Two dashboards from one file, where the only difference was the prompt. Then the room asks one question of forty rows and produces four different answers, and spends the afternoon finding out why.
Day 1 of 3 09:00 – 15:00 Arabella Hotel
The day
The morning answers how do I ask for what I want. You watch one prompt produce a picture and another produce a tool, then feel the gap on your own hands. The afternoon answers why does the answer keep changing, using the evidence the room itself produced at half past nine. You leave with a cleaned Work Order Set and a change log you can defend in six weeks, when nobody remembers why twenty rows are missing.
Three concepts
Cleaning is not a fourth. It is the practice of the third.
01
How a language model actually works
It predicts the next piece of text. It does not look up, and it does not calculate, until you make it write and run code.
02
The five-part data prompt
Role, data description, task, output shape, verification ask. Prompt A had one of these. Prompt B had all five.
03
Profile before you analyze
A short description of what a file contains, written before you analyze it. You cannot write part 2 of a prompt about a file you have not looked at.
Run of show
A guide, not a script, timings flex to the room.
1
Two prompts, one file09:00
Same file, same tool, same five minutes. One prompt makes a picture, the other makes something you can click. The room names the difference before anyone names a framework.
2
The same file, the same question, different answers09:30
Forty rows, one question, twelve minutes, no guidance. Every answer goes up on screen, one at a time round the room. We come back to that board twice more.
Break10:00
4
What an AI is, and how it actually works10:15
One mechanism explains both the impressive dashboard and the unreliable count: it predicts the next piece of text. It does not look up, and it does not calculate.
5
The five-part data prompt10:45
Not a new idea. A shape for the five consequences of the last block. Role, data, task, output, verification. Written live from a blank line in under two minutes.
6
Practice: the Site Downtime question, done properly11:15
Rebuild your morning prompt with all five parts. Then the three pitfalls surface: the duplicate, the negative value, and one asset written three ways.
Lunch11:50
8
Why dirty data breaks an answer12:35
Back to the board. Four defensible readings of the same forty rows, changed one at a time, and the winning site moves each time. None of them are wrong.
9
Profile before you analyze13:00
Six questions you ask a file before you ask it anything else. Three minutes to run, and they settle the argument you cannot win later.
Break13:25
11
Practice: profile and clean the full Work Order Set13:35
Forty-five minutes on 2,419 rows. Mixed date formats, 29 names for 17 assets, duplicates by two different methods, and one cost outlier you should not delete.
12
The change log14:20
The integration beat. The five-part prompt and the profile meet on one object: every change, the rows it touched, why, and the one decision you could not settle.
13
Close and hook14:45
Name what exists now that did not exist at nine o'clock. Then the first rung of a four-rung ladder, and the line that opens Day 2.
×
