The idea in one page

An automation has two parts. A trigger decides when it starts: a form is submitted, a file appears, it's Monday at 8:00. An action does the work: add a row, send a message, save a file.

In n8n each part is a node, a box on the screen. A workflow is a set of nodes joined by lines on the editor page.

Each form you build has two web links. The Test URL works only while you are testing; the Production URL works once the workflow is published.

Every time the trigger fires, n8n makes one execution, one run through the workflow. The Executions tab lists them with a status: Succeeded or Error, in red when something failed. That list is your run history. It is the only honest answer to "did it work?"

A workflow has two modes, and mixing them up causes most beginner trouble:

Three facts change how you work:

1. It runs on your own computer. Once published, the workflow keeps working as long as your computer is on and n8n is running, even after you close Desk's window. That's different from a service running somewhere else, and it's why you need to know how to stop it.
2. The trigger's reply tells you nothing about the rest. A form set to answer as soon as it is submitted says thank you before the workflow has done anything. If a later node fails, the person filling in the form never knows.
3. The off switch is Unpublish. Find it before you publish, not while something is misfiring.

Which tool? This book uses n8n, installed and started through Heidelberg AI Desk: a visual editor, strong enough to grow with you, running on your own computer from the first click. Make is a visual middle ground with many ready-made connectors, the links to other apps. Zapier connects the most apps with the least setup, but offers less control. All three teach the same ideas; what you learn here carries over.

Watch it once (30 seconds)

No sound. A yellow dot points at each button while the caption explains it. The first seconds show the earlier Desk; yours looks different, and n8n is the same version. Read the tour as text (opens in a new tab).

Set up (steps 1 to 3, once)

Step 1: Open Heidelberg AI Desk and find n8n. Open Heidelberg AI Desk from the Start menu, or with a click on its icon next to the clock, and click App Store.

You should see: a card for n8n. On the left, the window says Local apps can run; if it doesn't, click Set up Podman first.

Step 2: Click Install and open. On n8n's card, click Install and open. The first time, Desk downloads the app, about 1.2 GB, so this takes a few minutes.

You should see: n8n open in your browser, at the address http://127.0.0.1:35678/.

Step 3: Set up your account. n8n asks you to set up the owner account: an email address, your name and a password of at least 8 characters with a number and a capital letter. n8n sends nothing to the address. Write the password down: on your own computer, n8n can't mail you a new one.

n8n may then make two offers. You need neither: click Skip on the licence key, and Set up later in Settings on the assistant.

You should see: after you finish, n8n's Overview page, ready for your first workflow.

Worked example: a log that fills itself

The method comes from T01-L01 · From Idea to Working Result: The Four-Quadrant Method: start with the problem, explore it, then pick the tool and build the smallest version. Here it is on the freezer log, then on the call log.

Lab: the freezer-box log

Structure it. "Every Monday I spend 20 minutes typing last week's freezer notes into my log, and about one box a month goes missing from it. A good result: each box is logged the moment it goes in, with the time, and nothing is lost."

Understand it. She looks at a month of notes. The typing is dull, but the real damage comes from the delay: notes written on gloves and scraps don't survive until Monday. She also notices she always writes the same three things: box ID, position ("-80 B, shelf 3") and contents. She never writes the date, and she has needed it twice. Back to Structure it: "Log box ID, position and contents on the lab computer next to the freezer, and stamp the time automatically."

Choose tools. Going down the ladder: typing straight into a spreadsheet works, but nothing stops her from skipping a field, and the time is never filled in. A form fixes both: it makes every field required, and the workflow adds the time. n8n's built-in data table holds the log, so she needs no spreadsheet account and no passwords. Nothing here needs AI.

Build it. Seven steps, about 35 minutes.

Step 4: Make the table. On the Overview page, click the + at the top left and choose New data table. Name it freezer_log and choose From scratch.

Click Create, then click Add Column four times to add box_id, position, contents and logged_at, each type String.

You should see: your table with those four columns, next to the id, createdAt and updatedAt columns n8n adds by itself.

Step 5: Add the trigger and the form. Create a new workflow, name it Freezer log, click the plus on the empty canvas, and choose On form submission. Work down its settings:

• Authentication: choose n8n User Auth, so only someone logged in to your n8n can submit it.
• Form Title: Freezer log.
• Click Add Form Element three times for Box ID, Position and Contents. For each: type the label, click Add Attributes, add Custom Field Name (type box_id, position, contents to match), then add Required Field and turn it on.

You should see: three required fields, each with a matching Custom Field Name.

Step 6: Add the action. Click the plus on the form node, search Data table, and choose Insert row. Pick your table, keep Map Each Column Manually, and for box_id, position and contents, switch each to Expression and reference the matching field, for example {{ $json.box_id }}. For logged_at, use {{ $json.submittedAt }}, the time stamp the form adds to every entry.

Step 7: Test with a real entry. Open the form node and click Execute step. n8n opens the test form in a new tab; the first time, it asks you to confirm you trust this form, because it runs using your n8n login.

Click Allow access, then fill in a real box and submit. Back in the editor, the form node's Output shows every field, including submittedAt in UTC, for example 2026-09-26T15:20:44+00:00.

Step 8: Publish it. Click Publish, name the version, and click Publish again in the window that opens.

Copy the Production URL from the form node and submit one real box from the freezer. Open the Executions tab: you see a run marked Succeeded, with your data in it.

Step 9: Break it on purpose. Open Insert row again, click the box_id value, and replace it with a deliberately broken reference:

n8n's own preview under the field already warns you: Referenced node doesn't exist. Publish this version anyway, then submit a test box through the Production URL. The form thanks you exactly as before. It replies before the rest of the workflow runs, so it has no way to know anything is wrong. Open Executions: the newest run shows Error, in red, with Insert row outlined in red on the diagram and "Problem in node 'Insert row': Referenced node doesn't exist" underneath. No row was added to your table.

This is the failure you would never notice from the form. Fix it: put {{ $json.box_id }} back, publish again, and submit once more. The newest execution now shows Succeeded, with your row safely in the table.

Step 10: Find the off switch. Decide now: if you'll use the log this week, leave it published. If not, open the small arrow next to Publish and click Unpublish.

Either way, you now know where the switch is.

What it taught her: the thank-you page can lie. For a log only she uses, checking the Executions tab once a week is enough. If colleagues start using the form, she'll need a warning message when a run fails. That's the next trip round the quadrants (T05-L04 · Make an Automation Safe to Fail Before It Touches Real Systems).

Company: the call log

Structure it. "I type about twelve call notes into a spreadsheet every Friday, and half of them don't say whether I promised a follow-up. A good result: each call logged within a minute of hanging up, and a follow-up flag on every one."

Understand it. He compares a month of paper notes with his inbox. The missing follow-up flags are the real cost: two customers waited a week for a callback he had promised. The typing is a minor cost. Back to Structure it: "Log customer, topic and follow-up yes/no right after the call. The follow-up answer can't be skipped."

Choose tools. The same skeleton as the lab: a form, a data table, no AI. The one change is the follow-up field: a Dropdown element with yes and no, marked required, so the question can't be left blank.

Build it. He follows the same steps (4 to 10) with a table called call_log and fields customer, topic and follow_up. What it taught him: he'd rather know within the hour when a run fails, because a lost call note means a customer waiting. His next trip round the quadrants is a failure message, and after that a Monday summary of every open follow-up (T05-L02 · Put AI in the Middle of a Workflow, and Check Every Answer).

In both cases the automation is small: one trigger, one action. The value came from Understand it: the problem was never really the typing.

What goes wrong

Switching it on before testing. Symptom: you publish, submit a real entry, and the table fills with rows that read {{ $json.box_id }} instead of a real value. Fix: always run Execute step on the trigger with a real entry first, and check every field of the result before you publish.

Testing with the wrong link. Symptom: the Test URL says the form trigger "isn't listening yet." Fix: open the trigger node and click its own Execute step; only then is it listening for that Test URL. The Production URL only works once the workflow is Published.

Trusting the thank-you page. Symptom: weeks later you find gaps in the log, although every submission "went through". Fix: look at the Executions tab, filtered to Error, once a week (T05-L04 · Make an Automation Safe to Fail Before It Touches Real Systems).

Forgetting where it runs. Symptom: you close the laptop and assume the automation has stopped, or you're surprised it's still going days later. Fix: a published workflow keeps running as long as your computer is on, even after you close Desk's window, and Desk starts n8n again the next time you sign in to your computer. Unpublish to stop it for real.

n8n isn't running. Symptom: the Production URL doesn't answer, or n8n won't open at all. Fix: open Heidelberg AI Desk and click Open on n8n's card (it may read Start and open). Closing Desk's window doesn't stop n8n; the Stop button on its card does.

Do it yourself (45 minutes)

Use your own repeated copying job, not the freezer or the call log.

1. Structure it (5 min). Write one sentence: what you copy, how often, and what a good result looks like. List the three or four details you copy every time.
2. Understand it (10 min). Look at the last few weeks of that job. Where does it actually go wrong: the typing, the delay, or a detail that's often missing? Rewrite your sentence if you learned something.
3. Choose tools (5 min). Could a spreadsheet or a template do it? If a form plus a data table is the simplest thing that solves what you found, continue. Keep personal data about other people out of your practice log (T12-L01 · Decide What You May Paste Into AI at Work).
4. Build it (25 min). Follow steps 4 to 10 with your own fields. Test with a real entry, publish, submit once for real, then check the Executions tab and decide whether to leave it published.

Exit check

What you hand in: a screenshot of your workflow's Executions tab showing at least one Succeeded and one Error run, plus one line: "The off switch is ___, and while it's published it runs on ___."

It passes when you can point to the row your test created in the data table, and explain in one sentence why the form saying thank you doesn't prove the row was saved.

Rule to remember

Know how to stop it before you start it.

Details for experts

Checked on 12 October 2026 for Heidelberg AI Desk, on 27 September 2026 for the n8n steps. You don't need this to follow the book; it's for setting things up or answering a colleague.

Further reading and tools

Next books

• T01-L03 · Build the Smallest Thing Your Colleagues Can Run Without You: when an automation is the right choice, and when a template is enough.
• T05-L02 · Put AI in the Middle of a Workflow, and Check Every Answer: adding a model step, and checking its output before you use it.
• T05-L04 · Make an Automation Safe to Fail Before It Touches Real Systems: failure alerts, retries and a way to take over by hand.
• T12-L01 · Decide What You May Paste Into AI at Work: what may go into a form or a cloud service.

Tools

• Heidelberg AI Desk: Taught. Installs and starts n8n on your own computer, with no separate signup.
• n8n: Taught. Runs locally through Heidelberg AI Desk for this book.
• Make, Zapier: Catalogued. Same trigger-and-action ideas, different trade-offs.