What to give it: the facts, the format, the forbidden

The experiment that settled an argument at the desk

Almost everything written about working with these systems is written about the instruction. Say who it should be. Say what format you want. Be specific. All of that is true. Marasem's desk spent two weeks acting on it, and the log says it was aiming at the wrong half.

Ghaly ran the same task three times in one week, on three different orders of the same shape. The task was the 22.0-minute job from lesson 1: restate a revised size breakdown as a cutting-ready table. She changed only what she handed over.

The middle column is the one worth staring at. Ghaly did what every guide tells you to do. She gave it a role, named the format, told it to be careful, told it not to guess. The instruction went from two lines to fourteen. Checking time fell by 1.0 minute and writing the input cost 2.5 minutes more, so the net saving got worse. She had improved the thing that was already good enough, and paid for it.

The third column changed one thing. She pasted in the buyer's email and the attached grid in full, unedited, and left the original two-line instruction alone. Checking fell from 11.0 minutes to 7.0.

Why the source document is the lever

It is not that the output got better, although it did. It is that the output got cheaper to check.

With the grid pasted into the same window, every figure in the answer has something to be checked against, right there. Checking is a comparison, and a comparison needs two things side by side. Without the source in the window, checking a table of forty-two cells means opening an attachment, finding the row, reading across, and coming back. Forty-two times. Or, more realistically, six times and a shrug for the rest.

That is the general rule, and it is the one this whole course turns on:

The input that reduces your checking time is the input that puts the evidence next to the answer.

An instruction changes what it produces. A source changes what you have to do to believe it. Only one of those is on your side of the equation.

Anything you leave out will be filled in

Marasem logged 18 runs of this task. The first six produced seven figures that appeared in none of the inputs. Five were harmless restatements — a total that had been implied and was now written down. Two were not.

One supplied a size ratio for the bundles. The buyer's grid gave quantities, not a ratio. So the assistant produced a plausible one, correctly formatted, sitting in a column headed the way the factory heads that column. It was wrong for this order, and it looked exactly like every other figure on the page.

The other supplied a cutting allowance of 4%. Nobody had given it one. It needed a number to complete the table, so there was a number. That single habit costs this factory USD 5,544.00 in lesson 7. The mechanism is already fully visible here: a gap in the input becomes a figure in the output, and a figure in the output looks like every other figure in the output.

The repair is one line. Marasem's own measurement of it is why that line is in every prompt on this course:

If a figure is not in what I have given you, write MISSING. Do not supply one, do not estimate one, and do not carry one over from anything else.

After that line was added, the next 12 runs produced one such figure. The instruction does not make the system honest. It makes the gap visible, which is all you need, because a cell reading MISSING is a cell you go and fill.

The three parts of an input, in the order that pays

The facts. Everything you already hold, pasted whole rather than summarised. Summarising your own source before handing it over means doing the lossy half of the job yourself and keeping none of the benefit. If the document is confidential, that is a real constraint and it belongs to course 15.5. The point here is that if you can give it, giving all of it is cheaper than giving some of it.

The format. Say the shape you want back, and say it in the shape your own process uses: the column order of your cutting sheet, your own size names, your units. An output in your format is checked against your other papers by laying them beside each other. An output in a general format has to be translated first, and the translation is where you stop concentrating.

The forbidden. Three lines, and they are the same three lines for almost every task:

1. If a figure is not in what I gave you, write MISSING rather than supplying one.
2. Quote the line you took each figure from.
3. Do not change my units or my codes.

The second is the one desks skip, and the one that pays most. A figure with its source line beside it is checked in about four seconds. The same figure without one is checked by going to find the document, and going to find the document is exactly the step that gets skipped at half past five.

Act as an operations analyst who is sceptical of tools and interested only in measured time. I want one task scored honestly, not recommended. The task: [DESCRIBE IT IN ONE SENTENCE, AS YOUR DESK ACTUALLY DOES IT]. My own measurements: it takes me [MINUTES] by hand; preparing the input takes [MINUTES]; waiting takes [MINUTES]; checking the output to the standard this task needs takes [MINUTES]; repairing what checking finds takes [MINUTES]. My loaded cost is [AMOUNT] an hour. I ran it [NUMBER] times in [WEEKS] weeks. Do the following. First, compute the net saving a use and over the period, in minutes and in money, and say plainly whether the number is positive. Second, compute the check ratio, which is checking plus repair divided by manual time. Then compute the ratio it had to stay under, which is one minus preparing plus waiting divided by manual time. Say whether it clears it, and by how much. Third, classify the output: checkable against a source I hold, checkable only against my own judgement, checkable only by doing the task again, or not checkable in the time available. Justify the class from what would have to happen for me to disbelieve the answer, not from how hard the task is. Fourth, if the class is the third or fourth, say so and recommend retiring the task, and do not soften it. Fifth, if it is the first or second, name the single change to my input that would most reduce the CHECKING time, and estimate by how many minutes. Sixth, list every error this task could produce that my described check would not catch, and for each one name what it would cost if it left the building. Do not estimate any figure I have not given you: write MISSING instead.