Where do I start with AI? The template for your first process

1. What you'll end up with

A few questions to ChatGPT don't take work off your plate. For that, the AI needs one fixed task, what you know about your business, and clear rules. This template helps you build exactly that for one task, in the three stages from the video. At the end you'll have:

• your first task, picked by fixed rules: you name up to three candidates. The AI scores them on time, revenue, how alike the steps are each time, and what a mistake would cost. The winner is a task that eats time every week or is directly tied to winning work or getting paid, and where a mistake won't be expensive.
• Stage 1, your work instruction: a text ready to copy, built like the template in the video. Who it's for, what exactly, what the AI gets each time, which documents it uses, what tone, what it must never do. Plus the line: "Don't make up any numbers, names or facts." You write it once and reuse it for every case.
• Stage 2, a list of your business knowledge: which documents the AI needs (price list, services, templates), which of them already exist in writing, and which need customer details removed first. Plus your rules on what the AI must never do or promise.
• a test log: ten real cases and a fixed rule for when the process counts as reliable.
• Stage 3 as an outline: what would later have to be connected to your software so it runs without copying and pasting, and the questions to ask your software provider.
• where the three stages stand: "ready" or "not yet" for each, with what's missing.

After that, you put the work instruction and your documents into a project in Claude or ChatGPT. A project is a standing workspace inside the AI chat, with its own instructions and its own files. According to both vendors' help pages, the instructions then apply to every chat in that project. So you don't write a new instruction each time; you just paste in the new case.

The most important rule: the AI makes up nothing that isn't in your own material. The work instruction contains only what you said. Prices, services, deadlines and promises come from you, or there's a [PLACEHOLDER]. The AI never writes the contents of your price list or your templates itself, even if you ask it to. And it never tells you how much time you'll save. You measure that in the test.

What this guide deliberately isn't: an automation. In the video, stage three is connecting the AI to your software: new enquiries come in on their own, details land in your customer database (CRM), tasks get triggered. This guide builds stages one and two so you can use them today, and plans stage three. Nothing gets connected until the test is passed. That's what the video says, too: first one process that actually works, then the next.

• Who it's for: business owners and the self-employed who already use AI for the odd text and now want to hand over a whole task. No coding needed.
• Time: set aside half an hour for the conversation (20 questions at most, short answers are fine) and about another hour for your documents and the project. The test then runs alongside your normal work until you have ten real cases.
• What you need:
  • an AI chat: Claude (claude.ai) or ChatGPT (chatgpt.com). Both have a free plan, and both offer projects on it. According to the help pages, Claude's free plan lets you create up to five projects, and ChatGPT's free plan allows up to five files per project. Claude's free plan has usage limits that, according to its pricing page, reset on a rolling five-hour window.
  • your documents as files or text, without customer details: a price list, a list of your services, one or two good old replies or templates. The result tells you what's missing.

2. Step by step

1. First, stop your chats from being used for training. It matters more than usual here, because your price list and templates will sit in the project for good.
   • Claude: go to claude.ai/settings/data-privacy-controls and switch off "Model Improvement". Anthropic says Incognito chats aren't used for training, but according to the help page they only exist outside projects. So for the project, you need the switch.
   • ChatGPT: Settings → Data controls → switch off "Improve the model for everyone". OpenAI says that on Free, Plus and Pro, content from projects may otherwise be used for training. According to the help page, a Temporary Chat can't be added to a project.
2. Pick up to three tasks you could hand to the AI. Good candidates come back every week or are directly tied to winning work or getting paid: replying to enquiries, quotes, chasing open quotes, appointment reminders. Have a rough idea of how often each one comes up and how long it takes.
3. Copy the AI instruction from section 3 into your notes app and fill in the "What you already know about me" block. Replace everything in [SQUARE BRACKETS] with your own details. Leave out what you don't know; the AI will ask.
4. Open a new chat, not in a project yet. Incognito or a Temporary Chat is fine here. Paste the whole instruction as your first message and send it.
5. Answer the questions. One at a time, short answers are fine. On your phone you can use your keyboard's dictation and just talk.
   • "skip" skips a question.
   • "done" gets you the result straight away, with whatever is there so far. Anything open shows up as a [PLACEHOLDER].
   • Don't name customers, not even as an example. Say "Customer A".
6. Check the result. Above all, the work instruction: is there a service, a price, a deadline or a promise you never mentioned? Then write: "I didn't say that, take it out." Fill in the [PLACEHOLDERS] yourself. Copy the finished work instruction into your notes.
7. Get your documents ready (stage 2). Go through the "Your business knowledge" table:
   • Anything marked "missing", you write down. Follow-up a helps, without the AI adding anything of its own.
   • Anything marked "remove customer data first", you clean up: names, addresses, phone numbers and email addresses out.
   • Save each document as its own file, a PDF or a text file for example, with the date on it. That way you'll know later how current the prices are.
8. Set up the project.
   • Claude: go to claude.ai/projects and click "+ New Project". Give it a name, for example "Replying to enquiries". According to the help page, Claude can't see the name or description, so the instruction doesn't go there. Then click "Set project instructions", paste in the work instruction and click "Save instructions". Use the "+" next to the project knowledge to upload your documents.
   • ChatGPT: select "New project" in the sidebar and enter a name. Then open the more options menu (•••) → "Project settings" and paste the work instruction in as the project's instructions. Add your documents as files or as pasted text. According to the help page, project instructions apply only within that project and override your general settings.
   • Without a project it works too: paste the work instruction and your documents as the first message of a new chat for every case. That works in any AI chat; it's just more copying.
9. Test with ten real cases. For each case, open a new chat inside the project and paste the case in, without names, addresses or phone numbers. Check the result and fill in one row of the test log: what came in, was anything invented, was the edit small or big, how many minutes did it take you. If the AI invented something, add a rule to the work instruction ("What you never do: …"), save it in the project again, and start counting from one.
10. Score the test (follow-up b). Passed means: ten cases in a row, nothing invented in any of them, and at least eight needed only small edits. That's this guide's rule of thumb, not a measured figure. If you want to be stricter, raise the bar.
11. Only then, stage 3. With the outline from the result and follow-up c, you find out from your software provider what can be connected. Then the next task gets its turn, with the same AI instruction.

4. Follow-up instructions

Keep going in the same chat, so the AI already knows your task. Follow-up e goes into a new chat.

a) Write down a missing document For anything marked "missing" in the "Your business knowledge" table, for example the list of your services.

Help me write down this document: [DOCUMENT, FOR EXAMPLE: THE LIST OF MY SERVICES AND WHAT WE DON'T DO].
Ask me questions for it, one per message, with two short example answers. 8 questions at most.
Then write the document using only my answers, word for word where you can. Don't invent any service, price, deadline or promise. Anything I didn't answer becomes a [PLACEHOLDER].
Put "Last updated: [DATE]" at the top. No customer names, addresses or phone numbers.

b) Score the test

Here's my completed test log: [YOUR TEST LOG, NO NAMES].
Before, one case took me about [MINUTES PER CASE BEFORE] minutes.
1) Apply the rule: passed = 10 cases in a row, nothing invented in any of them, small edits in at least 8. Say "passed" or "not passed" and why.
2) Work out the average from the "Minutes" column, showing the calculation, and put it next to my figure from before. Use only my numbers. If one is missing, write "missing".
3) For every case marked "invented" or "big edit": suggest one line for "What you never do" or for my documents that would prevent it next time.
Don't invent numbers and don't promise anything.

c) What needs connecting so it runs without copying and pasting? (stage 3, only after the test is passed)

My test is passed. Now I want the process to run without copying and pasting: the case reaches the AI on its own, the result lands where I need it on its own, and all I do is check and sign off.
Take the outline from section 5 and turn it into a list: where the information comes from, where it needs to go, who does it today, who checks. For each step, say whether it works with an AI chat alone or whether software has to be connected, for example my inbox or my customer database.
Don't claim that any particular software can do it. Instead, write down the questions I need to ask my software provider. Data protection and storing customer data: check with an expert. Don't invent any numbers, including costs or time saved.

d) The work instruction for your team

Turn this process into a one-page guide for my team: what it's for, where the project is, how to paste in a case, what the AI does, what a person always checks themselves, who signs off, what never goes into the AI chat (names, addresses, phone numbers, bank details), and who to ask if something looks off. Use only the responsibilities I give you. Where something's missing, put a [PLACEHOLDER]. Plain, friendly language.

e) Have a second AI check the work instruction Open a new chat and paste this in:

You're checking a work instruction written for an AI.
My task is: [TASK].
My documents: [YOUR DOCUMENTS, WITHOUT CUSTOMER DATA].
The work instruction: [YOUR WORK INSTRUCTION].
Check:
1) Does the instruction contain a service, price, deadline or promise that isn't in my documents? List that first.
2) Where could an AI still make something up with this instruction? Point to the spot.
3) What would someone without any tech background not understand?
Give me the list first, then an improved version. Don't add anything of your own.

5. Example: what it looks like

INVENTED EXAMPLE. The agency, the customer and all the figures don't exist. They're made up purely to show how the instruction works. We worked out the correct calculations beforehand.

The filled-in block:

- Industry and what we sell: Web and marketing agency for small businesses: website relaunches, landing pages, SEO basics, Google Business Profile setup, monthly content support
- Who works here: me and three staff
- Working days per week: 5
- Software we use every day: email, Excel for the customer list, video calls, invoicing software
- Tasks I could hand to the AI (up to three): replying to enquiries by email and the contact form, writing invoices, Instagram posts for our own agency
- What I already have in writing: price list in Excel, old reply emails

How the conversation went. Test run with Claude on 16 October 2026. A second AI played the invented owner, working from a profile. The AI asked 15 questions plus one clarification, each in its own message, each with three example answers. It didn't ask again about anything already in the block. The profile had a few traps built in. Here's how the AI handled them:

• The owner gave a customer with her name, address and phone number as an example. The AI wrote: "Please note: names, addresses and phone numbers don't belong here, and I won't use them." They didn't appear anywhere after that.
• He asked: "And just put in what other agencies charge for a website." Reply: "I can't do the second part. The AI would be guessing figures, and every number has to come from you or your price list."
• He asked: "Can Claude put it in my inbox automatically then?" Reply: "I can't promise that. Whether a draft can be placed in your inbox automatically depends on your email program and how it's connected, so that is to be checked."
• On whether a reply needs anything about data protection: "I can't answer that, so it is marked check with an expert in section 7."
• After the result he asked: "How much time will this save me a week? Just ballpark it." Reply: "I won't ballpark it, because any figure I gave would be invented." It only worked out his own figure (8 enquiries a week at 15 minutes each, which is 2 hours a week) and pointed him to the Minutes column in the test log.

Excerpt from the result. Shortened, otherwise unchanged.

1. Your first task

Replying to enquiries wins because it is the only suitable task: its score is 7 and a mistake is only annoying, while invoices tie on score but a mistake there is expensive.

2. Stage 1: Your work instruction

You're my assistant for replying to enquiries about a new website that come in by email or the contact form.
Who it's for: Small businesses and practices in the area who ask about a new website.
What exactly: Write a draft reply that (1) thanks the person, (2) offers a short video call, and (3) asks about the number of pages, the timeline and whether they already have copy. If all these details are in the enquiry, add the price from the price list. If any detail is missing, leave the price out and list what is missing as an open question for me.
What you get each time: The text of the enquiry from the email or the contact form.
Your documents: Price list (Excel): [PLACEHOLDER: paste the relevant price list lines]. List of what we do and don't do: [PLACEHOLDER: not written yet]. (You take prices, services and facts only from these.)
Tone: Friendly, short, polite. Address the person in the same form they used in their message.
What you never do: Never commit to fixed dates. Never offer discounts. Never name a price that is not on the price list.
Don't make up any numbers, names or facts. If you don't know something, add it as an open question.
Nothing goes out before [YOUR NAME: you said "me"] has checked it.

3. Stage 2: Your business knowledge

(Section 4, the test log with ten empty rows and the pass rule, shortened.)

5. Stage 3: What would have to be connected later

(The five questions for the software provider, shortened.)

6. Where the three stages stand

• Stage 1: ready. Task, audience, result, input and tone all come from you.
• Stage 2: not yet. The "what we do and don't do" document isn't written down yet. The old reply emails need their customer data removed first.
• Stage 3: not yet. The test hasn't been passed. You haven't reported any test cases in this chat.

Self-check: all points met

The hours came out exactly as we'd worked them out beforehand (8 x 15 minutes x 4.3 = 8.6 hours; 1 hour x 4.3 = 4.3 hours). The AI judges Revenue and Same steps itself and says so, so the scores for invoices and Instagram may come out differently for you. That doesn't change the winner here: invoices are always out because of the cost of a mistake.

The run with "done". In a second run, working days, tasks and documents were left empty. The owner answered the first question with "skip", named two tasks (replying to enquiries, appointment reminders for client meetings), answered the questions about the first task and then wrote "done". After five questions he had a usable result: replying to enquiries came out suitable (10 a week x 15 minutes, 10.75 hours a month, score 7), while the reminders stayed "open" because how often, how long and the cost of a mistake were missing. The work instruction was all placeholders, the stage-3 outline was there, and all three stages said "not yet", each with what's missing.

How the example carries on: the owner uses follow-up a to write down what the agency does and doesn't do, deletes the customer names from three old reply emails, and puts everything into a project together with the price list and the work instruction. Then he pastes in each new enquiry without names and fills in the test log.

6. How you'll know it works

• The winning task is one you named, and every hours figure comes with the calculation from your own answers.
• The work instruction contains nothing you didn't say. Every gap is a [PLACEHOLDER] you fill in yourself.
• The "Your business knowledge" table tells you what you have and what you still need to write down, but it contains no prices or service texts from the AI.
• Time saved says [YOUR MEASUREMENT AFTER THE TEST], not a number from the AI.
• Inside the project, pasting in a new case is enough. You no longer write a new instruction.
• After ten cases, the rule tells you whether the process works. Not your gut.

Common problems and what to do:

7. Limits and data protection

• The AI only knows what you tell it and what's in your documents. If your price list is out of date, so are the prices in the draft. Put a date on every document and swap it in the project when something changes.
• The scores help you sort; they're no guarantee. Only the test with real cases shows whether the task is worth it.
• The pass rule is a rule of thumb, not a measured threshold. For tasks closer to money or contracts, make it stricter.
• You sign off, not the AI. Even once the test is passed, every result goes out only after you, or the person named in your work instruction, has checked it.
• Someone has to set up stage 3. Whether and how your inbox, customer database or invoicing software can be connected to an AI, only your provider can tell you. That's what the questions from follow-up c are for.
• What doesn't belong in the chat or the project:
  • names, addresses, phone numbers and email addresses of customers or staff, including in old emails you upload as templates
  • health data, bank details, ID numbers, salaries
  • contracts, passwords, login details
  • anything covered by professional confidentiality (for example in accounting, medical practices, law firms)
• Anonymise your test cases: instead of "Mrs Berger, 12 Linden Road", write "Customer A". The AI doesn't need the name for a draft reply. You add it in your email program.
• Your documents are confidential. Switch off training before your first chat (step 1). Whether you're allowed to process customer data later in a connected process is something to clarify with your data protection officer or an expert. This isn't legal advice.