This is the big beat of the second example. Instead of describing the writer's voice, we point the agent at nine finished issues and have it extract voice.md, with a quoted line from real work behind every rule. Then we seed slop.md by hand: the phrases and moves that never ship. Both go in a training folder at the repo root. We wire them into the skill: read the training material and the three most recent issues before drafting, and run a specific self-tightening checklist before handing anything over. SKILL.md gets unpacked into steps. Then the same idea, same skill, one thing changed, and we read the two first sentences back to back.
What's covered:
- Building voice.md from real past work instead of self-description
- Seeding slop.md and the maintenance rule that keeps it growing
- Why training lives at the repo root, not inside the skill
- The calibration step: written rules plus the three most recent pieces
- The self-tightening sweep, and why "make it better" is not a checklist
- Unpacking SKILL.md into step files
- Run two: same idea, same skill, and a draft that sounds like the writer
- Where to find the voice-training kit in your course assets