CODED × KJO · Day 1

×

Working With AI
on Real Data

Day one of three. Two dashboards from one file, where the only difference was the prompt, then a file you can defend, and a record of why.

Six hours · thirteen blocks

Arabella Hotel

We’ll begin shortly

Press → to begin

No preamble

I have 2,400 work orders. I want a dashboard.

Watch what I type.

Prompt A

What you’ll walk out able to do

By the end of today, here’s what changes.

Write a prompt that produces an answer you can check, not one you have to trust.

Tell the difference between a tool that is predicting and one that is computing.

Profile a file with six questions before you analyze it.

Keep a change log that still defends your numbers in six weeks.

The shape of the day

Two halves, two different questions.

The morning answers how do I ask for what I want. The afternoon answers why does the answer keep changing, using evidence you produce yourselves before eleven o’clock.

1

Two prompts, one file

Same tool, same five minutes. One makes a picture, one makes something you can click.

2

How a language model actually works

One mechanism explains both the good dashboard and the wrong count.

3

The five-part data prompt

A shape for the five things that follow. Seven lines, under two minutes.

4

Why dirty data breaks an answer

Four defensible readings of forty rows. The winner moves each time.

5

Profile before you analyze · the change log

Six questions, then the one thing that survives this room.

Topic 01 · Story

Nothing here is a mistake.

What Prompt A produced

Chart titles that name fields. “Sum of DowntimeHours by Site”.
Bars in alphabetical order, so the largest sits wherever the alphabet put it.
A count of work orders that includes the duplicates.
Nothing to click. It is a picture, not a tool.

Is this wrong? No, not wrong. Just not useful. That distinction is the whole day.

Demo · same file, same tool, same five minutes

Two dashboards. One difference.

The room · take answers for a full five minutes

Same file. Same tool. Same five minutes.
What was different?

Take every answer, one at a time round the room. Do not rush to the framework.

“The second prompt was longer.”

It was. Does length do it, or does something inside the length do it?

“You must be better at this.”

Everything in that prompt is on the sheet in your hand. By eleven o’clock you will write one.

Block 02 · your turn

Which site had the worst unplanned downtime last quarter?

The file is work-orders-40.xlsx, forty maintenance work orders. Paste it into Copilot Chat or Claude and ask that question however you would normally ask it.

Workbook · page 2

Write down the prompt you actually used.
Write your answer. The site, and the downtime hours next to it.
Do not look at anyone else’s first.

12:00

Timing flexes to the room. The clock is a guide, not a gate.

Demo · the answer board

Fifteen prompts. One file.

Forty rows. One question: which site had the worst unplanned downtime last quarter? Twelve minutes, no guidance. Go around the room. Each person reads theirs out, and it goes up here.

Name the felt problem

Nobody typed a bad prompt.

THE GAP Each of you typed a short prompt, and the tool filled the gaps for you.

It filled them reasonably…

→

reasonably is not the same as correctly.

And the answer does not tell you which one you got. The prompt is the product.

Pause

Three questions before we go on.

Keep that board. We come back to it this afternoon, and again tomorrow. It is the evidence for the whole second half of the day.

What is still unclear?

What would you push back on?

What does this change on Sunday?

Topic 02 · Story

How can both be true?

1

It built a working dashboard in five minutes.

2

Then fifteen of you asked it about forty rows and it gave four different answers.

The mechanism, in one line

It predicts the next piece of text.

Over and over. That is the whole mechanism, and it explains both the impressive dashboard and the unreliable count.

It predicts. It does not look up, and it does not calculate.

Four consequences

What follows from that one line.

01

It is not a search engine.It holds no database you can query.

02

It is not a calculator, unless it uses one.Asked to add a column of pasted numbers, it predicts what the answer looks like. When it writes and runs code, or works on an Excel table directly, it stops predicting and starts computing.

03

Confidence is a writing style, not a signal.The tone is identical whether the answer is right or wrong.

04

It only sees what you give it, and only so much at once.That limit is the context window.

Demo · prediction against computation

Twenty rows of numbers. Ask for the total.

Pasted into the chat

···

As an Excel formula

···

Mini check · your call

Which of these can you trust without checking the arithmetic?

So five things follow

What this changes.

1 Tell it what the work is for, because it cannot infer your intention.

2 Describe your data, because it has never seen it.

3 Define your terms, because it will define them for you if you do not.

4 Make it compute rather than predict, whenever a number matters.

5 Ask it to show its working, because you cannot judge the answer by reading it.

Topic 03 · Story

This does not start a new idea.

It gives the five consequences from the last block a shape. Say that out loud. It is what stops the morning feeling like two lectures.

“What would you add to Prompt A to fix the chart titles?” Every answer the room gives is right. What they do not have is a way to know what is missing before they look at the output.

The term & the link

This is the five-part data prompt.

1

Role“You are a maintenance operations analyst”

Output aimed at operations

2

Data description“Each row is one work order… dates are text in three formats”

It cleaned before it built

3

Task & terms“Unplanned means Emergency or Urgent”

The right rows counted

4

Output shape“Slicers… sorted largest first… titles state the finding”

Every visual thing you liked

5

Verification ask“State the row count and list every row you excluded”

The number at the bottom

Prompt A had one of these. Prompt B had all five. Nothing else was different.

Demo · write one live, from a blank line

Seven lines. Under two minutes.

Question first. Then my terms. Then what one row is. Then what I want back. Then the verification ask.

Mini check · your call

Which part will you drop first when you are busy?

The detail

Three things worth knowing.

02

Part 2 does more work than it looks like.One row here is one work order. Not one asset, not one failure, not one day. Skip it and the tool infers from column names. That guess is reasonable, and you will not see the difference in the answer.

04

Part 4 is where interactivity lives.Prompt B’s slicers exist because a line asked for them.

,

When you cannot define a term yet, say that.“I am not sure whether unplanned should include Urgent. Give me both and label them.”

This answer can still be wrong. The difference is that you can now find out.

Practice · 35 minutes

Same file. Same question. This time all five parts.

Your task

Rebuild your morning prompt using all five parts. Run it. Write your new answer and your new row count on page 4. Then write one sentence on what changed.

If you are stuck Which of the five parts is missing from what you typed?

Still stuck You have not told it what unplanned means. It is guessing.

Finished early Run it again with unplanned meaning Emergency only. Does your answer survive?

Block 06 · your turn

Rebuild it with all five parts.

Same forty-row file. Same question. This time you write the prompt properly, and you run it.

Workbook · page 5

Fill in all five parts on page 5.
Run it, and write your new answer and the row count the tool reported.
Then one sentence: what changed?

15:00

Timing flexes to the room. The clock is a guide, not a gate.

Unlock · the three pitfalls, in this order

Which of these did your prompt catch?

1

The duplicate.WO-1183 appears twice. Most answers counted its downtime twice.

2

The negative value.One row shows negative downtime hours. Most tools included it silently.

3

The name split.“Pump-A1”, “PUMP A1” and “pump a1” are one asset. Totals grouped by asset split three ways.

The correct answer

0hours

Ahmadi Refinery from 24 work orders

You did not get a better answer from a better tool. You got it because you told the tool what a row means and asked it to show its work.

Topic 04 · Story · go back to the board first

I did not tell you the whole reason.

This morning, fifteen of you asked one question of forty rows and produced that board. I told you nobody typed a bad prompt, and that was true.

Four of the differences on that board are not about prompting at all.

Demo · change one setting at a time, change nothing else

Every one of these is a defensible reading.

Name the felt problem · the hinge of the day

None of them are wrong. Only one is what you meant.

This morning’s line was the prompt is the product. That still holds. But a perfect prompt on a file you have not looked at gives you a defensible answer to a question you did not ask.

You cannot write part 2 of a prompt about a file you have not looked at.

The term & the link

So before you analyze anything, you look.

DATA PROFILE A short description of what a dataset contains, written before you analyze it.

Prompt B’s whole second paragraph is a profile…

→

which is why that dashboard cleaned the dates before it built anything.

There is a name for that, and there are six questions in it.

Demo · run all six on the full Work Order Set

Six questions. About three minutes.

It settles the argument you cannot win later, when somebody asks where your number came from.

Two things inside the profile

Where it gets interesting.

Q5

There is a trap inside the duplicates.Matched on the ID they give 20. Matched on the whole row they give 19, because WO-2087 is one ID used for two different work orders. Ask the tool which method it used. If it does not say, it picked one.

Q4

This is where the tool stops being useful.340 blank close dates. Are those work orders still open, or is the data incomplete?

The tool cannot answer that. You can, because you know the business. That gap is your job, and it does not go away with a better model.

Pause

Three questions before the long one.

Next is forty-five minutes on 2,419 rows. Take the break first. You want fresh attention going in, not at the end of it.

What is still unclear?

Which of the six will you actually run?

What would your team argue about?

Practice · 45 minutes · 2,419 rows

Profile it. Then clean it.

Your task

Run the six profile questions on the full Work Order Set. Write the answers on page 7. Then clean the file, recording every change and why on page 8.

If you are stuck Which of the six questions have you not asked yet?

Still stuck Ask it to list every unique value in the asset column. Read that list.

Finished early There are more than four defects in this file. Find one nobody has named yet.

Block 11 · your turn

Profile the full file. Then clean it.

The file is work-orders-full.xlsx, about 2,400 rows. This is the longest stretch of the day.

Workbook · page 7

Run the six profile questions. Answers on page 7.
Then clean the file, and record every change and why on page 8.
If you get stuck: which of the six have you not asked yet?

22:00

Timing flexes to the room. The clock is a guide, not a gate.

There are more than four defects in this file. If you finish early, find one nobody has named yet.

Unlock · surface the cost outlier last

Four findings. The last one changes the day.

1

Dates.Three formats. Ask who found all three.

2

Assets.29 names for 17 real assets.

3

Duplicates.20 by ID, 19 by whole row. Compare the numbers the room found, out loud.

4

The cost outlier.One work order costs about 40 times the next highest.

The room

Do you delete that row?

Let the room split. Both sides have a case.

Delete it

Your average looks sensible again. One row was distorting every cost figure in the file.

Keep it

It may be the single most important row in the file. A shutdown costs what a shutdown costs.

You cannot answer this from the data. You answer it by asking Operations what happened on that job. Write the question down. Do not delete the row today.

WO-2087 · a three-day compressor failure at Ahmadi Refinery · 390,000 KWD, and it is also one of the duplicate IDs. Removing duplicates by ID would have deleted it.

Integration, nothing new is introduced

The five-part prompt and the profile meet on one object.

CHANGE LOG The record of every change you made to a file, and why, including the row count it affected.

Every one of you now has a different cleaned file, because every one of you made different decisions. That is not a problem. Not writing them down is the problem.

Block 12 · your turn

Finish your change log.

Every change, the row count it affected, and why. Nothing new to learn here. This is where the five-part prompt and the profile meet on one object.

Workbook · page 8

Every change you made, and why.
How many rows went in, and how many came out.
One decision you are not certain about, and who you would ask.

12:00

Timing flexes to the room. The clock is a guide, not a gate.

Demo · walk your own log, including the part you could not settle

2,419 rows in. 2,399 rows out.

Name what it is for

The thing that survives this room.

In six weeks nobody remembers why 14 rows are missing. The log remembers.

Close

Three things exist now that did not exist at nine o’clock.

1

Two dashboards from one file, where the only difference was the prompt.

2

That board, where fifteen prompts produced four different answers to one question.

3

Your cleaned Work Order Set, and the change log beside it.

One line held this day together. The prompt is the product.

The hook into Day 2

You finished holding something you did not have this morning.

A file you can defend, and a record of why. That has a name. It is called Data, and it is the bottom rung of a ladder with four rungs on it.

1 Data

2

3

4

Tomorrow we climb one. Because here is the thing nobody told you about your clean file:
a number on its own still means nothing.

Working With AIon Real Data

I have 2,400 work orders. I want a dashboard.

By the end of today, here’s what changes.

Two halves, two different questions.

Nothing here is a mistake.

Two dashboards. One difference.

Same file. Same tool. Same five minutes.What was different?

Which site had the worst unplanned downtime last quarter?

Fifteen prompts. One file.

Nobody typed a bad prompt.

Three questions before we go on.

How can both be true?

It predicts the next piece of text.

What follows from that one line.

Twenty rows of numbers. Ask for the total.

What this changes.

This does not start a new idea.

This is the five-part data prompt.

Seven lines. Under two minutes.

Three things worth knowing.

Same file. Same question. This time all five parts.

Rebuild it with all five parts.

Which of these did your prompt catch?

I did not tell you the whole reason.

Every one of these is a defensible reading.

None of them are wrong. Only one is what you meant.

So before you analyze anything, you look.

Six questions. About three minutes.

Where it gets interesting.

Three questions before the long one.

Profile it. Then clean it.

Profile the full file. Then clean it.

Four findings. The last one changes the day.

Do you delete that row?

The five-part prompt and the profile meet on one object.

Finish your change log.

2,419 rows in. 2,399 rows out.

The thing that survives this room.

Three things exist now that did not exist at nine o’clock.

You finished holding something you did not have this morning.

Presenter & pointer controls

Working With AI
on Real Data

Same file. Same tool. Same five minutes.
What was different?