Blog URL → Viral X/Twitter Thread in 6 n8n Nodes

The workflow JSON in this guide was imported and saved on n8n cloud, with credentials and prompts verified through node panels. Paste it into a fresh workflow, open the OpenAI sub-node, and run — execution screenshots below show the actual AI output.

What You'll Build

A six-node n8n workflow that:

• Accepts a blog URL + tone + tweet count through a public form
• Fetches the article HTML server-side (no browser spoofing needed)
• Strips nav, footer, scripts, and styles down to the real article text
• Sends everything to gpt-5-mini through a LangChain chain with an Output Parser that forces valid JSON
• Returns a structured thread object: { hook, tweets[], cta, preview_card_headline }

Skip the Build — Import the Workflow

Create a new workflow, click the empty canvas, press Cmd+V / Ctrl+V. All six nodes load with their connections.

Prerequisites

Nodes used: Form Trigger, HTTP Request, Code, Basic LLM Chain, OpenAI Chat Model (sub-node), Output Parser Structured (sub-node) — the last one is the star of this guide.

Step 1 — Import the Workflow

From the home dashboard, click + Create workflow. On the blank canvas, paste the JSON. The workflow loads as "AI Twitter Thread Writer" with all six nodes in place.

Double-click OpenAI Chat Model and confirm the credential shows n8n free OpenAI API credits and the model is gpt-5-mini with temperature 0.7 — higher than a blog-writing chain because threads benefit from more creative hook-writing.

Step 2 — Fetch the Article

Double-click Fetch Article HTML. The URL expression pulls from the form:

={{ $json['Article URL'] }}

The response is set to text (not auto-parsed JSON) because you're getting raw HTML. The neverError: true option keeps the workflow going if a site returns 403, 500, or times out — the next Code node handles empty input gracefully.

The User-Agent Matters

Under Headers → Header Parameters:

n8n's default User-Agent is n8n.io which gets rate-limited by some publications. The generic Mozilla UA passes most filters.

When this step fails: If you test with a Cloudflare-protected site and get 403, swap the Fetch Article HTML node for a call to Firecrawl, Browserless, or ScrapingBee. The Code node downstream doesn't care where the HTML came from.

Step 3 — Extract Clean Text

Double-click Extract Clean Text. This is a Code node with ~40 lines of JavaScript. Three things happen:

1. Extract the title and meta description via regex on <title> and <meta name="description">
2. Strip noise — scripts, styles, nav, header, footer (these dominate the HTML but aren't the article)
3. Decode HTML entities so & becomes & and " becomes "

Then the body gets capped at 6,000 characters. Why? gpt-5-mini's context window can hold 40k+ input tokens, but past ~6k chars of prose, the thread starts paraphrasing the first half and ignoring the second. The cap keeps the thread focused on the article's argument.

const body = cleaned.slice(0, 6000);
return [{ json: { title, description, body, articleUrl, tone, tweetCount } }];

If you need the full article, switch to a chainMapReduce (LangChain's map-reduce chain) which summarizes chunks then combines summaries — but that's 3× the token cost.

Step 4 — The Prompt + Output Parser

Double-click Write Thread. The prompt enforces eight rules:

The Output Parser Magic

Scroll to the Output Parser sub-node (connected to the chain as ai_outputParser). Its JSON schema example:

{
  "hook": "string",
  "tweets": ["string"],
  "cta": "string",
  "preview_card_headline": "string"
}

The chain now enforces this shape. If gpt-5-mini returns malformed JSON or adds prose outside the JSON, n8n throws a clear error you can catch with an Error Trigger — or retry with maxAttempts: 2 in the node Settings tab.

Downstream nodes see the thread as a real object:

• {{ $json.output.hook }} → the hook tweet
• {{ $json.output.tweets[0] }} → first middle tweet
• {{ $json.output.cta }} → the closing CTA

No regex. No markdown stripping. Clean data.

⚠️ Known caveat we hit in testing: gpt-5-mini returned the schema wrapped as { "output": { ... } } instead of the bare object the Output Parser expects, causing strict-mode validation to throw "Model output doesn't fit required format". The AI's underlying response was correct — the strict parser just rejected the wrapper. Two fixes: (1) open the Write Thread node's Settings tab and set On Error to Continue (using error output) — the chain then surfaces the AI's raw text downstream and you parse it manually; (2) drop the Output Parser entirely and let the prompt's strict JSON instruction do the work alone. Option 2 is simpler for solo use; option 1 is better if you're piping into another node.

Step 5 — Run and Inspect

Click Test this trigger on the Form Trigger node, open the test URL, and submit a real article URL. Try one of your own posts or something from your favorite publication.

Watch the execution run node-by-node in the canvas. The final Write Thread node's output panel shows the parsed JSON. Copy the hook, tweets[], cta into X's compose pane and post as a thread — or send them downstream to an auto-post node (see Extensions).

In our live test, feeding a 6,000-character article excerpt produced a structured 5-tweet thread with hook, three substantive middle tweets, and a CTA — ~26 seconds, ~2,044 tokens via gpt-5-mini. Each middle tweet picked a specific concrete fact from the source article (course count, lesson topics, infra commit numbers).

Checking Thread Quality Before Posting

Before hitting publish, verify:

Extensions: Auto-Post to X, Schedule, or A/B

Auto-Post via X's API

After the Write Thread node, add an HTTP Request node for each tweet in the thread (or a Loop Over Items + single node). X's v2 API endpoint:

POST https://api.twitter.com/2/tweets
Authorization: Bearer <your OAuth2 token>
Content-Type: application/json

{ "text": "{{ $json.output.hook }}" }

Chain subsequent tweets with "reply": { "in_reply_to_tweet_id": "{{ $json.id }}" } to build the thread.

Schedule Threads with Typefully / Buffer / Hypefury

Most scheduling apps accept webhooks. Swap X's API for:

POST https://api.typefully.com/v1/drafts

…with your Typefully API key in the Authorization header and the thread as the body. You get drag-and-drop scheduling on top of AI generation.

A/B Test Two Threads From One Article

Duplicate the Write Thread and OpenAI Chat Model nodes. Give one temperature 0.4 (stable) and the other 0.9 (creative). Use a Merge node to zip the two outputs side-by-side, then pick the winner manually or feed both to an AI critic chain.

Bulk-Process a Backlog

Replace the Form Trigger with a Google Sheets Read node (or Airtable, Notion Database Read) that returns 50+ blog URLs. Add a Split In Batches node set to 1 item, and the rest of the workflow runs once per row. Cap concurrency to 3 to stay under API limits.

What's Next

You have a repeatable pipeline for turning any blog into a thread. Next, learn how to turn trending news into a daily digest → Multi-Source AI News Digest, or build an AI chatbot that knows your docs → Chat with a Webpage (RAG).