When a draft fails a quality check and gets rewritten, the best version is now kept rather than whichever came last. A rewrite can no longer leave you with something worse than the draft it started from.
What changed
Every draft is scored before it can send. A draft that falls short gets rewritten with the specific problem pointed out, and the rewrite is scored again.
The rewrite was always taken as the answer, because it was newer. That assumption is wrong often enough to matter. Fixing one problem can introduce another: a draft rewritten to remove an unsupported claim can come back shorter, flatter, and worse overall than the version that had the claim.
Now every version is scored and the best one wins. If the rewrite scores higher, it is used. If the original scored higher, the original is kept and the rewrite is discarded.
A second change went with it. Long drafts with long rewrite instructions could exceed what the writer can hold at once, and the instructions were the part that got cut, so the rewrite arrived without knowing what it was supposed to fix. Instructions and draft are now fitted together within the available space, so the rewrite always knows its brief.
Why it matters
A quality system that can make things worse is worse than no quality system, because it adds cost and variance without adding reliability. You cannot trust the output and you cannot predict it.
The failure was also specific in a way that hurt. Rewrites triggered by the strictest checks, meaning the drafts under the most scrutiny, were the most likely to come back flattened. So the emails receiving the most attention were sometimes the weakest ones sent.
The prompt problem compounded it. A rewrite without its instructions is not a rewrite, it is a fresh attempt at the same task, and it fails the same check for the same reason. Those drafts consumed two full attempts and then gave up.
How to use it
Nothing to configure. Keep-best applies to every draft that goes through repair.
Drafts that would previously have been abandoned after two failed rewrites now go out more often, because the attempts are spent usefully.
The quality record on a draft shows what was checked and what the final version scored, so a draft that scraped through is identifiable.
What the checks look for
Claims that research did not support. The most important one. An email cannot assert something about a prospect that was never established.
Length and structure. Emails that run long get less response, and a first email that reads like a page of prose usually gets none.
Tone. Openings that overclaim, closings that push too hard, and anything that reads as automated.
Personalisation that is not personalisation. A first line that merely restates the company name is not personal, and it is worse than a straightforwardly general opening because it advertises the attempt.
Two attempts, then it stops
Repair gets two attempts. After that, if nothing has reached the standard, the draft is marked failed rather than sent.
Two is a considered number. A problem that survives two targeted rewrites is usually not a writing problem, it is thin research or a brief that does not give the writer enough to work with, and a third attempt will produce a third version of the same shortfall.
Failed drafts show the reason, and the reason usually points at the input rather than the output. A run of failures on one campaign is a signal about the campaign brief.
What the budget problem looked like
The writer can only consider so much text at once. A long draft plus a long set of rewrite instructions could exceed that, and what got dropped was the end of the input, which was where the instructions sat.
So the rewrite received the draft and no brief. It rewrote the email in general terms, produced something different but not better, failed the same check, and consumed an attempt.
Fitting the two together within the available space means the instructions always survive. Where something has to give, it is now the less important part of the input rather than whichever part happened to be last.
What changed in the numbers
Terminal failures, meaning drafts that exhausted both attempts and never sent, dropped substantially after this went out, and the drop held.
The share of drafts passing on their first attempt did not change, which is what you would expect: this fixed the repair path, not the writing.
Those two together are the shape of a good fix. The thing that was broken improved and nothing else moved.
Reading a failed draft
A draft that failed both attempts shows what it failed on, and the reason is usually about input rather than output.
Repeated claim failures mean the research was thin, so the writer had nothing verifiable to build on and kept reaching for something it could not support. The fix is upstream, in research or in the lead record, not in the writing.
Repeated tone or length failures more often point at the campaign brief. A brief that asks for a great deal in one email produces long drafts, and long drafts fail length checks no matter how many times they are rewritten.
What did not change
The standard did not move. Keep-best selects among versions that were all scored the same way, and a draft that no version gets over the line still fails.
That is worth stating, because "keep the best one" could be read as lowering a bar. It is not. It removes a way of losing a good version, and nothing weaker gets through as a result.
Availability
Live now on all plans. Applied automatically to every draft that goes through quality repair.