# Teach It Once: hand AI a weekly job properly, one time

From tuitiv's free resources: tuitiv.ai/free/once

## The maths

For one job you do every week. Swap in your own numbers.

1. **Doing it yourself: 30 minutes, every week, forever.** And you're the only one who can do it, so it waits for you.
2. **Teaching it once: An hour, one time.** Then five minutes a week to read what it made and fix anything that's off.
3. **By week three: You're ahead.** Over a year, that's more than 20 hours back. For one job.
4. **Why it failed before: It didn't know what you know.** Who it's for, what good looks like, what you'd never write. It guessed, and it guessed like a bank.

## Pick the right job

Your first one should tick all four.

- **It comes round every week**, or more often. Proposals, client follow-ups, the Monday update, session notes.
- **It takes you 30 minutes or more.**
- **You've done it at least ten times**, so you'd know a good one when you saw it.
- **You have one you're proud of.** That's your example, and it does half the teaching.

## Your hour, in six steps

1. **Pick the job** and find the best one you've ever done. Take out client names before you paste anything in.
2. **Run prompt 01.** It asks one question at a time. Answer by voice if you can: you'll say more than you'd type, and the extra is where the useful detail is.
3. **Run prompt 02 with your best example.** It adds what made that one good to the instructions.
4. **Test it on this week's real job.** Paste the instructions into a fresh chat with this week's details, and see what comes back.
5. **Fix the instructions, not the draft.** Use prompt 03 every time something's off. That's how it stops needing you.
6. **Save it where you'll use it.** In Claude, prompt 04 turns it into a skill you run with a forward slash. Anywhere else, save the instructions in a project or a document and paste them in each week.

## The prompts

01 and 02 do the teaching. 03 is how it gets better. 04 makes it permanent in Claude.

### 01. Interview me about the job

Start a fresh chat. Name the job in the first line.

```
I want to teach you a job I do every week: [the job, for example: writing a proposal after a sales call]. By the end, I want instructions good enough that you could do it next week, in a fresh chat, without me explaining anything.

Interview me. Ask one question at a time and wait for my answer. Cover:
→ who it's for and what they need from it
→ what I'll hand you each time (notes, a transcript, a spreadsheet) and what you should do if something's missing
→ the steps I go through, including the ones I do without thinking (if I can't list them, ask me to walk you through the last one I did)
→ what a good one looks like, and what makes one bad
→ the rules: length, tone, things I'd never say or do
→ how I check it before it goes out

No more than ten questions, including follow-ups. If an answer is vague, ask once for a real example rather than filling the gap yourself. If I mention a client by name, use it to understand me but keep it out of the instructions.

Then write the instructions, headed "How to do [the job]". Start with two lines on who I am and what I do. Use "you" for the AI that will follow them and "I" for me. Put anything that changes each time (names, prices, dates) in a section called "What I'll give you each time", not in the steps. Use only what I told you. If something's missing, list it at the end under "Not covered yet" rather than inventing it.

Show me the draft and ask me what's wrong or missing before you call it finished.
```

### 02. Learn from my best example

In the same chat, straight after prompt 01.

```
Here's the best [job] I've ever done: [paste it, with client names taken out].

Work out what makes it good: the structure, the order, the length, the tone, the phrases that do the work. Compare it with the instructions you just wrote. What's in this example that the instructions don't capture?

Update the instructions with anything missing, and show me what you added. Don't copy the example's content into the instructions. Capture how it's done, not what it said.
```

### 03. Fix the instructions, not the draft

Every time it gets something wrong. This is the step that stops you redoing it yourself.

```
Here are the instructions: [paste]. Here's what you produced this week: [paste]. And here's what I changed before I used it: [paste my version, or describe the changes].

Work out why you got it wrong. Was something missing from the instructions, unclear, or did you ignore it?

Update the instructions so next week's version comes out the way I'd have done it. Show me only the lines you changed. Don't rewrite the parts that worked.
```

### 04. Turn it into a skill

In Claude, once the instructions have worked for two weeks running.

```
Turn these instructions into a skill called [job-name]: [paste the instructions].

When I run it, it should ask me for this week's details before it starts, follow the instructions exactly, and tell me anything it had to assume. Keep my rules about tone and what never to say word for word.
```

## The quieter reason

Underneath "it's quicker if I do it myself" there's usually something else: if I do it, I know it's right. That's fair. It's also how you become the only person who can do it, and then everything waits for you. Teaching it once is the same move good leaders make with a new team member. The instructions are the handover.

## When it's not quite right

**It asked all its questions at once**
Reply: "One question at a time, please. Start again from question one." It will.

**The first draft still sounded generic**
The instructions are missing your rules on tone. Run prompt 02 with your best example, and add three things you'd never write.

**I keep fixing the same thing every week**
You're fixing the draft. Use prompt 03 so the fix goes into the instructions and stays there.

**It made up details I didn't give it**
Add to the instructions: "If a detail is missing, ask me for it. Never invent a name, number or date."

Human-led. AI Amplified. · tuitiv.ai