Who here has used an AI tool?

Gemini ChatGPT Claude …and more

What did you make of these tools?

Who has asked an AI for something and got back something useless or low-quality?

AI is a mirror.

The quality of what you get depends on the quality of what you ask.

Picture the smartest assistant in the world. It has read every book… but it started at your company ten seconds ago.

It doesn't know how you want the document.

It doesn't know where to look for the information.

It doesn't know who it's talking to.

It doesn't know why you need it.

It'll do everything perfectly… if you tell it perfectly.

And it'll ruin everything if you don't.

And like any new hire who doesn't want to look bad… it sometimes invents the answer with total confidence.

So the skill isn't “using AI”.

It's knowing how to ask.

EXERCISE · ROUND 1
CLASSIFIED
“Generate me a report of this data.”
EXERCISE · ROUND 1
Download the file here [link] →
Paste your report into an email to pablosaitests@gmail.com
Subject: AWT1 - [Your Name]
In the body of the email, not as an attachment · Tool: Claude
EXERCISE · ROUND 1

You have a few minutes. When you're done, send your email.

AWT1
LIVE

Round 1 — results

[ project Cowork artifact · /grade 1 ]

It wasn't your fault. It was missing what only you knew.

Wrong assumptions — things you assume it already knows.

Don't assume it knows your context. ★ the most important

Don't assume it knows your process.

Don't assume it understands the task.

Don't assume it tells you the truth.

What a new hire needs fits in five letters

C R A T R

CRATR — the five things that turn a lost intern into an expert.

C Context what it's looking at, and why
R Role who you want it to be
A Audience who it's for
T Task the job, made specific
R Result the shape you want back
C R A T R
C — Context
what it's looking at, and why
WEAK
“this data”
STRONG
“a CSV with my company's monthly performance by channel: revenue, marketing spend, support tickets, satisfaction and returns.”
DECLASSIFYING…
C
R
A
T
R
C R A T R
R — Role
who you want it to be
WEAK
(no role — answers like a generic assistant)
STRONG
“Act as a skeptical data analyst.”
DECLASSIFYING…
C
R
A
T
R
C R A T R
A — Audience
who it's for
WEAK
(no reader — writes for no one in particular)
STRONG
“For a manager deciding where to invest next quarter.”
DECLASSIFYING…
C
R
A
T
R
C R A T R
T — TaskTHE BIGGEST LEVER
the job, made specific
WEAK
“generate me a report”
STRONG
“Don't summarize the business. Audit the file: how reliable is the data, what's missing, what's suspicious, and what decisions can it NOT support?”
DECLASSIFYING…
C
R
A
T
R
C R A T R
R — Result
the shape you want back
WEAK
(no format — usually a wall of text)
STRONG
“Return it as a list of findings, each citing the column or cell, and end with the single most valuable column that's missing.”
BRIEF DECLASSIFIED
C
R
A
T
R
WEAK · ROUND 1
CLASSIFIED
“Generate me a report of this data.”
STRONG · ROUND 2 · FULL CRATR
DECLASSIFIED
[C] “I have a CSV with monthly performance by channel: revenue, marketing spend, support tickets, satisfaction and returns.”
[R] “Act as a skeptical data analyst.”
[A] “For a manager deciding where to invest next quarter.”
[T] “Don't summarize the business. Audit the file: how reliable is the data, what's missing, what's suspicious, and what decisions can it NOT support?”
[R] “Return it as a list of findings, each citing the column or cell, and end with the single most valuable column that's missing.”
LIVE · DEMO

Same file. Same AI. Only how I asked changed.

Iterate.

The first answer is a draft, not a verdict.

“make it shorter” “cite the column” “what do you need to be sure?”
EXERCISE · ROUND 2
CRATR
Don't summarize the business. Audit the file itself: how reliable is this data, what's missing or suspicious, and what decisions can it NOT support? End with the single most valuable column to add.
EXERCISE · ROUND 2

Write your prompt with CRATR. When you're done, send your email.

AWT2
WARNING 01 / 05

Don't give AI anything you wouldn't tell a stranger — or a competitor.

If you wouldn't put it in an email to a stranger, don't paste it. Anonymize, or use only approved tools.

WARNING 02 / 05

AI invents data with total confidence. Always verify.

Every figure, date or quote: check it against the source before you use it.

WARNING 03 / 05

Sounding convincing doesn't mean it's true.

Don't confuse confidence with correctness. A polished text can be wrong too.

WARNING 04 / 05

You are 100% responsible for what the AI produces. Always.

If you sign it, you reviewed it. No exceptions.

WARNING 05 / 05

It tends to agree with you. Don't confuse it agreeing with you being right.

Ask it to push back, not just to confirm.

Brilliant intern.
Not an oracle.

LIVE

Before / After

The only thing that changed was how you asked.

[ project Cowork artifact · awt-resultados · /grade 2 ]

You didn't improve the AI. You improved how you ask.

AI is a mirror.

One thing: give it context. It's the difference between a lost intern and an expert.

Guide + examples [link to the site] →