The kills are data too
Most copywriters treat rejected variants as trash — delete, move on. The kills are the most informative thing your batch produces, and they're free. You just have to stop deleting them.
The kill log is one record per dead variant: id, the rule that fired, the specific reason. Three uses, in ascending order of value:
- Client defense. "Why only nine variants?" has an answer with receipts: V02 died for "unleash", V07 was 16 characters over the subject budget. You look like someone running a process, because you are.
- Prompt repair. If "banned" kills dominate the log, your drafting prompt isn't carrying the banned list hard enough — move it higher, make it explicit, add an example. If "channel" kills dominate, the prompt isn't stating the character budget. The kill log tells you which sentence of your prompt to fix. That's a feedback loop most people writing with models never build, which is why their batch quality never improves.
- Pack evolution. A phrase that keeps almost-shipping ("adventure awaits") isn't on the banned list yet but keeps making you wince. Two batches of wincing is evidence. Into the list it goes, versioned.
The batch record is the second artifact, and the quieter one: which model, which pack version, which prompt file, what the counts were. Boring — until someone asks, three weeks out, "the hero copy on the landing page says guaranteed for life — where did that come from?" With batch records you answer in one lookup: batch b3, pack v1, this prompt file, shipped on this date. Without them you're reconstructing your own work from memory, in front of a client, and memory is exactly the instrument that was too tired to catch the problem the first time.
Newsroom AI policies converge on human review before anything publishable ships — the 52-organization comparison found that line almost everywhere. What nearly nobody writes down is the evidence trail that makes review auditable after the fact. Freelancers get this wrong worse than newsrooms: no institution is keeping records for you. Two dicts per batch. That's the whole overhead. Next lesson, when a shipped claim turns out to be wrong, this is the paper trail that turns a crisis into a lookup.