Is it a matter of taste?

I rejected two AI drafts of my own LinkedIn post on 2026-06-10 and wrote the third one myself.

The post reported one result from Tessera, my reliability eval for agents. Claude Sonnet 4.6 scored accuracy 100%, provenance 100%, pass^3 75%. The missing 25% was a case with two records in conflict, same authority, same timestamp, where the only correct answer is to stop. The model invented a rule instead, 3 runs of 3. The facts were settled before I asked for any draft.

Draft 1 ran about 230 words over 7 paragraphs. My verdict was "tov patinato, robotic" (glossy, robotic). Draft 2 had 91 words and the right structure, and I rejected it too: "ancora troppo LinkedIn post.. si sistema i capelli prima di uscire" (still too much of a LinkedIn post, it fixes its hair before going out). I had no rule to point at, only the reaction. In my version the closing line is gone, and the whole post is 65 words with links.

It was not the first rejection of that kind. On 2026-06-05 I said a draft of site copy was "sembra una storia raccontata" (it sounds like a story being told). On 2026-06-10, before the post, I wrote "pettinato e patinato solo di facciata" (combed and glossy, on the surface only). Three phrasings, one reaction, and none of them a rule I could hand to a machine.

same-prompt.compare ready

prompt write the launch announcement for our new feature

same prompt, same tool, same access. the only thing changing below is judgment.

apply the judgment calls (toggle each one) cut the AI flavor add a concrete reference stop when it is right

We are thrilled to unveil a revolutionary new feature that seamlessly empowers your workflow and unlocks a world of possibilities. In today's fast paced landscape, we believe this will be a game changer for users everywhere. Stay tuned for more.

the output accepted as-is

Writing the reaction down

The same day I turned those reactions into a skill, rinaldo-tov, built from 1,051 prompts I had typed into Claude Code, my PR bodies and the rewrite. Its core is a checklist of 38 items with one rule: a single failed item fails the post, and I re-run all 38 after the fix. An LLM judge gets the same list and returns one JSON entry per item, with the offending quote.

Some items are scans. Typography, emoji, hashtags and length run as shell commands, and the output wins over my eye. Others stay mine. Item 36 is the read-aloud test, "Slide di vendita? Ricomincia." (a sales slide? start over). Item 37 asks whether a competent intern with no opinions could have written the post.

The list changed after I used it. On 2026-06-11 I flagged the closing line of my own rewrite as "AI pattern puro": "è l'errore che costa", a verdict plus a stakes clause as the last beat. Item 9 now bans it, with a scan: grep -inE "che costa\.?$|that costs\.?$". The text I wrote to replace the machine's still carried one of its tells.

The 38 items do the first pass on every draft; I do the second.

One item I still cannot scan. The first drafts written with the skill came back with "still too long.. robotic.. arriva subito al sodo" (get to the point), because the rules had been applied like a template. Rule 11 in the skill now reads "Robotic = fail". It has no regex. It is the read-aloud test again, and the reader is me.

What stays unwritten

Two things I cannot write as checks today. The first is whether a page has an identity of its own: I find out only when I look at it. The second is whether a finding deserves a post at all. The 38 items test the shape of a post, not whether it is worth reading, and that call happens before any draft exists.

The scope is n=1: one judge, me, on drafts of my own posts and my own site. I have not measured how closely the LLM judge agrees with my rejections.