Stepthrough Start free

ChatGPT: Write and Test the Prompt for an AI Step in n8n, Make or Zapier

Practice it step by step

Start the free practice

Take the free module: why it matters, practice and a quick check →

You do every step yourself in a practice copy of ChatGPT. It checks your work as you go, and “Do it for me” finishes any step you get stuck on. Free, about 9 minutes, and nothing touches your real accounts. Open full screen

Free · Live nowChatGPT for Everyday WorkChatGPT~9 min20 stepsBeginner

Updated . Built against the tool’s current screens and checked step by step. How we check this lesson

The AI step is the part of an automation most likely to go quietly wrong. In this free, interactive lesson you attach six sample emails from an insurance agency's shared inbox, have ChatGPT write the prompt for an AI step in n8n, Make or Zapier that returns only JSON, uses only five desk names and never invents a policy number, and then have ChatGPT test that prompt on every email and fix what the test shows. You get the full prompt to copy. The agency is made up.

At a glance

QuestionAnswer
What you buildAttach six sample emails and the desk list, get a JSON-only routing prompt with a two-desk rule and a never-invent-a-policy-number rule, then have ChatGPT run it on every email, spot the weak results and fix the prompt.
ToolsChatGPT
TimeAbout 9 minutes, 20 steps in 4 parts
LevelBeginner, no coding needed
CostFree
Ways to learn itPractice it step by step

The 4 parts, in short

  1. Start from real emails: Read the six sample emails, Turn on Temporary chat, Open the + menu, Add photos & files, and 2 more steps.
  2. Ask for the AI step prompt: Click the message box, Ask for the prompt, Send it.
  3. Read the rules it wrote: Read the desks and the two-desk rule, Read the policy number rule, Read the output rules.
  4. Test it before you trust it: Click the message box, Ask it to test its own prompt, Send it, Read the results table, and 4 more steps.

Start here first: Summarize a Meeting Transcript with ChatGPT (module 1 of ChatGPT for Everyday Work)

In the course: Part 3 · ChatGPT for Everyday Work (Module 41 of the full course)

The business case: one misrouted email is a lost client

A shared inbox that one person sorts by hand is slow, and it stops when that person is out. An AI step can sort it in seconds, but only if its answer is the same shape every time and it never guesses. A claim sent to billing, or a policy number the AI made up, does more damage than a slow inbox.

Most people write a one-line prompt like "which desk is this for?" and find out what it does wrong after it is live.

What changes: ChatGPT writes a strict prompt with exact fields, allowed values and rules for the hard cases, then runs it on your sample emails so you see the weak spots before a client does.

The example: an insurance agency's shared inbox

Calloway Insurance Agency (made up) gets about 90 emails a day at one shared address. Five desks handle them: CLAIMS, QUOTES, CERTIFICATES, BILLING and OTHER. The sample has six emails: a truck rear-ended this morning, a bakery opening a second location, a certificate of insurance needed by Thursday, a customer charged twice, a vendor pitch, and a basement flood with a billing question in the same email.

Download the file and follow along: [sample-inbox-emails.txt](/content/files/chatgpt-write-test-ai-step-prompt/sample-inbox-emails.txt).

Who this lesson is for

Anyone adding an AI step to an n8n, Make or Zapier automation: office managers, operations staff, agency owners and freelancers who build for them. If you have not built an automation yet, start with your first AI automation. This is the second of three lessons in [Plan and Fix Automations with ChatGPT](/courses/plan-and-fix-automations-with-chatgpt).

What ChatGPT produced in the real run

The first answer was a full routing prompt in a code block, with sections for desks, routing, urgency, policy numbers, summary, confidence and output, and three placeholders for the email's sender, subject and body. The lesson shows it in full so you can copy it.

In the test, all six emails got a valid desk and no policy number was invented. Email 6 (the basement flood plus a premium question) went to CLAIMS with medium confidence, and came back not urgent, which a person at the agency would probably want flagged. ChatGPT said emails 6, 3 and 4 should still get human review, found two weaknesses (relative deadlines like "Thursday" and two-desk follow-up), and wrote replacement routing and urgency rules that keep the same five fields.

ChatGPT's test was simulated by ChatGPT itself. Its own words: "These are simulated AI-step outputs, not results from a live automation." Nothing in this lesson was run inside n8n, Make or Zapier. Paste the prompt into your tool's AI step and run the same six emails there before you trust it.

What this saves you

  • Misroutes: the two-desk rule and the human-review list catch the emails that would land on the wrong desk.
  • Bad data: the policy number rule returns null instead of a made-up number.
  • Rework: you find the weak spots in a chat, not after a week of live emails.
  • What it does not replace: a real test in your own tool, and a person who checks anything below high confidence.

Gotchas

  • Ask for JSON only with exact field names and allowed values. Free text breaks the next step in your automation.
  • Say what to do when an email fits two desks, or the AI will pick differently each time.
  • Say "never invent" for any ID, like a policy number, and ask for null when it is missing.
  • A test inside ChatGPT is a rehearsal. Run the same emails through your real AI step.
  • Read the results yourself. In the real run a basement flood came back not urgent.
  • Words like "Thursday" mean nothing without the date the email arrived. Pass the received time into the prompt.

Next business case

Still deciding what to automate? Turn a manual process into an automation plan first. Something broke later? Read an automation error and fix it. To build the routing flow itself, see the [n8n beginner course](/courses/n8n-beginner) or [AI automation with Make](/courses/ai-automation-with-make).

Try it yourself, free

Reading the steps is a start. Doing them is how it sticks. The interactive module walks you through every click in a practice copy of ChatGPT, checks each step, and never touches your real accounts.

Start “ChatGPT: Write and Test the Prompt for an AI Step in n8n, Make or Zapier”

Step by step: ChatGPT: Write and Test the Prompt for an AI Step in n8n, Make or Zapier

All 20 steps of the interactive module, in order. In the module you watch each one, then do it yourself in a practice copy that checks your work.

1. Start from real emails

  1. Read the six sample emails. About 90 emails a day land in one shared inbox, and one person forwards each to a desk. Five desks, six real-looking emails. Email 6 is the tricky one: water damage and a billing question. Read them, then click ✕.

    About ninety emails a day land in one shared inbox, and one person forwards each one to a desk. Five desks, six sample emails. Email six is the tricky one. Water damage, and a billing question. Read them, then close the file.

  2. Turn on Temporary chat. Client emails have names, phone numbers and policy numbers. Click the Temporary chat icon at the top right so they stay out of your history.

    Client emails have names, phone numbers and policy numbers.

  3. Open the + menu. Click the + next to the message box.
  4. Add photos & files. Click Add photos & files.
  5. Pick sample-inbox-emails.txt. Pick sample-inbox-emails.txt.

    Pick the sample emails file.

  6. Open it. Click Open. The prompt will be written for these desks and these emails.

    The prompt will be written for these desks, and these emails.

2. Ask for the AI step prompt

  1. Click the message box. An AI step inside an automation must answer the same way every time, so the next step can read it. Spell out the exact fields, the allowed desk names, what to do with a two-desk email, and never invent a policy number. Click the message box.

    An AI step inside an automation has to answer the same way every time, so the next step can read it. Spell out the exact fields, the allowed desk names, what to do with a two desk email, and never invent a policy number.

  2. Ask for the prompt. Paste in the request. It asks for JSON only with five fields (desk, urgent, policy number, one-line summary, confidence) and the prompt in a code block you can paste.

    This asks for JSON only, with five fields, and the prompt in a code block you can paste into your tool.

    I'm adding an AI step to an automation (n8n, Make or Zapier) that reads each new email in our agency's shared inbox and routes it to a desk. The attached file lists the desks and six sample emails. Write the prompt for that AI step. It must: return only JSON with the fields desk, urgent (true/false), policy_number, one_line_summary and confidence (high/medium/low); use only the five desk names; say what to do when an email fits two desks; and never invent a policy number. Show me the prompt in a code block I can paste. Keep everything outside the code block short.
  3. Send it. Click Send (the arrow on the right).

3. Read the rules it wrote

  1. Read the desks and the two-desk rule. Exactly one of five desk names, never two. For mixed emails, rule 3 puts an actual loss (CLAIMS) ahead of a billing question, and rule 5 keeps the second issue in the summary. Click the rules.

    Exactly one of five desk names, never two. For mixed emails, a real loss goes to claims first, ahead of a billing question. And the second issue goes in the summary, so it isn't lost.

  2. Read the policy number rule. Only a number that is really in the email, copied exactly. Not a phone number or dollar amount. If there is none, it returns null, so the next step sees "no number" instead of a made-up one. Click the rule.

    Only a number that is really in the email, copied exactly. Not a phone number, not a dollar amount. If there is none, it returns null, so your automation sees no number, instead of a made up one.

  3. Read the output rules. ONLY valid JSON, exactly five fields, no extra words or code fences, so your automation can read it directly. The three {{email_...}} placeholders at the end are where your tool maps in the sender, subject and body. Click the output rules.

    Only valid JSON, exactly five fields, no extra words, so your automation can read it directly. The three placeholders at the end are where your tool puts in the sender, the subject and the body.

4. Test it before you trust it

  1. Click the message box. Never paste a prompt into a live inbox untested. Have ChatGPT run it on all six emails first. Click the message box.

    Never put a prompt on a live inbox untested. Have ChatGPT run it on all six emails first.

  2. Ask it to test its own prompt. Paste in the request. ChatGPT acts as the AI step, shows the JSON for each email and a summary table, says which emails a person should still see, and fixes the prompt if a result shows a weakness.

    This has ChatGPT act as the AI step. It shows the result for each email, a summary table, which emails a person should still see, and fixes the prompt if it finds a weakness.

    Now test it. Act as that AI step and run your prompt on all six sample emails. Show the JSON for each, then a table: email, desk, urgent, confidence. Tell me which emails a human should still look at and why, and change the prompt if any result shows a weakness. Keep the explanation short.
  3. Send it. Click Send (the arrow on the right).
  4. Read the results table. Six emails, six desks, and no policy numbers invented. Email 1 (the truck) is the only one marked urgent. Email 6 went to CLAIMS with Medium confidence. Click the table.

    Six emails, six desks, and no policy numbers made up. The truck accident is the only one marked urgent. Email six went to claims, with medium confidence.

  5. Check email 6 yourself. Two inches of water in the basement since Sunday night, ruined carpet, a request to call, and it came back "urgent": false. Would you want that flagged at your agency? Probably yes. This is why anything that is not high confidence should go to a person. Click the JSON for email 6.

    Two inches of water in the basement since Sunday night, and a request to call. It came back not urgent. Would you want that flagged at your agency? Probably yes. That's why anything that isn't high confidence should go to a person.

  6. Read which emails a person should see. Three: email 6 (two desks), email 3 (the Thursday deadline, and the prompt cannot see when the email arrived) and email 4 (charged twice, two months running). Click the list.

    Three of them. Email six, because it needs two desks. Email three, because of the Thursday deadline, and the prompt can't see when the email arrived. And email four, charged twice, two months in a row.

  7. Read the two weaknesses and the new rules. It found two weaknesses, relative deadlines and two-desk follow-up, and wrote replacement routing and urgency rules. They use the received timestamp and keep the same five fields. Paste them over the old rules. Click the new rules.

    It found two weaknesses. Deadlines like Thursday, and two desk follow up. It wrote new routing and urgency rules, using the time the email arrived, and keeping the same five fields.

  8. Remember what this test was. Its own words: "These are simulated AI-step outputs, not results from a live automation." ChatGPT pretended to be the AI step. Before you trust the prompt, paste it into your tool’s AI step and run the same six emails there. Then send anything that is not high confidence to a person. Click the line.

    In its own words. These are simulated A I step outputs, not results from a live automation. ChatGPT pretended to be the AI step. Before you trust it, paste the prompt into your own tool, and run the same six emails there. Then send anything that isn't high confidence to a person. You wrote and tested an AI step prompt before it touched a real inbox.

Frequently asked questions

What should the prompt for an AI step in n8n look like?
Give it a role, the exact list of allowed values, rules for the hard cases, and an output section that asks for only valid JSON with named fields and no extra text. End with placeholders for the data your tool maps in, like the email's sender, subject and body. The same shape works in Make and Zapier.
Can ChatGPT write a prompt for my automation's AI step?
Yes. Attach real sample inputs, list the fields and allowed values you need, and say what to do in the hard cases. Ask for the prompt in a code block so you can paste it into your tool.
How do I test an AI step prompt before it goes live?
Ask ChatGPT to act as the AI step and run the prompt on your sample inputs, then read every result yourself. That test is simulated, so run the same samples through the AI step in your own tool before switching it on.
How do I stop the AI from making up a policy number or ID?
Tell it to extract the number only when it is written in the email, copy it exactly, and return null when there is none. In the real run no policy number was invented, but check every test result anyway.
Which emails should a person still check after AI routing?
Anything below high confidence, anything that fits two desks, and anything with a deadline the AI cannot date. In the real run that meant the flood with a billing question, the certificate due Thursday and the customer charged twice.

Start this module free

More in ChatGPT for Everyday Work

Google Docs meeting transcript with the Edit menu open and Select all ready to copy the textSummarize a Meeting Transcript with ChatGPT

Paste a long meeting transcript into ChatGPT and get the decisions, an action items table with owners and due dates, and a follow-up email ready to send.

Free · Live ~9 min Beginner

Stepthrough lesson with a Gmail customer thread about a birthday cake order selected for copyingReply Faster: Turn a Messy Customer Email Thread into a Clear, Friendly Reply

Turn a messy customer email thread into a clear, friendly reply: give ChatGPT the thread and your facts, check every promise, fix the tone, and send it from Gmail.

Free · Live ~9 min Beginner

ChatGPT Create project dialog for Brightwater Cleaning with the Project-only memory optionHow to Create a Project in ChatGPT for Your Brand Voice and Files

Set up a ChatGPT Project with your brand voice instructions, your price list and your style guide, so every new chat already knows your business.

Free · Live ~9 min Beginner

Windows Open dialog choosing a garden center sales CSV file to upload into ChatGPTAnalyze a Spreadsheet with ChatGPT: Upload a CSV, Get a Chart and Three Takeaways

Upload a CSV of sales to ChatGPT, have it check the data first, ask your questions, get a chart you can download, and finish with the three takeaways that matter.

Free · Live ~10 min Beginner

ChatGPT with Work selected at the top, the heading What should we work on? and the Work on anything boxChatGPT Work: Have ChatGPT Do a Multi-Step Task

Switch ChatGPT to Work, hand it a real job (compare three suppliers into a spreadsheet), approve the sites it opens, skip a login, answer its question, check the file…

Free · Live ~8 min Intermediate

◷Get a Morning Invoice Summary in Slack using ChatGPT Dots

Set up a ChatGPT dot to check Gmail for new invoice emails and post one summary to Slack every weekday morning. A 7-step practice based on OpenAI's published screens.

Free · Live ~4 min Intermediate

All 28 modules in ChatGPT for Everyday Work

More ChatGPT tutorials

Plain-English definitions, each with an example from a free lesson.

The full AI glossary

Browse related tutorials