🧹 Shiny Gmail Automation — AI inbox cleaner (3 workflows, 55 nodes, ₱0 budget)

Built with n8n
, , ,
1

Hey everyone! :waving_hand:

Sharing my first n8n build here!

I had around 7,000 emails sitting in my Gmail inbox and I was too lazy to clean them manually — so I built an automation to do it for me.

What it does

Every morning at 6 AM, it:

  • Fetches unprocessed inbox emails
  • Checks my Telegram sender/domain rules first
  • If no rule exists, asks Gemini AI to classify the email
  • Applies a Gmail label (priority, receipts, unimportant, email-subs, or please-review)
  • Archives it out of my inbox
  • Sends me a Telegram summary when it’s done

If AI is unsure about an email, it goes to please-review — it never guesses silently.

I can also manage rules straight from Telegram:
/whitelist boss@company.com → priority
/blacklistdomain spammy-site.com → unimportant
/receipt orders@store.com → receipts

Architecture

Workflow Nodes Purpose
A — Daily Orchestrator 25 Fetches inbox, applies rules, calls child, summarizes
B — Email Processor 19 Classifies and labels one email at a time
C — Telegram Rules Manager 11 Manage rules via Telegram bot

55 nodes total. Budget: ₱0.

Stack

n8n (self-hosted) + Gemini AI + Gmail API + Telegram Bot + Neon Postgres + ngrok

Everything runs on free tier.

Some n8n gotchas I discovered (might help others!)

During deployment I ran into a few issues that took a while to figure out:

  1. Import wiring corruption — IF, Switch, and SplitInBatches nodes can have internally corrupted output mappings after JSON import. The wires look correct visually but route data to the wrong outputs during execution. Fix: delete the node and recreate it from scratch.

  2. SplitInBatches infinite loop — If your Wait node accidentally loops back to a node before SplitInBatches instead of directly back to it, the batch restarts from item 1 forever. Caught this one after 60+ loops. :sweat_smile:

  3. Gmail metadata capitalization — Gmail returns From, Subject, Date (capitalized) but code might expect lowercase from, subject, date. Need to check both or all your emails show empty sender and “(no subject)”.

All documented in the deployment guide with step-by-step fixes.

Links

:link: GitHub: GitHub - chanrylejay/shiny-gmail-automation: AI-powered Gmail inbox cleanup on n8n — Gemini sorts your emails, Gmail labels track everything, Telegram lets you set the rules. 54 nodes, 3 workflows, zero inbox stress. · GitHub

Full deployment guide, troubleshooting, and schema included. MIT License.

About me

I’m not a developer — I’m a non-technical automation builder from Manila. This is my first GitHub project and my first n8n community post!

Built it because I was lazy. Stayed because it actually works. :grinning_face_with_smiling_eyes:

Would love feedback, ideas, or questions! :rocket:

SS1
SS1501×696 337 KB

SS2
SS21044×891 137 KB

2 Me gusta
2

Nice first build - the orchestrator/processor split across Workflow A and B is a smart pattern, keeps the per-email logic clean and reusable. The Telegram rule manager on top makes it genuinely practical to use daily without touching n8n. How are you handling Gemini classification accuracy - do you tune the prompt per label or use a single prompt with all categories listed?

1 me gusta
3

Thank you! :grinning_face_with_smiling_eyes: Yeah, for Gemini I’m using one main prompt with all the categories listed, not separate prompts per label.

I didn’t really try to make Gemini perfect. I made it conservative instead. If it’s confident, it labels the email. If it’s unsure, low confidence, or the API fails, it sends the email to please-review instead of forcing a bad category.

A few things that helped:

  • I use a strict JSON response with fixed categories:
    unimportant, receipts, priority, email-subs, and please-review

  • I keep the temperature low at 0.1 so the output stays consistent instead of getting too creative

  • I only send metadata to Gemini, like sender, subject, and snippet — no full email body

  • I have fallback rules for uncertainty:
    confidence below 0.8 → please-review
    invalid category → please-review
    API or parse issues → please-review

Then for repeat senders or domains, I use Telegram rules so they can skip Gemini completely.

So the setup is basically: one structured prompt, strict categories, conservative fallback, and rules for the obvious stuff.

4

The metadata-only approach is smart - keeps token usage low and avoids any privacy concerns with full email content going to the API. The Telegram bypass for repeat senders is a nice touch too, saves unnecessary API calls for known senders. That confidence threshold setup should make the system pretty reliable in practice.

2 Me gusta