ScaffoldDocsSkill
scaffold-docs-skill is a Claude Code skill that generates and maintains technical documentation for code libraries through an iterative, author-reviewed process. Rather than producing documentation in a single unreviewed pass, it structures the work into phases and gates each phase behind explicit author approval before advancing.
AI-generated documentation tends to be either too shallow (summaries of what the code does) or misaligned with the intended audience. It also drifts: once generated, it falls out of sync as the codebase evolves. scaffold-docs-skill addresses both problems -- the iterative review loop enforces quality, and the .scaffold-docs.yml tracking file records the last sync commit and audience definition so updates can be diff-scoped to only what changed.
Documentation is built in three tiers:
- Getting Started: A narrative tutorial focused on a single use case, written for a reader encountering the library for the first time.
- Diving Deeper: Topic-organized files explaining design decisions and intent behind the API surface.
- Reference: Per-module API specifications.
Each tier is produced in passes: structure first, then headers, then full prose. The author reviews and approves each pass before the skill advances. Nothing is finalized without a human sign-off.
The skill has 78 stars and 4 forks on GitHub as of June 2026, with 15 commits and no formal releases. It is installed by symlinking the folder to ~/.claude/skills/scaffold-docs/ and invoked via /scaffold-docs.
Placed at Assess. The iterative human-in-the-loop documentation pattern is sound and addresses a genuine problem with AI-generated docs. The skill is early-stage (no releases, small community) and has not been used in this pipeline. Trial requires running it against a real codebase and evaluating whether the output quality justifies the review overhead.
