Workflow Architecture · n8n Social Posting Agent

Building a headless social media engine with human-in-the-loop Telegram control

Managing social media distribution across seven different platforms is a masterclass in friction. TikTok and X require snappy, compact text under strict character limits; LinkedIn and Instagram reward structured, long-form storytelling; YouTube Shorts demands curiosity-driven, punchy titles to capture click-through rates. Attempting to manage this manually results in hours lost to copy-pasting and formatting.

Most automation tools force a binary choice: either pay for expensive, rigid SaaS scheduling suites, or build headless scripts that broadcast content blindly without manual quality control. The posting agent solves this by decoupling scheduling from publishing. It uses Google Sheets as a stateful ledger, LangChain for real-time content adaptation, and a dual-workflow Telegram architecture that puts an interactive approval gate directly in your chat app.

The architecture — stateful orchestration without a database

Instead of standing up a dedicated PostgreSQL instance and a custom React frontend just to manage content queues, the posting agent leverages Google Sheets as an accessible, multi-user CRUD dashboard. Editors drop video links, draft captions, and select target platforms using simple checkboxes.

Workflow architecture showing the dual-workflow setup (polling, Telegram approval loop, and multi-platform publishing)

An automated cron trigger polls the sheet every 15 minutes, filtering specifically for rows where Status == 'Ready To Post' and the timestamp has matured. Once picked up, the workflow transitions the row state to prevent duplicate processing and writes critical execution metadata—specifically the instance Exec_ID and the asynchronous Resume_URL—back into the sheet.

// Filter By Schedule (n8n Code Node)
const items = $input.all();
const now = new Date();

const valid = items.filter(item => {
  const scheduledTime = item.json.Scheduled_At;
  if (!scheduledTime || scheduledTime.toString().trim() === '') return true;
  return new Date(scheduledTime) <= now;
});

if (!valid.length) return [];
return [valid[0]]; // Process strictly one row per execution tick to guarantee idempotency

Human-in-the-loop control via asynchronous Telegram webhooks

True automation requires knowing when not to automate. Rather than broadcasting unverified AI-generated content, the Main Workflow halts execution immediately after preparing the content payloads. It constructs a rich markdown preview of the captions, platform targets, and AI-optimized titles, pushes it to an authorized Telegram chat, and enters an asynchronous Wait Approval state.

To handle interactions without polling or timeouts, a second decoupled workflow—the **Telegram Listener**—acts as a persistent command gateway:

// Command Parser — Router & Security Layer
const ADMIN_CHAT_ID = 'YOUR_TELEGRAM_CHAT_ID';
const msg = $input.first().json;
const chatId = String(msg.message?.chat?.id || '');
const text   = (msg.message?.text || '').trim();

if (chatId !== ADMIN_CHAT_ID) return []; // Drop unauthorized traffic silently

// Match pattern: /command_execId optional_payload
const re = /^\/(approve|reject|edit_title|edit_long|edit_short)_([a-zA-Z0-9]+)(?:\s+([\s\S]+))?$/;
const match = text.match(re);

if (!match) return [{ json: { command: 'unknown', execId: '', payload: '', chatId } }];

return [{ json: {
  command: match[1],        // approve | reject | edit_title | edit_long | edit_short
  execId:  match[2],        // Unique workflow execution identifier
  payload: (match[3] || '').trim(),
  chatId
}}];

Content adaptation & multi-platform fan-out

Once the wake-up webhook fires, the Main Workflow resumes execution. If the action was a rejection, it marks the ledger as Rejected and terminates cleanly. If approved, it dynamically overwrites the baseline content with any custom drafts submitted during the review pause, clears the scratchpad columns, and begins media syndication.

To satisfy platform-specific formatting rules without creating duplicate database records, content adaptation happens on the fly during data preparation:

• AI Title Generation: If the YouTube flag is checked, the workflow feeds the base title into a LangChain OpenAI node running gpt-4o-mini with a strict system prompt designed to maximize curiosity while capping output at 100 characters.
• Intelligent Truncation: The system maps full-length storytelling copy to Instagram, Facebook, and LinkedIn, while automatically generating a fallback 280-character substring for TikTok, Threads, and X if an explicit short caption wasn't provided.
• Concurrent API Syndication: A Blotato media node downloads the video binary directly from Google Drive using the extracted file ID, hosting it on a CDN to dispatch parallel API requests across all selected social networks.

Engineering War Stories: what broke and how it was fixed