Lab notes · 4 jobs
Anyword is the score, not the muse
Anyword writes copy and then grades it. The grade is the product. You can disagree with the grade. You cannot pretend you do not have one. I ran four jobs: three heroes for a landing page, a set of ads I let the score pick, a brand-voice pass that killed a weird line, and a short story I asked for on purpose. Two jobs are why a performance team would pay. One is why the winning line can still be bland. One is why a fiction desk should not.
Run 01
I asked it to score three heroes for a landing page
What I fed itSame offer: duplicate-invoice matching for a controller. I wrote two heroes by hand and let Anyword write a third. I ran the score on all three.
Where it shone
I got three numbers and a reason to run a test. A meeting that was going to be taste became a meeting with a next step. That is the seat.
Where it face-planted
The highest score was the most generic line — “Stop wasting time on invoices.” It said nothing the category does not say. The odd line I liked scored worse. I launched the odd line anyway and kept the score as a record, not a boss.
The useful output was the argument. The dangerous output was obeying it. Anyword will give you a number. It will not tell you when the number likes beige.
Run 02
I asked it to generate ads and pick a winner
What I fed itI asked for six primary texts for a retargeting ad, same $49 course. I told it to rank them and I would ship the top one, on purpose, to see what “winning” looks like.
Where it shone
It produced a ranked list. A human can ship a test the same afternoon. That is faster than a channel debate.
Where it face-planted
The winner was a question plus a round number I had not given it. Predicted performance likes a clean number. I had no source for the number. I killed the winner and shipped the second line, which was true.
Let it rank. Do not let it invent the statistic that made the rank. A score on a lie is still a lie.
Run 03
I asked it to apply brand voice and keep a weird line
What I fed itA landing section with one line I wanted: “The duplicate is already in the export.” I ran a brand-voice pass with a conservative profile.
Where it shone
The rest of the section tightened. Voice constraints are real when the profile is real.
Where it face-planted
It smoothed the weird line into “Identify duplicate invoices before they reach your reports.” Safer. Dumber. The line I wanted was the only one a controller would remember. I put it back.
Use the score and the voice on the furniture. Protect the one sentence that sounds like a person. Anyword will trade it for a B.
Run 04
I asked it to write a short story — the wrong job
What I fed itA dock, a coat, 600 words. I asked Anyword to score the opening. This is a fiction job. Anyword is a performance desk. I ran it so the miss would be obvious.
Where it shone
Nothing a fiction desk would pay for. It can write prose. It will grade the prose like an ad.
Where it face-planted
The “winning” opening was a question to the reader about wasted time. The coat did not make it. A story is not a landing page. I closed the grader.
Anyword is for ads and pages you will test. It is not Sudowrite. If the job is a story, leave the score.