Twelve by Thursday
The Drift 28 brief, expanded: four email subject lines, three landing-page heroes, five social posts. Twelve variants, one launch, Thursday. This is exactly the work the models get handed — in the Reuters Institute's UK survey, one journalist in six generates parts of articles at least monthly (the heavier traffic is transcription and copy-editing, but one in six is a lot of variant batches), and the New York Times' reported newsroom rules explicitly put machine-suggested headline and social-copy options on the allowed side of the line, with a human picking. Machine proposes, you dispose. Fine. The craft question is what "disposing" looks like when there are twelve of them.
Stop pasting variants into a doc
The amateur move is a Google Doc with twelve lines and some highlighting. The problem isn't aesthetics — it's that a line in a doc can't be checked, sorted, killed, or reported on. So the first act of this lesson is a data modeling decision: a variant is a record. An id you can cite in a kill log ("V07 died for a reason, here it is"), a channel that determines which rules apply, and the text itself.
Because the channel is half the spec. A subject line that runs 80 characters isn't off-voice — it's truncated mid-thought in every inbox preview, which is worse. Run the editor: CHANNEL_SPECS carries a character budget per channel — 55 for subjects, 90 for heroes, 140 for the social teaser format — and the check is one comparison. Your real numbers will differ per client and per platform; the point is that they're numbers in a dict, not lore in your head. When the email platform changes its preview length, you change one integer and every future batch inherits it.
What a variant has to survive
By the end of this lesson every variant faces three checks, in order:
- Banned phrases — the sludge list from your voice pack.
- Voice rules — exclamations, sentence ceiling. Measured, not felt.
- Channel budget — the character math above.
Survivors ship to the client. Everything else gets killed with a recorded reason — and the reasons turn out to matter as much as the survivors. That's the kill log, step 6. First, though, you need to see what you're defending against: the specific, predictable way generated copy goes wrong at scale.