Where the two hours go
Ask how long a newsletter takes and most people think of the drafting. Yet the measurements say drafting is precisely the part a machine speeds up most: on office writing tasks the time spent fell by roughly 40 percent, and on email by 31. So the two hours are not going into the sentences.
They go into three other things. The first is gathering: reviewing what happened this week and deciding what is interesting. The second is deciding: what the lead item should be and what stays out. The third is checking, if it happens at all. Of those, the machine helps not at all with the third, barely with the second, and a great deal with the first — provided you tell it in advance what to look for. The secret of the forty minutes is that the decisions get made at the start of the week, not while writing the newsletter.
What gets built in advance, once
Four durable pieces, and with them most of the weekly work stops being work. Roughly two hours to build, then good for months.
- The style sheet. Rhythm, register, banned list, structure, stance. This goes into every request; the details are in fixing your own voice, which covers all five elements.
- The section template. How many sections, how many words each, what is mandatory in them. For example: lead item 200 words with one claim and one number; three short items at 40 words each; one tool tip at 60. The template is what makes a newsletter recognisable, and what narrows the request to the machine.
- The source list. Five to ten places the week's material comes from, in order. Without it you decide where to look afresh every week — and that step takes the most time.
- The subject-line criteria. Not templates but rules: how many characters, whether a number goes in, what we never write. This is needed because the subject line is where the model's default is flattest.
The forty minutes, stage by stage
Five stages, each with its own budget. If a stage overruns, it means the previous stage was not finished properly.
- Gathering — 10 minutes. Down the source list, and from every hit one sentence plus one link. Do not ask the machine for a summary but for filtering against your own criteria: "which of these affects [your audience], and why". The output is a raw list, not prose. If a longer piece does have to be summarised, separate rules apply: the sectioned method from reliable summaries works here too.
- Outline and selection — 5 minutes. This step is yours. Pick the lead item and write in one sentence what you are claiming about it. From the rest, mark the three that stay. These five minutes cannot be replaced by a machine, and they decide the quality of the whole issue.
- Writing — 10 minutes. A separate request per section, with the section template and the style sheet. Do not ask for the whole thing at once: the sections then become variations on the same text, and the length instruction is not respected either. The same logic sits behind the outline-first workflow for longer pieces.
- Subject line — 5 minutes. A separate task with a separate request, based on the finished text. There is a section on this below.
- Checking — 10 minutes. Every number, name, date and link. Click every link — not because the machine invents them but because old links die. The method is set out in checking before you publish, step by step.
The subject line, which is its own job
The subject line needs a separate request because the model lacks exactly the information that decides it: what your readers have seen over the past few weeks. So its default is the genre average — the subject line that would work on any newsletter, and therefore stands out on none.
Four rules help, and all four are about narrowing the model's room. Give the character count, because without it it will write more than fits on a phone. Require the lead item's claim, not its topic — "the cost of hybrid work moves somewhere else" works, "on the costs of hybrid work" does not. Ban the question form and the exclamation mark if that is not your voice. And ask for ten, not three: the subject line is the format where the first three suggestions are almost always the dullest, and things start getting interesting around the seventh.
The prompt set
Two requests cover stages three and four. Fill in the templates once; after that only the material changes each week.
=== SECTION REQUEST (stage 3, one per section) ===
Write one section for my weekly newsletter.
Section: [lead item / short item / tool tip]
Length: exactly [200] words, plus or minus 20
Must contain: [one claim + one number + one link]
Forbidden: an opening sentence, a summarising close, exclamation marks
The material: [raw sentences and links from stage 1]
The claim this section stands for: [one sentence from stage 2]
Style: [your style sheet]
Anything the material does not tell you, do not invent: write [to check].
=== SUBJECT LINE REQUEST (stage 4) ===
Write TEN subject lines for the newsletter below.
At most [45] characters. Every one must contain the lead item's CLAIM.
No questions, no exclamation marks, no emoji, and never the words:
[crucial, must-have].
Make three factual, three with a concrete number, four from an unusual angle.
The newsletter text: [the finished text]
What never to hand over
Three decisions where the machine's suggestion is reliably worse than yours — not because the model is weak, but because it does not have the information.
- What the lead item should be. Deciding this requires knowing what you wrote last month, which pieces got replies, and what actually matters to your readers right now. The model knows none of the three.
- What stays out. Leaving something out is a more valuable editorial decision than putting it in, and the model's default is always to put it in. Give it ten items and it will process ten.
- What you promise in the subject line. The distance between subject line and content is the one thing readers learn about you within a few weeks. You own that, not the machine.
When it is not worth it
If the newsletter goes out less often than weekly, the preparation does not pay back — the four durable pieces still have to be built, you just use them less. For a monthly, the style sheet and the section template are worth building; the source list and the prompt set are not necessarily.
And there is one case where the machine actively hurts: when the sole value of your newsletter is the personal viewpoint. There a machine-written section is not merely neutral but a step backwards, because it smooths out exactly the variation the reader subscribed for. In that case the machine's job is gathering and checking — stages one and five — and the middle three stay yours. That is still fifty or sixty minutes instead of two hours.
What to take away
The forty minutes do not come from faster drafting but from building four things once in advance: a style sheet, a section template, a source list and subject-line criteria. After that the weekly run is five stages, each with its own request: gather, decide, write section by section, subject line, check. Stage two — what leads and what stays out — is yours, and if you hand that over, the remaining thirty minutes are worth nothing. And for the other formats you make from the newsletter, do not work from the newsletter but from the source: that is what the repurposing workflow is about.