Template · 2 to 3 daysFoundersSolo operatorsAgency owners

Gmail AI Triage + Draft Reply

Gmail push-notifies on every new message. Claude Haiku classifies in 400ms into one of 7 labels (client, sales-inbound, sales-outbound-reply, vendor, admin, newsletter, noise). Haiku applies the label, then drafts a reply in your voice (fine-tuned on your last 200 sent emails) as a Gmail draft inside the thread. If urgent, a Slack DM fires. Inbox-zero becomes a 15-minute review job, not a 2-hour typing job.

Book the audit · 30 minutes · No pitch

Tool-interchange map · every layer is swappable

Gmail AI Triage + Draft Reply · tool-interchange map

01 · TriggerGGmail pushOutlook pushiCloud MailIMAP poller+3 more

02 · ClassifyCClaude HaikuGPT-4o miniGemini FlashMistral Small+3 more

03 · Label + fileGGmail labelsOutlook categoriesNotion inboxAirtable+3 more

04 · Draft replyCClaude SonnetGPT-4oGemini 2.5Fine-tuned voice model+3 more

05 · Notify (urgent only)SSlack DMSMS via TwilioPush notificationEmail flag+3 more

Every layer is swappable · the architecture is not

What it does

Every Gmail message push-notifies the workflow in real time (no polling, no 5-minute lag). Claude Haiku classifies in about 400ms into one of seven labels: client, sales-inbound, sales-outbound-reply, vendor, admin, newsletter, noise. The right Gmail label is applied and the message filed into the right category view.

For anything that warrants a reply, a second pass runs: Claude Sonnet drafts the response in your voice, fine-tuned on your last 200 sent emails. The draft is saved inside the Gmail thread, not sent. You open the thread, skim the draft, edit if needed, press send. If Haiku flags a message as urgent, a Slack DM fires with the subject line and a link.

Inbox-zero becomes a 15-minute review job, not a 2-hour typing job. The workflow runs on Gmail push, not cron, so the draft is usually waiting before you finish your coffee.

Real-world use cases

Who actually uses this, and what changes for them.

Founder with 300+ inbox emails a day

Your inbox has become a second full-time job. You spend more time typing replies than doing the actual work.

Before

You wake up to 140 unread messages. You spend the first 90 minutes of every morning reading and typing. Half of what you write is a variation of the same 12 replies you send every week.

After

Every message already has a label and a draft sitting in the thread. You run a 15-minute triage pass at 9am and another at 3pm. You edit drafts instead of writing from scratch. Inbox-zero is actually achievable by lunch.

Investor-relations lead at a Series-B startup

Every LP update, every warm intro, every press query comes through the same inbox. The moment one goes unanswered for 48 hours, a relationship takes a small hit.

Before

Urgent messages from LPs and press sit next to newsletter noise for hours because everything looks the same in the unified inbox.

After

LP and press messages are auto-labelled and the Slack DM fires inside 30 seconds of arrival. Routine vendor stuff is labelled and left for the batch review. Nothing urgent slips past the first hour anymore.

What changes for your team

The classifier catches sales-outbound-reply separately from sales-inbound, so real buyers never sit behind cold pitches.
Draft quality improves over time · you can swap in a fine-tuned model on your last 200 sent messages and the voice match gets eerie.
The Slack urgent ping runs on Claude Haiku at 400ms, cheap enough to run on every message all day.
You stop needing a dedicated EA for inbox triage. The workflow is the EA.

Real runs, observed

Trigger

Inbound cold pitch arrives at 2:47am from an unknown sender

What landed

Labelled "sales-outbound-reply" (filed), no draft generated, no Slack ping. You see it in the batch review at 9am with 40 others like it.

Time

400ms classify, zero human cost

Trigger

Existing client email at 10:12am asking about Q2 deliverables

What landed

Labelled "client", draft reply waiting in thread referencing your Notion project page and the last update sent, Slack DM fires with subject

Time

8 seconds end-to-end

Default tool stack

GGmail pushCClaude HaikuGGmail labelsCClaude SonnetSSlack DM

Sample prompts

Time + scope

2 to 3 days

To ship

Fixed

Pricing model

Included

Source JSON

14 days

Tweak window

Live in these engines

Engine

timothy armoo

Founder inbox triage layer

→

Related templates

Template · 2 to 3 days

Invoice + Receipt Extractor

A Gmail label or Drive folder triggers the run. LlamaParse extracts text from the PDF (handles tables, multi-column, bad scans). Claude Sonnet returns a strict schema: vendor, invoice number, date, subtotal, tax, total, currency, line items. A QuickBooks bill is created against the vendor record (or the vendor is auto-created); line items map to the right expense accounts via a lookup table you own. An audit row lands in Notion with the original PDF attached. Month-end reconciliation drops from 6 hours to 20 minutes.

View →Template · 2 to 3 days

Prospect Research + Send

A webhook takes a LinkedIn URL or email. Perplexity pulls 90 days of company news, funding, and headcount shifts. Claude Sonnet scores against your ICP JSON and writes a 3-sentence opener referencing one specific signal. Smartlead queues the email at 9:45am local; Expandi fires the LinkedIn touch 36 hours later. End-to-end: under 3 minutes per prospect.

View →Template · 2 to 3 days

Meeting to Tasks

Fireflies emits the transcript.ready webhook. n8n passes the transcript to Claude Sonnet with a strict JSON schema (decisions, actions with owners and due dates, blockers, follow-ups). Owners resolve by fuzzy-matching speaker names to your Notion people DB. Each action writes a row in the project page; Slack pings every assignee with a deep link. P95 latency call-ends to Slack-ping: 72 seconds.

View →

Ready to ship this?

The audit takes 30 minutes.
The first template ships in days.

Book a call. I review your stack, confirm the fit, and we start the same week.