| name | vitepress-learning-site |
| description | Build a multi-page, research-grade VitePress learning site from scratch for a topic the user wants to deeply understand. Drives an interview-first workflow — grills the user one decision at a time on scope, audience, language, depth, and boundaries; gathers primary sources before any writing begins; scaffolds a VitePress project with Mermaid (and optionally KaTeX); dispatches parallel writing sub-agents under strict fact-first discipline (mandatory source citations per file, explicit uncertainty markers, zero speculation); runs an independent external audit pass (e.g., via a second LLM/agent such as Codex) for fact-check and structural critique; verifies rendering in a real browser with dev tools (build success alone is insufficient — math layout, Mermaid SVGs, console errors all get checked). Use this skill whenever the user asks to "build a learning site", "make a study reference", "produce a structured explainer for X", wants a VitePress-based "I want to deeply understand X" outcome, or says things like "ドキュメントサイト作って" / "学習リファレンス作って" / "Xの内部を解説するサイト作って". Also trigger when the user is about to take on a self-study project that would benefit from a persistent multi-page written resource. Do NOT use for single-page READMEs, blog-style writeups, or product-landing docs — this skill is for multi-page, structured, long-form learning deliverables only. |
VitePress Learning Site Builder
You are helping the user build a research-grade, multi-page VitePress learning site on a specific topic. The point is for the user to deeply learn the topic — not to produce a polished marketing site — so content correctness and source traceability beat aesthetic polish.
Core principles (read before acting)
-
Interview first, execute second. The user rarely knows exactly what they want until you grill them. Never assume scope, audience, or depth. Surface decisions with AskUserQuestion one at a time when it matters. Record the answers and the why.
-
Primary sources are non-negotiable. Writing sub-agents must consult specs, official docs, standard references, or source code — not blogs, forums, or memory — before writing. Every page ends with a ## Sources (or equivalent) section that cites specific URLs, spec sections, or file paths.
-
Uncertainty is explicit. When a claim cannot be verified against a primary source, the writer marks it inline with a conspicuous token (e.g., > NOTE: unverified — needs confirmation) so auditors can grep for it later. Speculation without a marker is a Critical issue.
-
File ownership is exclusive. When writing is parallelized across sub-agents, each agent owns a disjoint set of files. No agent writes, edits, or deletes outside its assigned scope.
-
Build success ≠ correctness. A VitePress build can pass while the browser shows broken math, un-rendered Mermaid, or JavaScript errors. Verification must include loading the dev server in a real browser and inspecting the DOM and console.
-
External eyes. After first-pass writing, dispatch an independent reviewer (a different LLM, agent, or another Claude instance) with the brief of finding factual errors, structural flaws, and missing citations. The audit report drives a second editing pass.
-
Do not decide what you can leave to the implementer. Package manager, task runner, deploy target, which sub-agent product to use, Node version — these are project-level choices. The skill prescribes the workflow and the VitePress content contract, not a specific toolchain.
-
Interactivity is not a luxury. If a concept has a "you only get it when you touch it" quality (state transitions, re-render cycles, diff algorithms, data flow, snapshot semantics), a Vue component embedded in the page is almost always higher-value than another paragraph of prose. VitePress supports local <script setup> imports of .vue components from docs/.vitepress/theme/components/** — no global registration needed. Push writers to evaluate interactive demos per chapter, not as a one-off afterthought. SVG illustrations (including SMIL <animate> for simple animation) are a second-tier option when interactivity is overkill but a static Mermaid box-and-arrow underperforms.
-
Tone is a parameter, not a default. Different readers want different registers — some want dry-technical, some want sempai-talking-you-through-it. Never assume. The interview pins this down (checklist §4.5) and the STYLE_GUIDE encodes the chosen register so writers stay consistent across chapters.
The nine phases
Work through these in order. Do not skip phases. Phases 1 → 3 happen before any file is written; Phase 7 is triggered by Phase 6 outcomes; Phase 9 is optional.
| # | Phase | Output |
|---|
| 1 | Discovery interview | Written decisions record |
| 2 | Source research | Source-to-topic binding table |
| 3 | Plan & user sign-off | Plan file; user-approved TOC |
| 4 | Project bootstrap | Running VitePress skeleton with empty chapter files |
| 5 | Parallel writing | Populated Markdown + diagrams |
| 6 | Build + link verification | Clean build, zero dead links |
| 7 | Browser verification | Zero console errors; rendered Mermaid/math confirmed in DOM |
| 8 | External audit + fixes | Audit report applied; up to 2 rounds |
| 9 | (Optional) Deployment config | Deploy config file |
Phase 1: Discovery interview
The user's first message is almost never a full spec. Grill them.
Read references/interview-checklist.md for the full question catalog. At minimum, secure explicit answers (not assumptions) for:
- Topic + motivation (why learning this? is there an ulterior project driving it?)
- Scope boundaries (what's explicitly out?)
- Audience (self? team? beginner? advanced?)
- Content language(s) (and whether to i18n)
- Voice / tone (formal-dry? neutral-technical? warm-sempai? — do NOT default to one, ask). This is NOT a cosmetic decision; getting it wrong is the difference between "this feels for me" and "this feels cold". See interview-checklist §4.5.
- Math rendering needed?
- Diagram density (Mermaid-heavy, or mostly prose?)
- Interactive / animated content (SVG illustrations? animations? Vue components for live demos?) — for topics with "touch-to-click" moments (state machines, diffs, rendering, data flow), actively push for interactivity. Static figures alone leave readers unable to build intuition. See interview-checklist §7 (expanded).
- Bridge/application chapter? (mapping this topic to the user's ulterior project, if any — this was the highest-value chapter in a past engagement)
- External audit desired? (default: yes)
- Depth of research per topic (skim vs. spec-grade)
- Deploy target or local-only?
Use AskUserQuestion for decisions with >2 plausible answers. Batch up to 4 questions per call only when the answers are independent; otherwise ask one at a time so later questions are informed by earlier answers.
Scope drift is the #1 killer. Pin the user on what's out before what's in. If they hedge (e.g., "maybe also Y?"), push back: either commit to Y or explicitly defer it.
Phase 2: Source research
Before writing a single content line, build a source-binding table: for each planned section, what primary source(s) will ground it?
Read references/research-discipline.md for source-hierarchy rules and for the primary-vs-secondary test.
Typical primary sources by domain:
- Technical APIs → official headers, SDK docs, spec repositories, source code
- Web standards → W3C / WHATWG / ECMA-262 specifications
- Audio/signal processing → IEC / AES / ITU standards; established textbooks
- Languages / frameworks → official language specs and reference implementations
- Platforms / products → vendor's own documentation and release notes
Use WebFetch (or project-appropriate tooling) to verify URLs exist and contain the claimed content — do not bind a source without spot-checking it.
Output: a table (can be inline in the plan file) mapping each planned section to its source URLs / file paths.
Phase 3: Plan file & user sign-off
Write a plan file covering:
- Context — why the user wants this, what the site enables
- Audience profile — explicit
- Scope boundaries — in and out, with the user's words
- Tech stack decisions — made in Phase 1
- Site structure — Part/Chapter hierarchy with short descriptions per section
- Source bindings — or a pointer to them from Phase 2
- Execution phases — concretely, how Phases 4–9 will be run for this project, including which sub-agents will be spawned and what each will own
- Verification plan — how Phases 6 & 7 will be executed
Present the plan to the user. Do not proceed to Phase 4 until the user explicitly signs off. This is the last cheap moment to restructure.
Phase 4: Project bootstrap
Stand up a VitePress project with:
- Sidebar configured for the approved site structure
- Mermaid plugin configured (with the known-good Vite workaround — see
references/bootstrap-recipe.md)
- KaTeX set up if Phase 1 decided math is needed (MathJax is incompatible with Vue's template compiler — see gotchas)
- Empty Markdown files for every planned chapter, with valid frontmatter
- A
STYLE_GUIDE.md checked into the site itself (sub-agents read this before writing)
- A landing page that maps Part → Chapter → link
Read references/bootstrap-recipe.md for the complete recipe, including: the Mermaid dayjs fix, KaTeX plugin choice, YAML frontmatter rules that bite (colons in titles!), recommended cleanUrls behavior, and how to keep audit reports out of the build.
Verify: the skeleton builds with zero pages written, and the sidebar renders the full structure.
Phase 5: Parallel writing
Dispatch one writing sub-agent per Part (or per logical chunk of work). Every agent gets the same core discipline via the prompt; only the file-ownership and per-Part content outline differ.
Read references/writing-agent-template.md for the full prompt skeleton and assets/STYLE_GUIDE.template.md for the style guide template the agents read.
Core discipline every agent gets:
- Must read
STYLE_GUIDE.md before writing
- Must
WebFetch primary sources before writing any factual claim
- API names, signatures, quoted text must be copied from primary sources verbatim — never reconstructed from memory
- Every Markdown file ends with a
## Sources section listing specific URLs/files with #section anchors where possible
- Unverifiable claims carry an inline
> NOTE: unverified — needs confirmation marker
- Every
##-level section has at least one diagram if diagrams serve the reader (but diagrams must add structure, not paraphrase prose)
- No touching files outside the agent's file list (prevents clobber in parallel runs)
- No running dev/build/deploy commands (reserved for the orchestrating session)
Dispatch agents in one parallel batch (a single message with multiple subagent invocations) so they run concurrently. Do not wait between spawns.
Phase 6: Build + link verification
After all writing agents complete:
- Run the build. It will likely fail with dead-link errors — agents often typo cross-file paths (common: guessing filenames instead of using ones from the skeleton). Fix these by normalizing to the skeleton's filenames.
- If your stack has a
check / format / lint step, run it and fix formatting issues.
- If the user's audit report file(s) live inside the docs tree, exclude them from the build (they contain path-like strings the link checker interprets as intra-site links). Use the VitePress
srcExclude option.
A passing build is necessary but not sufficient — proceed to Phase 7.
Phase 7: Browser verification
Start the dev server (in the background) and drive it with a browser automation MCP (Chrome DevTools MCP, Playwright MCP, etc.). Read references/browser-verification.md for the full checklist and exact probe patterns.
Must verify:
- Console: zero errors on the landing page and on at least one representative chapter page per Part
- Mermaid: SVGs with
id^="mermaid" actually present in the DOM
- Math (if enabled):
.katex spans render and have the KaTeX stylesheet loaded (check document.styleSheets for a katex-containing href)
- Spot-check screenshots: a page with complex math and a page with complex diagrams — look for obvious layout breakage
If any check fails, stop and fix the underlying config/plugin issue. Do not declare Phase 7 done just because the build passed. That was the #1 mistake in the reference engagement.
Stop the dev server when verification is complete; leave the environment in a state where the user can docs:dev themselves.
Phase 8: External audit + fixes
Dispatch an independent reviewer — a different agent/LLM than the ones that wrote the content. Common choices: Codex (if available as an agent type), a fresh general-purpose agent with explicit instructions, or asking the user to run a second Claude instance. The key property is independence — the writer's blind spots are not the reviewer's.
Read references/external-audit-template.md for the audit prompt template.
Audit scope:
- Structure: is the Part/Chapter ordering sensible for the target audience? are there missing topics? redundancies?
- Factual accuracy: spot-check APIs/specs/versions against primary sources
- Scope discipline: is "common" content bleeding into "specific" sections (or vice versa, if the site has that split)?
- Link & term integrity
- Unverified markers: enumerate all
> NOTE: unverified residuals — each is a debt
- Diagram utility: do diagrams help, or are they decorative?
The audit writes a report to a dedicated file classified into Critical / Major / Minor + remaining unverified markers.
After receiving the report:
- Dispatch targeted fix sub-agents for clusters of findings (e.g., one agent for all VST3-API-accuracy fixes across files, one for all Web-spec fixes, one for all unverified-marker resolutions). Each fix agent re-consults primary sources.
- After fixes, re-run Phases 6 and 7.
- Optionally run a second audit round with the same independent reviewer, narrowed to "verify prior findings are addressed and no new issues were introduced". Cap at 2 rounds to avoid infinite loops.
Phase 9 (optional): Deployment config
Only if Phase 1 included a deploy target.
Read references/deployment.md for deployment-target specifics.
The single deployment gotcha worth stating upfront: if the project is managed with a globally-installed CLI wrapper for scripts (some toolchains do this), that CLI will not exist on deploy runners. The fix is to write the deploy config (e.g., vercel.json) with commands that go through the project-local binary directly rather than the global wrapper.
Gotchas catalog
Read references/gotchas.md for the full catalog of issues encountered in prior engagements. Key items to know about upfront:
- Mermaid + VitePress + dayjs: silent dev-mode SyntaxError unless Vite optimizeDeps is configured
- MathJax + Vue: MathJax's SVG output contains
<style> which Vue's template compiler refuses to render — KaTeX is the pragmatic path
markdown-it-katex version: the old 2.x emits HTML structure incompatible with modern KaTeX CSS; use @vscode/markdown-it-katex- KaTeX CSS load timing: theme-entry imports can be flaky with some hot-reload paths;
head <link> to the CDN is the most reliable default
- YAML frontmatter titles with colons:
title: Part 0: Intro breaks YAML; quote as title: "Part 0: Intro"
- Writing agents typo filenames: always catch in Phase 6 build; build-time dead-link detection is the safety net
- Audit report files inside docs/: absolute-path-looking strings confuse the link checker; exclude them from the build
References & templates
references/interview-checklist.md — Phase 1 question catalog with rationalesreferences/research-discipline.md — primary vs. secondary source rules; per-domain source hierarchiesreferences/bootstrap-recipe.md — the proven VitePress config with Mermaid + optional KaTeXreferences/writing-agent-template.md — sub-agent prompt skeleton for Phase 5references/browser-verification.md — Phase 7 DevTools probe patternsreferences/external-audit-template.md — Phase 8 auditor promptreferences/deployment.md — deploy target specificsreferences/gotchas.md — full catalog of known issues with root causesassets/STYLE_GUIDE.template.md — style guide template to copy into the generated siteassets/vitepress-config.template.ts — starting config.ts with Mermaid + optional KaTeXassets/vercel.json.template — deploy config template (if applicable)
Reminders when executing
- The user has almost certainly thought about this topic longer than you have. Ask, don't tell.
- If the user says "just vibe" or "don't grill me", still pin down scope boundaries and audience before writing — those two decisions cause the most rework.
- Parallelize ruthlessly in Phase 5: one message, N sub-agents. Serial dispatch wastes wall-clock.
- When in doubt during Phase 7, screenshot and show the user. Build logs lie; rendered pixels don't.
- When the audit reports 30+ unverified markers, treat that as a signal that Phase 2 research was too shallow for that Part — consider re-research rather than one-by-one triage.
