Tested the publish-content skill on this project and immediately realized: the agent writes like an agent, not like a person. The skill could read transcripts, identify inflection points, and produce structurally correct MDX — but the voice was wrong. "In this session, we implemented the core pipeline" is how a bot talks. "55 AI Studio apps and no idea what to do with them" is how I talk. The fix wasn't a prompt tweak — it was a rule: read 3-5 existing log entries in content/log/*.mdx before drafting new ones. Match the voice. Bold declarative lead sentences. First person. Specific data. No "phase," no "implementation," no "session," no "execute."
The /learn session after this conversation codified it: the batch-project-pipeline skill now requires voice calibration from existing entries before any drafting begins. The skill also got the cascading provenance format — conversations feed log entries, log entries build threads, threads compose projects. Every claim traced to a specific conversation ID and step range. The meta-lesson: automating content production is easy; automating content quality requires teaching the agent what good looks like by example, not by instruction.