DEXIUM/LABS

The prompts from the videos

Hidden Layer

The hidden AI inside famous brands, and how to steal it for your business.

File 01 · Duolingo

OWL

Five versions of one reminder

Run your own bandit: test, keep the winner, bench the rest.

Paste this into ChatGPT

You are a copywriter for a small business. Write 5 versions of ONE customer reminder.

My business: [what you sell, in one line]
Who gets it: [e.g. past customers who haven't booked in 60 days]
The one action I want: [e.g. book a visit, reorder, finish checkout]
Channel and length: [e.g. SMS under 160 characters / email subject + 2 lines]
My voice: [e.g. warm, a little dry, no hype]
Facts you may use: [real offers, real deadlines, real stock]. Use nothing else.

Rules:
1. All 5 versions ask for the same single action. Only the angle changes.
2. Use a different angle for each: friendly nudge, specific benefit, light humor, honest urgency (only from the facts above), short question.
3. Do not invent discounts, deadlines, scarcity or testimonials.
4. Stay inside the length limit.

Output a table: Label (A to E) | Angle | Message. Do not pick a winner. I will test them.

How to use it

  1. Fill the brackets. Pick one reminder and the one action you want (a booking, a reorder, a finished checkout). Replace each [bracket] and paste the prompt into ChatGPT.

  2. Send each version to a small slice. Split your opted-in list into slices, one per version (for example 5 slices of 10%), and send A to E at the same time. Count the action, not just opens.

  3. Keep the winner, bench the rest. Send the winner to everyone left. Bench the others (rest them, do not delete them). Next week, add one fresh version and run it again.

Download the PDF

Example (made-up business)

The brief

Maple Mobile Grooming (example). Dog owners not booked in 8+ weeks. Action: book a groom. SMS under 160 characters. Voice: warm, a little silly. Facts: 4 Saturday slots left this month; standard groom $65.

What comes back

A · Nudge: "Your dog is due for a groom. Want us to pencil one in? Reply YES and we'll find a time. - Maple Mobile Grooming"

B · Benefit: "A fresh groom means less fur on your couch. $65 standard, and we come to you. Reply YES to book. - Maple Mobile Grooming"

C · Humor: "Your dog called. Says the fluff is getting out of paw. Reply YES to book a groom. - Maple Mobile Grooming"

D · Urgency: "4 Saturday slots left this month. Want one for your dog? Reply YES. - Maple Mobile Grooming"

E · Question: "Ready to book your dog's next groom? Reply YES for a time. - Maple Mobile Grooming"

Then

500 opted-in clients: send each version to 50 people (5 slices, 250 total). Next day, count the YES replies. The winner goes to the other 250.

Run your own bandit: checklist

  • One reminder, one action to measure (click, booking, reorder).
  • Five versions, same ask, different angle. Label them A to E.
  • Send only to people who opted in, and keep an unsubscribe in every message.
  • Send each version to a small slice (for example 10% each). Same day, same time of day.
  • Count the action (bookings, clicks, replies), not just opens.
  • Keep the winner and send it to the rest of the list.
  • Bench the others. Rest them a few weeks, then retest. Do not delete them.
  • Rest the winner too after a few rounds. A message that worked yesterday works less today.
  • Every week: add one fresh version, bench the weakest, run the next round.
  • Log it: version, number sent, number who acted.

Slice sizes and rest times are starting points, not Duolingo's numbers. On a small list, two versions a few actions apart is a tie: run another round before you pick.

File 02 · Netflix

STARS

Find the two parts doing the work

Judge them by what customers do, not what they say.

Paste this into ChatGPT

You are a business analyst for a small business. Help me find the TWO parts doing most of the work, judged by what customers DO, not what they SAY.

Parts to review (channels, products, offers, steps or campaigns, 4 to 10): [list them]
Time period: [e.g. last 90 days]
Result I care about most: [orders / revenue / bookings]
What customers did (numbers for each part): [clicks, opens, orders, revenue, repeat purchases]
What customers said (optional): [survey answers, reviews, comments, favorites]

Rules:
1. Rank every part by behavior only: buys first, then clicks, then opens. Opens are the weakest signal.
2. Treat what customers said as a hint to check, never as proof.
3. Show your math. Use only the numbers I pasted. If a number is missing, tell me what to pull. Do not guess.
4. Name the two parts that account for the most results, and what share of the total they make up.
5. Flag any part where what customers say and what they do disagree.
6. For every other part: keep, test or pause, with a one-line reason.

Output: (1) a ranking table, (2) the two parts and their share, (3) any say-vs-do gaps, (4) one cheap test to run this week on each of the two parts. Do not tell me to build anything new yet.

How to use it

  1. Gather what customers did. One row per part (channel, product, offer, step). Add what each one produced over the same period: clicks, opens, orders, revenue. A spreadsheet export is enough.

  2. Run the prompt. Paste it, fill the brackets, and add any survey answers or reviews under "said". They are hints, and the prompt treats them that way.

  3. Check the math, then test. Re-add the numbers on a calculator (ChatGPT can slip on arithmetic). Then run the one cheap test on each of your two parts this week. Re-check next month.

Download the PDF

Example (made-up candle shop)

Numbers pasted in (last 90 days)

Email (welcome + restock alerts): 96 orders, $4,320

Local café wholesale (3 accounts): 9 repeat orders, $3,150

Market stall (6 Saturdays): 52 sales, $1,560

Instagram posts: 31 orders, $1,395

Google Ads: 28 orders, $1,260 (ad spend $900)

Survey: 62% named Instagram as their favorite way to find us.

What comes back

Total revenue $11,685. Your two parts: Email (37%) and café wholesale (27%), together 64%.

Say vs do: 62% say they love Instagram. It brought 12% of revenue.

Google Ads: $1,260 in revenue minus $900 spend = $360. Test, or pause.

This week: add a restock alert for your top 3 candles. Ask each café about a standing monthly order.

File 03 · ChatGPT

SOURCE

Make ChatGPT quote its source

Answer only from this source, and quote the line.

Paste this into ChatGPT

Answer ONLY from the source below. Do not use outside knowledge.
For every claim, quote the exact line from the source in "quotation marks".
If the source does not say, reply exactly: "The source doesn't say." Do not guess.
If the source is unclear or contradicts itself, say so and quote both lines.

SOURCE:
"""
[paste the real document here]
"""

QUESTIONS: [one or more questions]

How to use it

  1. Paste the real document. Your policy, contract, price list or notes, between the triple quotes. The model can only stay inside what you give it.

  2. Ask your questions. Type them under QUESTIONS. Every answer should come with a quoted line, or the exact words "The source doesn't say."

  3. Check the quote. Search your document for each quoted line (Cmd/Ctrl+F). No match, no trust. This prompt lowers made-up answers. It does not guarantee zero.

Download the PDF

Example (made-up candle shop policy)

The source you paste

"Unused items can be returned within 30 days of delivery for a full refund. Opened candles cannot be returned. Gift sets can be exchanged within 14 days. Refunds go to the original payment method in 5-7 business days."

Question 1: Can I return a candle I already lit?

"Opened candles cannot be returned." So no, a candle you have opened is not eligible.

Question 2: Is return shipping free?

The source doesn't say.

Why it matters

Without the rules, question 2 usually gets a confident, plausible answer. With them, you get the gap.