Week of 19 October 2026: writing that reads like a person, and the first fourteen replies

91 invitations, 60 accepted, 14 replies, across three accounts whose owners had agreed to publish their weeks. The first real week. The accept rate was good and the reply rate was not, and this is the week I spent reading the messages that got no answer and rewriting how they are written.

Tuesday: the replies start teaching the drafter

Three commits on Tuesday changed what the model is told. The replies it had got, by source, angle and title, were counted and fed back: the titles that reply nudge the queue, the angles that reply are preferred. An acceptance was re-dated by the ledger event rather than the lead's stage change, because the stage moves again when the first message goes out and last week's acceptances had reappeared in this week's column on the numbers page. The daily run's reads and writes moved to the model's batch endpoint at half the price, which is the start of the fifty-cents-a-day story. And the landing page lost its em dashes and a third of its words.

Wednesday to Friday: four parts, in the sender's voice

Wednesday's two commits are the ones I would show somebody who asked what the product writes. The first message now has four beats and the drafter is told them in order: an anchor, which is a thing the person did or wrote, said plainly and dated; one line on who the sender is, and nothing else about them; the value, which is what the problem costs them now and later, with one true result to cite if there is one; and an ask small enough to answer in a sentence. A backtest sheet in the repository runs the drafter against a set of real profiles and scores the output against the shape everybody deletes, so a prompt change can be measured before it ships.

The second commit is about voice. The message is written as the account owner, whoever they are, in their own language and tone from the targeting form, and never as a generic founder. Thursday fixed the hot-lead ask, which the model had been turning into an invitation to a meal it had chosen; the ask is a call on the person's own calendar. Friday made the first message arrive as short paragraphs rather than a wall, because the reply rate on a wall is the reply rate on a wall.

• A rewrite requested from the queue now gets the same profile the daily run gets, rather than a thinner one; the two had drifted.
• The brand got a logo kit and a banner built from one SVG source, black and white, no gradients.
• A billing refusal from the model provider became a top-severity incident, because the day it happened the product stopped writing and nothing said so for hours.
• A refused profile read is retried with the person's other identifier before being given up on; a tenth of reads had been failing on one of the two.

Saturday and Sunday: the inbox is a list of people and one thread

The replies page was a feed of messages. By Saturday it was a list of the people who answered on the left, newest first, and the thread of whoever is selected on the right: what we sent, what they said, what you said back, and the box to answer in, with the header and the reply box pinned so the thread scrolls between them. On Sunday Today got the week as a coloured funnel against last week, the status strip became three tiles, a tab older than the deploy was taught to reload itself instead of failing every click, and a message with half an emoji in it stopped crashing the daily run. That last one was found by a real person's headline.

The week's number, 14 replies to 91 invitations, is on the numbers page under each account's first name. It was the first week with enough acceptances to say anything about the writing, and what it said was that the writing had to change, which it did, which the next post measures.