Claude Has the Patterns. I Have the Ear.

Blog post #59

Yesterday was my first day working for myself. I marked it with a LinkedIn post, which sounds trivial until you consider that I almost never write on LinkedIn and have no idea what I sound like there. So I did what I now do with most things I am unsure about: I opened a session with Claude and we worked it out together.

What came out of that was more interesting than the post itself.

What shipped

  • A five line LinkedIn post. A good morning coffee selfie, a callback to the pandemic when I sent one to my team every morning, and the news that I am on a twelve month leave to build an AI company.
  • Replies to every single comment. There were thirty two by the evening.
  • Updates to three of my writing skills, based on what the session exposed about them.

What’s working

The division of labour turned out to be sharper than I expected.

Claude knows the patterns. It knows that opening a question with a frame (“du som startat Wrycom”, roughly “you who started Wrycom”) creates distance, and it knows why a rule of three reads as machine written. It has read more text than I ever will.

I know how things sound in my mouth. That turns out to be the half that cannot be delegated, because Swedish is not the language Claude thinks in first.

Three examples from the session:

Claude suggested “kul med marginal” to mean something like “fun by a margin”. I could not explain what was wrong with it, only that no Swede says that. It is English idiom wearing Swedish words.

For a former colleague who had just been made redundant, and who I had gone for a beer with after her last day, Claude wrote “kul att vi tog en öl”. Kul means fun. The beer was good, but the reason for it was that she lost her job, and the word sits crooked on top of that. We changed it to fint, which carries warmth without turning it into a party.

And when a comment complimented me, Claude’s reply talked the compliment down. That is the Swedish reflex, and it feels humble from the inside. From the outside it reads as doubt. Take the compliment.

I caught five of these before Claude pointed at them. That is not a complaint about the tool. It is the whole point: I could hear that something was off every time, and I could explain why almost none of the time. Claude could explain every time, once I pointed.

What’s unclear or broken

My humanizer skill had a rule telling me to strip hyphens from compound words, so “data-driven” would become “data driven”. That is wrong in English and often ungrammatical in Swedish, where the hyphen is mandatory. Following that advice makes text worse, not more human. The actual tell was never the hyphen. It was the density of corporate compounds.

The same skill banned emojis outright. In a LinkedIn comment an emoji is ordinary human punctuation, and removing it makes the reply read cold. The rule was written for headings in articles and then applied to everything.

Neither skill knew any Swedish. Both were built on an English guide, so they had nothing to say about the fingerprints that actually show up in Swedish AI text: em dashes where Swedish uses a shorter spaced dash, translated English idiom, the same three connectors opening every paragraph, and the LinkedIn dialect of “så otroligt stolt” and “vilken resa”.

And all of it assumed the text was an article. That is why none of it ever triggered on a four word comment.

Decisions made

  • Keep the post short and name nothing. No company name, no product. A teaser that holds something back gives me a second post to write later.
  • Do not update the job title the same day. LinkedIn generates its own “new position” card that competes with the post.
  • Answer every comment, and vary the answers. Everyone reads the thread top to bottom, so nine identical thank yous look worse than nine short different ones.
  • Move the real conversations offline. I dislike turning private talk into public performance. Suggesting a beer works better than a DM, and it is worth more to the other person.

Tooling & process

I rewrote all three skills at the end of the session, while the examples were still fresh. humanizer got a length calibration at the top, five Swedish specific patterns, and a fix for the hyphen and emoji rules. no-ai-slop got the Swedish list and stopped opening with an em dash of its own. social-writer grew two modes it was missing entirely: writing an original post when there is no source text, and answering comments.

That last one is the part I keep thinking about. The skill existed to turn a finished blog post into social copy. But what I actually needed was help writing four words to an old friend without sounding like a stranger. The tool was built for the wrong shape of problem, and I only found out by hitting it.

The post got 1,585 impressions and fifty two reactions, which is far more than I expected from a network I have barely spoken to in years. But the line people replied to was “Lite nervös. Mest taggad.” Slightly nervous. Mostly excited.

That was the one line Claude told me not to touch.

— Stefan