Course resource

Step-by-Step Build Guide

Your first real automation, built one step at a time. The example is lead research; the method transfers to anything.

What you will build: a new row in a sheet triggers research on that company, an AI-drafted opening message, and a notification for you to review.

Allow ninety minutes the first time.

Before you open the tool

Write the line:

New row in "Prospects" sheet
  -> look up the company
  -> AI drafts an opening message
  -> post to my chat channel for review

Set up the sheet with these columns. The last three are written by the automation:

Company	Website	Contact name	Contact email	Research	Draft	Status

Step 1 — trigger, and nothing else

Create the scenario. Add only the "new row" trigger. Run it.

It fires when you add a row
You can see the row data in the output
The column names match what you expect

Do not add step two until this works. Debugging one thing is easy; debugging five is not.

Common problem: the trigger fires on every existing row the first time. Expected. Let it clear, or start with an empty sheet.

Step 2 — the filter

Add a filter immediately after the trigger:

Continue only if: Website is not empty AND Status is empty

Test: add a row with no website. Nothing should happen.

This step is what stops the automation re-processing rows forever and running on incomplete data. It is the step people skip and regret.

Step 3 — gather the input

Add a step that fetches the company's site content, or use a search tool module.

It returns something for a real company
It fails gracefully for a bad URL

Look at the actual output. It will be messier than you expect — navigation, cookie banners, footers. That noise goes into your AI step unless you trim it.

Step 4 — the AI research step

Here is content from a company's website:
{{content}}

Summarise in under 100 words:
- What they do
- Who they sell to
- Anything that suggests they are growing or changing

Use ONLY the content above. If it does not say, write "not stated".
Do not use any outside knowledge about this company.

Temperature 0.2. This is extraction, not creativity.

Output is about the right company
It does not contain facts that were not in the input
Run it twice on the same input — is the output similar?

That last check is how you catch a temperature set too high.

Step 5 — the draft

Write an opening message to {{contact name}} at {{company}}.

What they do: {{research from step 4}}
What I offer: [YOUR ONE SENTENCE]

Rules:
- Under 90 words
- Open with something specific from the research
- One small ask
- No "I hope this finds you well", no "I wanted to reach out"
- If the research does not give me a genuine reason to contact them,
  write "NO CLEAR ANGLE" instead of a message

Temperature 0.6 — this one is writing.

That last rule is the quality gate. A message with no real angle should not be sent, and the automation should say so rather than manufacturing one.

Step 6 — write back and notify

Write the research and draft back to the sheet
Set Status to "drafted"
Post to your chat channel with the company name and the draft

Do not send anything. The output is a draft for you to review.

Step 7 — the error path

Add a route: if any step errors, post to your channel with the step name, the company, and the error.

Test it by putting an invalid URL in a row.

Step 8 — the ten-run test

Add ten real rows. Then check:

All ten processed
Research matches the actual companies
No invented facts
At least one "NO CLEAR ANGLE" — if every single one produced a message, the model is stretching
Drafts sound like a person
Cost per run is what you expected

What you have learned

The pattern generalises:

This build	The general pattern
New row	A trigger you filter tightly
Website content	Gather input, trim the noise
Research summary	An extraction step at low temperature, constrained to the input
Draft message	A generation step at moderate temperature
Post for review	Output to a human, never direct to a customer
Error route	Fail loudly

Now change one thing

The fastest way to understand it is to modify it:

Swap the trigger to a form submission
Add a classification step that routes by company size
Change the output to a draft email instead of a chat message

Each change teaches you more than reading another guide.

Your build log

Step	Worked first time?	What went wrong	Fix
1 Trigger
2 Filter
3 Gather
4 AI research
5 AI draft
6 Output
7 Error path

Back to dashboard