There’s a specific way n8n automations fail, and it’s worse than crashing.

They just… stop. An API rotates a token, a schema changes, a rate limit kicks in and the workflow errors out on one execution, then the next, then the next. Nothing tells you. You find out three weeks later when someone asks why they stopped getting the report.

The fix is a workflow whose only job is to catch other workflows failing. You build it once, point every other workflow at it, and you never silently lose an automation again.

This one takes about fifteen minutes and needs one credential. Workflow JSON is on GitHub at the bottom.

The Error Trigger is different from every other trigger

Most n8n triggers fire on an external event a webhook, a schedule, new mail. The Error Trigger fires when another workflow on your instance fails.

That’s the whole trick. You don’t add error handling to each workflow. You build one handler, then in each workflow’s settings set Error Workflow to point at it. One handler, all your automations covered.

The payload it receives contains the execution context: which workflow died, which node it died on, the error message, and a link straight to the failed execution.

How it works

1. Error Trigger No configuration. It sits there until something breaks.

2. Shape Error Data (Code node) This is where most of the value is. The raw error payload is nested and inconsistent the message can live in execution.error.message, execution.error.description, or trigger.error.message depending on which node type failed. The Code node checks all of them, falls back to a default, and pulls out the fields that actually matter:

const failedNode = e.execution?.lastNodeExecuted || 'unknown node';
const message = String(rawMessage).split('\n')[0].slice(0, 300);

lastNodeExecuted is the single most useful field in the payload. It tells you exactly where in the chain things broke, which is usually enough to diagnose without opening n8n at all.

The .split('\n')[0].slice(0, 300) matters more than it looks. Stack traces are long, and an alert you have to scroll through on your phone at 2am is an alert you’ll start ignoring.

3. Production Only? (IF node) Checks execution.mode. When you’re building a workflow and clicking Execute manually, it fails constantly that’s what building is. You don’t want a Telegram notification every time.

Manual failures still get logged. They just don’t buzz your phone.

4. Alert Telegram Workflow name, failed node, trimmed error, and a direct link to the execution. Telegram because a bot token takes two minutes and needs no OAuth swap it for Slack or email, the logic is identical.

5. Log Error (Google Sheets) Every failure, manual or production, appended with a timestamp. This turns into the interesting artifact over time: which workflow breaks most, whether failures cluster around a particular API, whether that “one-off” error is actually weekly.

Gotchas

• Setting the Error Workflow is a per-workflow setting, and it’s easy to forget. Building the handler does nothing on its own. Each workflow needs Settings → Error Workflow → point at the handler. Make it part of your checklist when you build anything new.

• The error handler can’t catch itself. If your handler has a bug — bad credential, wrong chat ID — it fails silently and you’re back where you started, except now you think you’re covered. Test it deliberately: build a throwaway workflow with a node that’s guaranteed to fail, run it, confirm the alert lands.

• Don’t alert on everything. The fastest way to make error alerts useless is to send too many. The manual/production split is the minimum filter. If a workflow is expected to fail sometimes, handle that inside the workflow rather than letting it hit the global handler.

• lastNodeExecuted isn’t always the culprit. It’s where execution stopped, which is usually but not always where the problem started — a node can fail because an upstream node passed it something malformed. Treat it as a starting point, not a diagnosis.

• Trigger failures behave differently. If a trigger node itself can’t run (expired OAuth on a Gmail trigger, say), the error surfaces differently than a mid-workflow failure, and the message lands in a different part of the payload. That’s why the Code node checks trigger.error.message too.

Fork it

Import the JSON, add a Telegram bot credential and your chat ID, create a sheet with an Errors tab, then point your existing workflows at it.

👉 [GitHub repo link]

Then go break something on purpose. An error handler you haven’t tested is a guess.

Reply and tell me what your worst silent failure was. I collect these.

Every morning I open Gmail and do the same three seconds of triage on every email: is this urgent, can it wait, or is it noise? Multiply that by 40 emails and it’s 10 minutes of my day spent just deciding what deserves attention before I’ve actually done anything.

So I built a workflow that does the sorting for me. It watches my inbox, has GPT classify every incoming email, pings Slack if something’s urgent, and logs everything to a sheet so I have a running record of what’s landing and how it’s being handled.

This is the kind of workflow you can steal as-is for a personal inbox, or fork for a support/sales inbox with almost no changes. Full JSON is at the bottom import it straight into n8n.

The problem it solves

Inbox triage doesn’t need a human brain —it needs pattern matching. “Is this a billing complaint or a sales lead?” is a classification task, and classification tasks are exactly what LLMs are good at and cheap to run. The workflow doesn’t reply to anything or make decisions for you it just does the sorting so you open your inbox already knowing what’s on fire.

How it works

1. Gmail Trigger Polls the inbox every minute for new mail. Nothing fancy just watching for new messages to land.

2. Extract Email Data (Set node) Pulls out the fields we actually care about: subject, sender, body text, and the message ID. This keeps the payload clean before it hits the AI step.

3. Classify Email (OpenAI node) This is the core of the workflow. A single prompt to gpt-4o-mini with a strict instruction: return only JSON, nothing else.

{"category": "urgent | billing | support | sales | spam", "priority": "high | medium | low", "summary": "one sentence"}

Keeping the model’s output to a single JSON object (no markdown, no chit-chat) is what makes this reliable enough to route programmatically. If you let the model ramble, you’ll spend more time parsing its response than it spent classifying the email.

4. Parse Classification (Code node) A small JS step that parses that JSON safely and merges it back with the original email fields. I added a fallback (category: "support") in case the model ever returns something malformed cheap insurance, one line of code.

5. Route by Category (Switch node) Splits the flow into three branches: urgent, spam (dropped, does nothing), and everything else (other, falls through to logging).

6. Notify Slack (Urgent branch only) Sends a formatted message with sender, subject, and the model’s one-line summary — so you get the gist without opening the email.

7. Log All Emails (Google Sheets) Every non-spam email (urgent or not) gets appended to a sheet with category, priority, summary, and timestamp. Six months from now, this sheet tells you what kinds of emails actually eat your time.

Gotchas worth knowing before you build this

• Force JSON-only output. The single biggest failure mode with AI classification nodes is the model adding a sentence of preamble before the JSON. Be explicit in the system prompt (”respond with ONLY a JSON object”) and still wrap the parsing in a try/catch don’t trust it blindly.

• gpt-4o-mini is plenty here. This is a cheap, high-volume classification task, not a reasoning task. Don’t reach for a bigger model you’re paying more for zero benefit.

• Polling vs. push. Gmail Trigger polls on an interval, which is fine for personal use but adds latency. For a support inbox where minutes matter, look at Gmail’s push notifications via Pub/Sub instead.

• The Switch node’s fallback branch is doing real work. Anything that isn’t explicitly urgent or spam falls through to other and still gets logged this is what stops the workflow from silently dropping emails it’s unsure about.

Fork it

The full workflow JSON is on GitHub — import it into n8n, drop in your own Gmail, OpenAI, Slack, and Google Sheets credentials, and it’s live in about five minutes. Swap the categories in the system prompt to match your own inbox (support tickets, sales leads, whatever you’re drowning in) and the routing logic stays identical.

👉 [GitHub repo link]

Next issue, I’ll cover how to add a second AI step that drafts a reply for the “support” category so triage becomes triage and first-draft response.

If you build on this, reply and tell me what you changed. I read every reply.

Subscribe now

Hey there 👋

Welcome to n8n Explained where we turn automation ideas into real, working workflows.

If you subscribed weeks or months ago and thought, “Is this thing ever going to start?” it’s finally happening 🎉

I’m Sepideh, and I run this newsletter together with Amir.

My background is in data and analytics using AI-Powered solutions, while Amir is a CCNP-certified cloud and DevOps expert. We’re both deeply interested in AI and AI agents, and we actively use automation and AI tools in our daily work.

We’re starting this newsletter to share practical, easy-to-adapt automation examples built with n8n, the open‑source workflow automation tool.

Each post will walk you through real use cases, from simple to advanced, with takeaways you can apply instantly.

Here’s what you can expect:

• 🧩 Step‑by‑step breakdowns of real n8n workflows

• ⚙️ Mini‑automation templates you can copy and reuse

• 📬 One concise email per week — no spam, only value

To kick things off, I’d love to hear from you:

👉 What’s the biggest automation you wish existed?

Hit reply and tell me! your answers might shape the next post (and I might even build your idea in n8n).

Let’s automate something great together.

Catch you soon,

Sepideh Tajik And Amir Soleimani

Creators of n8n Explained

Subscribe now