The scaffold is in

Docent exists, as of today, as a single commit: the spec, the Claude skill that drives it, the prompt files that shape what it writes, the JSON schemas that validate what it produces, and an Astro site template that renders the result. The whole thing fits in a few dozen files.

What’s in the scaffold

The project has three visible layers:

• The spec — SPEC.md, a working document that describes how Docent is meant to behave and why. It is explicit about the design principles (zero hosted infrastructure, PR-gated output, structured content, same-repo versioning) and about what the project is deliberately not trying to be (not an API doc generator, not an issue tracker replacement).
• The skill — skills/docent/. Five modes (init, update, digest, release, triage), each defined in its own markdown file. System prompts for overview writing, journal writing, issue summarizing, and theme classification. Schemas for everything it writes.
• The site template — skills/docent/templates/site/. An Astro project with seven routes, four distinct visual vibes (editorial, technical, clinical, expressive), and a small set of components.
• The plugin packaging — .claude-plugin/marketplace.json and .claude-plugin/plugin.json. Makes install a two-command job: /plugin marketplace add calumjs/docent and /plugin install docent@docent. No npm, no git submodules, no copy commands.

Scheduling decision

Early drafts of the spec assumed GitHub Actions would run the content updates. That changed during the scaffold: Docent now uses Claude Code Routines for content regeneration and leaves GitHub Actions doing only the GitHub Pages deploy. The practical effect for a maintainer: no ANTHROPIC_API_KEY to configure, no YAML to wire up — just two /schedule commands once, and the site maintains itself on the maintainer’s Claude Code subscription.

A GitHub Actions fallback stays in the spec (§6.7) for users who can’t use Routines, but the primary path is Routines.

Design inheritance

The spec reserves theme.json as a per-repo visual identity the skill chooses once at init. There are four vibes. For this repo, the editorial vibe was the obvious pick — Docent’s name literally means museum guide, and the project is about readable prose. Visual identity is frozen at init; scheduled runs can’t change it.

Coming up

The interesting next question is how the generated content holds up in practice. This post is the first real test — read it, look at the overview, and if the voice feels off, the skill’s prompts need tuning, not the template.