Amplify

Turns an Obsidian second brain into published content. Wiki in, video out.

What this is

Amplify is a CLI (amplify) and an MCP server that lives outside the vault and reads into it. It saves sources immutably, maintains an LLM wiki over them, shapes wiki material into stories, and renders those stories into video, carousels, articles, and diagrams.

This repo stores no user data. Every note, source, story, timeline, and rendered asset lives in the user's Obsidian vault. The repo holds code, templates, and skills.

Architecture

• TypeScript + Node.js, ES modules ("type": "module")
• Build: npm run build (tsc). Output goes to dist/.
• Dev: tsx for running without build
• Tests: vitest — 117 tests across config, vault, dates, signals, memory, and diagram
• No frontend. CLI plus MCP server. Remotion compositions are React, rendered headless.

Key directories

src/
  types/          ← TypeScript interfaces (config, source, wiki, amplify, signal, memory,
                    story, studio, director, diagram, brand, meeting)
  vault/          ← Read/write Obsidian markdown with gray-matter frontmatter
  cli/            ← CLI commands (init, studio, storyboard, listen, obs, import-sparks)
  mcp/            ← MCP server (stdio transport, 5 tools)
  ai/             ← Humanizer pass over generated prose
  amplify/        ← Content amplification context builder (hooks, structures, persuasion)
  carousel/       ← LinkedIn carousel HTML and PDF rendering
  diagram/        ← React Flow diagram generation (self-contained HTML output)
  studio/         ← Video production (HeyGen, ElevenLabs, Google Imagen, Remotion, assets)
  capture/        ← OBS screen and camera recording
  config.ts       ← Loads .amplifyrc with fallbacks to env vars
  index.ts        ← Public API re-exports

Conventions

• All vault paths are relative to vaultPath and resolved through src/vault/paths.ts
• Frontmatter is parsed/serialized with gray-matter via src/vault/frontmatter.ts
• Readers return typed objects, writers accept typed objects
• New vault content types follow the pattern: types file, path resolver, reader, writer, re-export in vault/index.ts and index.ts
• Config changes go in three places: types/config.ts (interface + DEFAULT_PATHS), config.ts (loadConfig merge), .amplifyrc.example. Then run npm run check:vault — it catches a path that resolves to nothing, which is otherwise silent
• The vault has two ownership layers Amplify must respect. Sources/ is immutable: read it, cite it, never rewrite it. Wiki/ is agent-owned: create and rewrite freely. Anything the user keeps outside those two is theirs; do not ingest it and do not write into it unless asked directly
• Slugs go through src/vault/slug.ts, not a local copy
• Dates: use src/vault/dates.ts, never raw Date math. Two traps it exists to avoid. new Date("2026-08-19") parses as UTC midnight and lands on the previous day west of UTC. new Date().toISOString().split("T")[0] is the UTC day, which rolls over mid-evening in the Americas. Frontmatter dates are a third trap: YAML parses an unquoted date: into a Date object, not the string the type claims, so read them through normalizeDateKey()
• CLI commands live in src/cli/commands/ and register in src/cli/index.ts

The flywheel

Every skill sits at one stage. Know which stage you are working on.

Sources/  →  Wiki/  →  Work/Stories/  →  formats  →  published  →  signals
 /ingest     /wiki      /craft-content    /draft-*   studio render   record_signal
 /meeting    /lint                                                        │
                                                                          ▼
                                                              suggest_content reads
                                                              signals + wiki to rank
                                                              what to publish next

The wiki is the input to content, not a side archive. A page with sources behind it is most of a draft already. That is why suggest_content ranks wiki pages by source count and whether a story folder for them exists yet.

Skills

Claude Code skills in .claude/skills/. Each reads .amplifyrc for vault paths.

Second brain — /ingest, /meeting, /wiki, /lint, /add-inspiration Story — /craft-content Formats — /draft-video, /draft-carousel, /draft-article, /draft-diagram, /draft-images Publish — /linkedin-post, /youtube-description, /youtube-short, /youtube-upload, /listen, /humanizer Identity — /voice, /brand, /heygen-avatar, /heygen-video Setup — /init

The wiki schema

{vaultPath}/CLAUDE.md is the vault's own schema, the contract that turns an agent with file access into a disciplined wiki maintainer. It defines the layers, the single frontmatter schema, the page shapes, the ingest/query/lint workflows, and the orphan rule ("every new page must link to at least one existing page"). The template lives at templates/vault/CLAUDE.md and is copied on /init.

Identity files

Skills read two identity files that shape all output:

• Voice.md (paths.voice) — Who the user is and how they sound
• Brand.md (paths.brand/Brand.md) — How their work looks

Keeping personal data out

This repo is public and stores none of the user's content.

• npm run check:personal scans tracked and untracked-but-not-ignored files.
• A pre-commit hook in .githooks/ runs it on staged changes and blocks the commit on a hit. npm install wires it up via the prepare script.
• Generic patterns (home paths, emails, API keys, vault paths) live in tools/check-personal.mjs. Real names live in .personal-patterns, which is gitignored — the names themselves must not enter the repo.
• package.json, README.md, and LICENSE are exempt from the name rules only, since a public repo needs its own URL and attribution. The hard rules still apply to them.
• Docs and skill examples must use invented data. Never paste a real source, colleague, story slug, or vault path into an example.
• If a match is genuinely fine, add a personal-ok comment on that line.

Writing style

When writing prose that lands in the vault (wiki pages, stories, drafts), follow the humanizer rules:

• No AI vocabulary (delve, landscape, crucial, leverage, robust, innovative)
• No filler phrases, no significance inflation, no sycophantic language
• No em dash overuse. Use commas, colons, or periods.
• Vary sentence length. Be specific. Have opinions.
• Preserve the user's voice. The humanizer is for Amplify's writing, not theirs.

Commands

npm run build        # Compile TypeScript
npm run dev          # Run CLI with tsx
npm run serve        # Start the MCP server
npm run lint         # Type check without emit
npm test             # Run vitest
npm run check:vault  # Assert every configured vault path exists
npm run lint:wiki    # Wiki structure: broken links, orphans, index drift, aliases
amplify init [path]  # Connect Amplify to a vault
amplify studio list  # List director cuts across stories
amplify studio render# Render a story's timeline to video
amplify studio preview
amplify storyboard <slug>
amplify listen <note>

Docs

• Vault — vault structure, the layers, Voice.md
• Signals — event log connecting pipeline stages
• Memory — session history, preferences
• MCP Server — all tools, Claude Desktop config
• Studio — video production pipeline, API setup, Remotion
• OBS Capture — OBS integration, screen/camera recording