” behavior works — the workflow simply doesn’t continue until both files exist.

The Merge

Both files are concatenated with clear section headers:

# MEETING TRANSCRIPT

[00:01:29] shakur: How's it going, Tom?
[00:01:30] Patrick Chouinard: Hey, hello.
...

# ZOOM CHAT LOG

<raw chat log content>

This merged text is what the first LLM processes.

LLM Prompt Chain

Step 1: Extract Signal

Input: Merged transcript + chat log (~100-200KB for a 2-hour call) Output: Structured analysis with sections for Resources, Q&A, Insights, Tools, Follow-Ups

The prompt instructs the LLM to focus on information that would be useful one week later — filtering out greetings, filler, and casual banter. It explicitly handles the dual-source nature: “You are analyzing the transcript and Zoom chat log of a weekly AI community call.”

Step 2: Community Post

Input: Extracted signal from Step 1 Output: Polished, ready-to-post community recap

Critical: The prompt explicitly instructs plain text output for Skool compatibility:

“Output PLAIN TEXT only. Do NOT use markdown syntax like # headers, bold, - bullet lists, or links. Instead use emoji for section headers.”

Section headers use emoji: 📎 SHARED RESOURCES, ❓ KEY Q&A, 💡 KEY INSIGHTS, 🛠️ TOOLS AND CONCEPTS, 🔄 FOLLOW-UPS, 📝 SUMMARY.

Step 3: Compress Post

Input: Community post from Step 2 Output: 30-40% shorter version preserving all high-value content

Same plain text constraint. Maintains the emoji section headers from Step 2.

Step 4: Weekly Invite

Input: Compressed post + calculated next Tuesday date Output: Community invite post with unique flavor from last week’s content

The prompt defines a rigid structure:

1. Opening hook — playful reference to last week’s content
2. Standard section — reproduced exactly (HOW THE CALLS WORK, rules, etc.)
3. Flavor bridge — references specific unresolved topics/people
4. Closing — Zoom link and date

Tone rules: playful and welcoming, no negative call-outs, no inside jokes that exclude non-attendees.

Next Tuesday Calculation

const callDate = new Date(datePrefix + 'T12:00:00Z');
const dayOfWeek = callDate.getDay();
const daysUntilTuesday = (2 - dayOfWeek + 7) % 7 || 7;

The || 7 ensures that if the call is on a Tuesday, the invite targets the next Tuesday (7 days later), not the same day.

Date formatting includes ordinal suffixes: “September 9th”, “March 3rd”, “January 21st”.

LLM Configuration

All four LLM steps use identical settings:

Setting	Value
Model	anthropic/claude-sonnet-4.6
Max tokens	8192
Temperature	0.3
Provider	OpenRouter

Temperature 0.3 provides consistent, focused output while allowing enough variation to keep weekly invites fresh.

Output Files

All saved to /home/node/output/YYYY-MM-DD/:

File	Size (typical)	Purpose
transcript.txt	~150KB	Formatted raw transcript
extracted-signal.md	~20KB	Structured signal extraction
community-post.md	~15KB	Full community post (plain text)
community-post-compressed.md	~12KB	Compressed version
YYYY-MM-DD-weekly-invite.md	~1.5KB	Next week’s invite