Plain first, then voice: what we changed after a padded draft came back unchanged

A heavily padded draft is the hardest thing we're asked to rewrite. We found out by pasting one in and reading what came back. It looked like a rewrite, but it still said "Dear Valued Client", "delighted to confirm", "committed to delivering an exceptional experience" and "bringing your vision to life". This note covers what we changed, what it cost, and what it did not fix.

01What went wrong

Sealis edited the draft in place. When the draft is mostly stock phrasing, the model keeps the phrasing, because it is in the text it was handed. The meaning check made this worse: it flags any clause the rewrite drops, so removing a filler sentence read as a content error and sent the rewrite back toward the original wording.

There was a second gap. Sealis names the stock phrases it knows to the model, and checks afterwards which survived. Its list caught 6 phrases in that draft and missed 7, so those 7 were never named.

02What we changed

• More phrases. We added the service-email boilerplate that was missing: "Dear Valued Client", "delighted to confirm", "committed to delivering", "exceptional experience", "complete satisfaction", "bringing your vision to life" and "thank you for choosing us for your…". We left out formal wording that is plain rather than stock, such as "please be advised".
• Plain first, then voice. If a draft has three or more flagged phrases, a first call reduces it to what it says. Every fact, number, date, name, request and condition stays. Pleasantries, flattery and promises with no content go. The voice is then written from that plain version. If the plain version loses a hard fact, or most of the text, the original draft is used instead. The facts are checked in code.

The same draft and the same samples, before and after:

Before: Dear Valued Client, Quick update on your kitchen renovation. The total cost is $12,400… Our team is committed to giving you an exceptional experience… We're looking forward to bringing your vision to life. After: Hello, Right, so the kitchen. It's $12,400 all in, materials and labour included. That's the total. Work starts Monday 3 March and should be done by Friday 21 March. If you've got any questions or concerns, get in touch.

03What it cost

On the 24-rewrite development set (the same AI-written drafts as the voice transfer report), against the previous version of the pipeline:

• Facts kept: 100% before, 100% after.
• Machine phrasing cleared: 97% before, 94% after (inside the noise).
• Points kept: 99% before, 90% after. This drop is the point of the change, because the dropped points are mostly filler, but it is also a real loss when the filler mattered to someone.
• Length: 0.94 of the draft before, 0.74 after. Only half of the rewrites now stay inside our old 85 to 120% band, so by our own pass bar fewer rewrites "pass" (33% before, 29% after).
• Voice match, by the topic-independent measure: 46% before, 50% after the plain-first step, and 58% once the voice pass was also told it may lead with the main point and reorder. Both are inside the noise at this sample size, so we do not claim a gain. Facts kept stayed at 100% in all three runs.

04What it did not fix

Voice matching is no better. What improved is that the result no longer reads like the draft it started from, and a short rewrite in the right voice is easier to judge than a long one in none.

We ran the development drafts only, on one model, and judged the readability of the examples ourselves. Nobody blind has rated these. The change is on by default, and can be switched off with one setting.