$ ~/archive/ play make-email-triage
transcript_decrypted.log
0.0 I haven't manually triaged an email in six
2.94 months.
3.56 A workflow built in Make.com sorts every
6.86 incoming message into one of three buckets,
9.439 then takes the right action without me touching
11.859 it.
12.339 Today I'm building it with you, in 15
14.839 minutes.
17.1 Here's the picture.
18.32 A new email arrives.
20.179 Make.com grabs it, sends the subject line
23.32 embodied to ChatGPT with one specific prompt.
26.859 ChatGPT returns one word, lead, support, or noise.
30.8 Based on that word, three different actions fire.
34.179 Lead goes to Notion and pings me on
36.38 Slack.
36.92 Support gets a draft reply waiting in my
39.14 inbox.
39.78 Noise gets archived without me ever seeing it.
42.359 Total build time today, 15 minutes.
44.539 Time saved, about five hours every week.
47.06 Let's go.
49.619 Step one, the trigger.
51.6 Open Make.com, click Create New Scenario.
55.1 Search for Gmail.
56.06 Email in the module list, click it.
57.7 Choose Watch Emails.
59.119 Connect your Gmail account if you haven't already.
61.899 Now the settings.
63.219 Folder, Inbox.
64.64 Mark messages as read.
66.319 No, we want them untouched until our workflow
68.879 decides what to do.
70.26 Limit, 10, so it processes 10 emails per
73.2 run.
73.78 Filter, add one filter is unread, so we
76.739 don't reprocess things.
78.159 Click OK.
79.079 The trigger is done.
80.379 Next we need the classifier, the brain that
82.78 decides what kind of email each one is.
86.98 Step two, the classifier.
89.26 Click the plus icon to add a module.
91.859 Search for OpenAI.
93.62 Choose Create a Completion.
95.299 If you haven't connected OpenAI yet, you'll need
98.219 an API key, one-time setup, two minutes.
101.34 In the model dropdown, pick GPT-4-0
104.12 Mini.
104.56 It's cheap and fast.
106.06 We don't need the heavy model for this.
107.859 Now the prompt, paste this exactly.
109.959 Read this email and respond with one word
112.599 only.
113.239 Lead, support, or noise.
115.079 Lead means a sales inquiry or a real
117.98 business opportunity.
119.14 Support means an existing customer with a question
122.06 or issue.
122.84 Noise means newsletter, promotion, automated notification, or anything
127.42 else.
128.159 Email subject.
130.699 Step three, the router.
132.58 The router is what makes Make.com powerful.
135.84 Click the wrench icon on the connector after
138.319 OpenAI.
139.18 Choose Add a Router.
140.74 Three branches will fan out from it.
143.039 On each branch, we set a filter.
145.08 That's the condition for taking that path.
147.159 First branch, filter where the OpenAI response equals
150.319 lead.
151.179 Second branch equals support.
153.02 Third branch equals noise.
155.319 This is the decision tree.
157.18 From here, the email gets sent down exactly
159.539 one path based on what ChatGPT decided.
162.539 Now we wire up what happens on each
164.68 branch.
166.879 Step four, the lead branch.
169.219 This is the high-value path.
171.199 On the lead branch, add two modules.
173.58 First, Notion.
174.74 Create a database item.
176.24 Connect your Notion.
177.419 Pick your CRM database and map the fields.
180.12 Subject from Gmail goes to Title.
182.12 Body goes to Notes.
183.62 Sender email goes to Email.
185.46 Set Status to New Lead.
187.5 Save.
188.24 Second module on the same branch, Slack.
190.599 Send a message.
191.599 Pick your channel.
192.56 Write the message, New Lead, just hit the
194.699 inbox plus the subject and sender.
196.84 Map those fields.
198.199 Save.
198.96 Now when ChatGPT classifies an email as a
201.74 lead, two things happen automatically.
203.879 A CRM email is a lead.
204.719 A CRM entry is created and I get
206.06 pinged.
206.539 Time from email arriving to me knowing about
209.039 it.
210.64 Step five, the support branch.
213.199 On this branch, I don't want full automation
215.819 because customer support requires my actual
218.719 judgment.
219.419 So instead of replying directly, the workflow drafts
222.8 a reply that sits in my inbox waiting
224.819 for me to review.
225.9 Add a Gmail module.
227.62 Create a draft.
228.819 Two, map the sender email.
231.28 Subject, re plus the original subject.
234.08 Buy.
234.719 I send another OpenAI call here that takes
237.52 the email content and writes a friendly first
239.699 draft response based on a template.
241.659 The draft sits in my drafts folder.
244.06 I open it.
245.02 Edit if needed.
246.219 Hit send.
247.12 20 seconds of work instead of three minutes
249.34 of starting from scratch.
252.259 Step six, the noise branch.
254.46 This one is the simplest.
256.3 Add a Gmail module.
258.0 Modify a message.
259.339 Map the message ID from the trigger.
261.819 Set.
262.379 Add labels to a label I created called
265.199 Noise and Mark as read to Yes.
268.199 That's it.
269.24 Newsletters, sales pitches, automated notifications, all of them
273.199 get archived to the noise label
274.839 without ever showing up in my main inbox.
277.42 I never see them.
278.66 They never break my focus.
281.959 Step seven, test it.
283.959 Click run once at the bottom of make
286.079 .com.
286.8 Send yourself a test email.
288.56 Try one obvious lead.
290.3 One support question.
291.519 One promo.
292.68 Watch the modules light up green as the
294.899 email flows through.
296.139 Three common errors.
297.8 One open AI returns more than one word.
301.04 Fix.
301.779 Add reply with own word only.
303.819 No punctuation.
304.939 No explanation twice in the prompt.
307.319 Two, the router branches don't fire.
310.18 Fix.
310.62 Check that your filter values are lowercase exactly
313.139 matching lead, support, and noise.
315.66 Three, you hit the make free tier limit.
318.079 Fix.
318.74 The pro plan is $29 a month.
321.199 Five.
321.5 For five hours a week back, do the
323.259 math.
325.339 Three ways to extend this once you have
327.86 the base running.
328.939 One, replace the open AI module with Claude.
332.06 Same idea, different model.
334.04 I prefer Claude for nuance when the email
336.319 is ambiguous.
337.579 Two, add a fourth branch for VIPs.
340.319 If the sender is on a list of
341.879 important people, route to a separate priority.
344.74 Slack channel that pings me even on weekends.
347.699 Three, connect this workflow to your custom GPTs.
351.319 Host the rest of your tools, publicize the
352.439 GPs.
352.439 Level up and add more types of INPUT
354.899 to make that completely adequate.
355.48 Essex is one of them to add your
358.079 example in the tutorial.
358.439 preview and output page get too empty, check
358.439 forming the false focus in your output and
358.54 Stereo tool.
360.0 It's all there to youtube.ją
360.56 The be過去 it would be like speaking from
368.639 your computer or a VA device.
368.639 Alaska is the nation of fun and music.
368.899 All thatu, you'll see it later.
371.6 Azure, Maori, PSA,iertelega.N hitch that's only
379.139 different to Cocoa and all the others I
381.3 will
381.319 inside Notion automatically.
383.04 That's the loop where the AI generates the
385.3 lead, my workflow
386.439 processes it, and the CRM updates without me
389.439 ever touching it.
390.399 See you there.

I Built an AI Email Triage System in 15 Minutes (Saves Me 5 Hours/Week)

RELATED INVESTIGATIONS
RELATED INVESTIGATIONS