skill-release

# Skill Release — converge the publish surfaces A release is **convergence**, not a pipeline: every publish surface derives from the skill's own files, so running this at any time is safe and idempotent. Surfaces: the collection index (repo `README.md`), per-skill interface metadata (`agents/openai.yaml`), the [cankun.me/skills](https://cankun.me/skills) page, and — rarely — a flagship repo. **Write boundary**: this skill edits metadata surfaces only — frontmatter keys (`metadata.summary`, `compatibility:`), `agents/*.yaml`, the index README, and `cankun-blog/app/data.ts`. The SKILL.md body belongs to the author: a content finding (weak description, missing triggers) is a flag routed back, never an edit. The route starts at a committed skill — moving a directory into `skills/` is the author's (or skl's) move, not this skill's. The human-facing summary of a skill resolves through one chain, highest first: 1. frontmatter `metadata.summary` in `SKILL.md` 2. `agents/openai.yaml` → `interface.short_description` 3. first sentence of the frontmatter `description` Index prose is always derived through this chain, and convergence flows toward the **better text in either direction**: when a row reads badly, fix the skill's `metadata.summary` and reconverge; when existing row prose beats the derived value, promote that prose into `metadata.summary` rather than degrading the row. ## 1. Inventory Enumerate `skills/*/SKILL.md` frontmatter (committed skills only — an untracked skill ...