How to Use n8n with ._Template 113 Social Listening Monitor Resposta

Your brand gets mentioned dozens of times a day across Reddit, X, forums, and review sites — and by the time someone forwards you a screenshot, the conversation is already three days cold. A frustrated customer went unanswered. A competitor comparison thread ranked on Google. A feature request buried in a subreddit died without a reply. Manual social listening doesn't scale, and paid tools like Brandwatch or Mention start at hundreds of dollars a month for features you'll barely touch. Template 113, the Social Listening Monitor with automated response, closes that gap: it watches the channels that matter, scores each mention for urgency and sentiment, and either drafts a reply for your review or posts it automatically — all inside n8n, on infrastructure you already control.

The Problem: Mentions Move Faster Than Your Team

Speed is the whole game in social listening. Research consistently shows response time is the single biggest driver of whether an online complaint escalates or defuses. But the practical reality for a lean team is brutal:

• Fragmentation. Mentions arrive across five or six platforms, each with its own dashboard, notification model, and API quirks.
• Signal-to-noise. For every mention that needs a human, there are ten that are neutral chatter, bot spam, or your own marketing being re-shared.
• Context switching. A founder or ops lead checking six tabs between meetings will miss the one thread that actually mattered.
• Inconsistent tone. When replies do happen, they're written under pressure with no shared voice, so the brand sounds different every time.

The cost isn't just a missed reply. It's the compounding effect: an unanswered negative review keeps ranking, an ignored feature request signals you don't listen, and a competitor thread with no counter-narrative becomes the default answer prospects find. Template 113 exists to make sure the right mentions surface within minutes and get a consistent, on-brand response — without a human babysitting a dashboard.

The Solution: An Always-On Listening and Response Loop

Template 113 is a scheduled n8n workflow that runs a four-stage loop: collect → filter → classify → respond. It polls your chosen sources on an interval, deduplicates against mentions it has already seen, uses an LLM to score sentiment and intent, and then routes each mention down one of three paths — ignore, draft for human approval, or auto-respond. Everything a human needs to review lands in a single Slack channel or database row, not scattered across platforms.

The design principle is tiered autonomy. High-confidence, low-risk mentions (a simple "does this integrate with X?" question) can be answered automatically. Anything sentiment-negative, high-follower, or ambiguous gets drafted and held for a human. You decide where the line sits by adjusting a single confidence threshold, so the workflow can start fully human-in-the-loop and graduate to more automation as you trust it.

Step-by-Step Setup in n8n

1. Trigger — Schedule Trigger node. Set the interval to every 15 minutes for active brands, or hourly if volume is low. Avoid going below 10 minutes: most platform APIs rate-limit aggressively and you'll gain nothing but 429 errors. Store a lastRun timestamp in a Data Store or a small Postgres table so each run only fetches new content.

2. Collect — HTTP Request nodes per source. Add one HTTP Request node for each channel:

• Reddit: GET https://oauth.reddit.com/search with your query in q, sort=new, and a Bearer token from the Reddit app credentials.
• X/Twitter: the recent search endpoint filtered by your brand keywords and -is:retweet to cut noise.
• Hacker News / forums: the Algolia HN API (hn.algolia.com/api/v1/search_by_date) needs no auth and is great for technical brands.
• RSS/Google Alerts: an RSS Feed Read node pointed at a Google Alerts feed catches blog and news mentions cheaply.

Merge all sources with a Merge node (Append mode) into a single item stream, normalizing each into a common shape: { id, source, author, text, url, timestamp }. A Set or Code node handles that normalization.

3. Deduplicate — Code node + Data Store. Hash each mention's ID and check it against the set of IDs you've already processed. Drop anything seen before. This is the piece people skip, and it's why their first version spams the same reply five times. Persist the seen-IDs list; don't hold it only in memory.

4. Classify — AI Agent or Basic LLM Chain node. Feed each surviving mention into an LLM node using an Anthropic Chat Model (Claude Sonnet 5 is a strong default here — fast, cheap, and accurate at short classification). Prompt it to return strict JSON:

{
  "sentiment": "positive | neutral | negative",
  "intent": "question | complaint | praise | comparison | spam",
  "urgency": 1-5,
  "requires_human": true | false,
  "confidence": 0.0-1.0
}

Use a Structured Output Parser so malformed responses get retried rather than crashing the run. Keep the temperature near 0.2 for consistent scoring.

5. Route — Switch node. Branch on the classification: spam → discard; confidence > 0.8 AND requires_human = false → auto-respond path; everything else → human-review path.

6. Respond — second LLM node + platform post node. The response generator gets a system prompt containing your brand voice guidelines, a few example replies, and the original mention. For the auto path, pipe its output straight into the platform's post/reply node (e.g., an HTTP Request POST to the Reddit comment endpoint). For the human path, send a Slack node message with the mention, the draft reply, and two buttons — approve or edit — wired back through a Webhook node so an approval posts the reply.

7. Log — append every action (mention, classification, action taken, timestamp) to a Google Sheet or database. This becomes your audit trail and your training data for tuning the confidence threshold later.

The Benefits: What Changes Once It's Running

Coverage without headcount. A single workflow watches every channel continuously, so nothing depends on someone remembering to check a tab.

Response time measured in minutes. The 15-minute poll plus instant classification means urgent negatives surface before they gain momentum.

Consistent brand voice. Because every reply flows through the same prompt with the same guidelines, your tone stops depending on who happened to answer.

Cost you control. Self-hosted n8n plus per-token LLM calls typically runs a few dollars a month at moderate volume — a fraction of enterprise listening suites — and your data never leaves your infrastructure.

Graduated trust. Start with 100% human review, watch the drafts for two weeks, then flip the threshold to let the safe cases through automatically. You never have to take the leap blind.

Common Pitfalls and How to Avoid Them

Skipping deduplication. The number one failure mode. Without a persistent seen-IDs store, every run re-processes and re-replies to the same mentions. Build the Data Store check before you connect any posting node.

Auto-responding too early. Do not turn on the auto path until you've reviewed at least 50–100 drafts and confirmed the classifier isn't mislabeling complaints as questions. A wrong automated reply to an angry customer is far worse than a slow one.

Ignoring rate limits. Reddit and X will throttle you hard. Add a Loop Over Items node with a small wait between posts, and wrap HTTP nodes with retry-on-fail and exponential backoff. Never fire dozens of API calls in a tight burst.

Vague brand-voice prompts. "Be friendly and helpful" produces generic replies. Give the response prompt three or four real example answers and explicit rules — no over-promising, always link to docs, never argue publicly — so the output sounds like your team, not a chatbot.

No kill switch. Add a single boolean in a Set node at the top of the workflow that disables all posting. If the classifier drifts or a source floods, you can freeze responses instantly without editing the workflow under pressure.

Treating classification as fire-and-forget. Review the log sheet weekly for the first month. Every misclassification tells you how to sharpen the prompt or move the confidence threshold. The workflow gets meaningfully better with two or three tuning passes — then it largely runs itself.

Set it up human-in-the-loop first, watch it for a couple of weeks, and let the data tell you where to hand it more autonomy. Within a month you'll have moved from finding out about mentions days late to answering the ones that matter within the hour — on infrastructure you own, at a cost you barely notice.