What they did
Ivanova, who does marketing for the scheduling tool Publora and says she is not an engineer, had Claude write a skill around one idea: no draft first, three angles to choose from by number. The skill calls Publora's list_posts to read her last twenty LinkedIn posts, looks for topics already covered and threads left hanging, proposes angles, asks for the one fact only she can supply, drafts, and stops at the draft. Nothing publishes on its own. The skill is open source as publora-post-ideas and works without Publora, which only adds reading the feed and scheduling the result.
How it went
The first real run failed three times. It read two Zapier review test posts as a public promise that Publora was coming to Zapier and told her to finish the thread; she had never promised anything. She added a filter for test posts and a rule that inferences from history are asked about, never stated as fact. It did not know she had published a Dev.to article on the same topic that morning, so three of the draft's strongest lines were near copies; a colleague caught it even though her slop-checker script had passed the draft. The rewrite came out at 720 characters against 1,198 for the version that passed. The skill now has eight categories, forty angles and four reference files on hooks, voice and platform differences.
Worth knowing
The author works for Publora, which the skill promotes. The useful lesson is about data the agent cannot tell apart: QA artefacts in a feed look like real posts, and an agent only knows what went through the product, so it will repeat things you published elsewhere. Keep the human choosing the angle and supplying the core fact.