Lessons · Lesson 5 of 11
- 01 · The unit of account is a task, not a tool
- 02 · What to give it: the facts, the format, the forbidden
- 03 · Checkability, and the four classes of output
- 04 · Reading a long document: ask for pointers, never a summary
- 05 · Drafting a message: the saving is on the blank page
- 06 · Reconciling a list: change the output's shape, change the review
- 07 · Setting up a calculation: the task where the arithmetic says no
- 08 · Translation, checked the way it actually fails
- 09 · The review budget, and the rate at which checking stops working
- 10 · The log, and why the task you use most pays least
- 11 · Six weeks, honestly totalled
Drafting a message: the saving is on the blank page
Measure where the time actually goes in writing to a buyer, and price the two drafts that quietly conceded something nobody had agreed.
Lesson 5 of 11 · 12 min
Thirty-one drafts, and a very small number
Writing to a buyer is the task every desk reaches for first, so Marasem logged it carefully. 31 drafts over six weeks: chasing an approval, explaining a delay, asking for a size set, answering a claim.
| Term | Minutes |
|---|---|
| Manual time | 8.5 |
| Input time | 2.0 |
| Waiting time | 0.5 |
| Checking time | 2.7 |
| Repair time | 1.0 |
| Net saving a draft | 2.3 |
2.3 minutes a draft. Over six weeks that is 71.3 minutes and USD 11.41.
It is the smallest positive number on the log. It is also the task the desk enjoyed most. Both of those facts matter, and lesson 10 is about the gap between them.
Where the time actually goes
Ghaly timed her own two halves for a fortnight. A hand-written message took 8.5 minutes. Of those, 3.4 were spent before a word was typed: deciding the shape, choosing the opening, deciding how much of the problem to admit. The other 5.1 were spent writing and rereading.
With a draft in front of her, the first number fell to 0.4. The second became reading and editing.
The saving is on starting, not on writing. That is a real thing to save. The blank page is where a merchandiser's afternoon disappears, and having something to react to is worth two minutes and a better mood. But it also caps the saving, permanently. A message to a buyer is a class 3 output. Nothing in the room settles whether it is right except your own judgement, so the check is you reading it as the person who would have written it. Marasem's check ratio here is 43.5%, and no move gets it much lower.
The two drafts that would have cost real money
Of the 31, two contained a commitment nobody at Marasem had made.
One offered to hold a ship date the plan could not support. The other accepted a price. Two emails earlier, in a paragraph of its own, Ravenscar had proposed a reduction of USD 0.11 a garment on purchase order RO-9142. The draft was replying to a question about labels, and it closed with a sentence agreeing to that reduction. On 24,000 pieces that is USD 2,640.00. It is 0.51% of the FOB price of USD 21.40, and 231.4 times the entire six-week saving on drafting.
Both were caught. That is not luck, and it is not vigilance. It is a specific forty-second check, described below, and it is the only part of the review that has ever caught anything on this task.
Neither draft was inventing anything dramatic. The buyer's price had been mentioned in what Ghaly pasted. The ship date was in the tech pack. The system did what it does, which is produce the most plausible continuation of a merchandiser's letter. And a merchandiser's letter, at that point, plausibly agrees.
The dangerous drafts were the good ones
Marasem's desk rated every draft on the log: needs work, needs an edit, or sendable as it stands. Six of the 31 were rated sendable.
Both commitments were in those six. The other twenty-five contained none.
That is the finding, and it is uncomfortable, because nobody did anything wrong. A draft you have to rework gets read as a draft. You are inside it, arguing with it. A draft you would send gets read as a letter, which means it gets read for tone and not for content. The better the writing, the shallower the reading.
The check that works, in forty seconds
Do not review the prose. The prose is not the risk, and it is the thing these systems are best at.
- List every number, date, quantity, price and promise in the draft.
- List the same from what you gave it.
- Anything in the first list that is not in the second is invented until you prove otherwise.
- Read only the sentences those items sit in.
Marasem logs this at 0.7 minutes of the 2.7-minute check. It caught both commitments. Over the six weeks it also caught four smaller items: a quantity restated in the wrong unit, two dates shifted by a weekend, and a style code from a different order.
The remaining 2.0 minutes are the judgement half. Is this the right thing to say to this buyer today? That half cannot be reduced, because it is the reason you are the one sending it.
Check yourselfA draft to a buyer reads perfectly and you have nothing to change. What do you do before sending it?Show the answer
The forty-second check, and precisely because you have nothing to change. A perfect draft has skipped the reading that a rough one forces on you, so the mechanical pass is the only reading it will get. At Marasem, two of the six drafts nobody wanted to edit carried a commitment worth USD 2,640.00 between them.
Prompt · Diff the figures and the terms, and nothing else
Before a drafted message leaves your outbox, and before a translated work instruction reaches the floor. It is the ninety seconds of the review that catches almost everything a review ever catches.
Act as a proofreader with one narrow job. You are given a SOURCE and an OUTPUT. Do not comment on style, tone, grammar, fluency or completeness. Do the following and nothing else. First, extract from the SOURCE every number, quantity, price, percentage, measurement, date, style code, purchase order number, size code and named commitment about who will do what by when. List them. Second, extract the same from the OUTPUT and list them. Third, produce three lists: items present in both and identical; items present in both but changed, showing each value; and items present in the OUTPUT that are in no part of the SOURCE. Fourth, treat everything in that third list as an invention until proved otherwise, and quote the sentence each one sits in. Fifth, if I have given you a glossary, check each of its terms: where the source term appears, the output must carry its listed partner, and any substitution must be reported even where the substitute is correct in general use. My glossary: [PASTE IT, OR WRITE NONE]. Sixth, report any sentence in the OUTPUT that states or accepts an obligation: a date, a price, a quantity, an agreement. Quote it, whether or not it appears in the source. Do not rewrite anything. SOURCE: [PASTE]. OUTPUT: [PASTE].
AI can make mistakes — check anything you act on.