content-vault

The consolidated GTM Architecture Kit is now maintained in timscheuerai/gtm-architecture-template. Start there for the shared skills library, editable functions, blank context and first-batch demo. This repository retains its earlier material and history.

19 Claude Code skills + a pre-configured Obsidian vault. Same setup we use to run our content.

Free. MIT. Fork it.

researcher            repurpose             lead-magnet-creator
linkedin-copywriter   x-copywriter          long-form              newsletter-writer
youtube-script        youtube-description   youtube-thumbnail      youtube-publisher
graphics-designer     launch-video          video-use
shorts-edit           shorts-cut            shorts-audio           shorts-motion   shorts-qa

Two ways to run this

Oxygen (hosted). Team setup, one-click import, integrations already wired up. oxygen-agent.com

Claude Code (local). Clone the repo into your skills folder.

git clone https://github.com/timscheuerai/content-vault.git ~/.claude/skills/content-vault

Then open that folder in Obsidian — it ships a pre-configured vault (Kanban board, graph, backlinks, search; no plugin setup). That's where your content lives and is tracked. See The Obsidian side.

Then in Claude Code:

/linkedin-copywriter   draft a post about my Q2 launch
/researcher            find 10 ideas for next week
/repurpose             turn yesterday's webinar into LinkedIn + X

The setup that actually matters

Without this step the writer skills sound like generic AI. Run it once and you're done.

Connect LinkedIn

Oxygen: already done.

Claude Code: plug in Unipile or whatever you use.

export UNIPILE_DSN=...
export UNIPILE_API_KEY=...
export UNIPILE_ACCOUNT_ID=...

Tell Claude to fill your corpus

Pull my last 20 LinkedIn posts via Unipile, sort by reactions, write them into skills/linkedin-copywriter/corpus.md using the format in corpus.md.example.

Claude pulls, sorts, writes the file. From the next draft, linkedin-copywriter writes in your voice.

Same for X

export X_BEARER_TOKEN=...

Pull my last 30 tweets, originals only, sort by engagement, write them into skills/x-copywriter/corpus.md.

Done. Skills are personalized.

Video + YouTube skills (optional)

Only if you use them. Each has a ## SETUP block in its SKILL.md.

export ELEVENLABS_API_KEY=...   # youtube-description (transcripts), video-use, shorts-*
pip install rembg               # youtube-thumbnail (face cutouts)
brew install ffmpeg             # shorts-*, video-use
pip install pillow numpy        # shorts-* (the graphics + caption renderers)

youtube-publisher talks to your own channel, so it needs your Google Cloud project + a YouTube Data API OAuth client (client_secrets.json). Walkthrough is in the skill.

What's in the vault

linkedin-copywriter. Drafts LinkedIn posts in your voice. Hook frameworks, body shapes, CTA patterns, AI-slop blacklist.

x-copywriter. Drafts X tweets and threads. Different game than LinkedIn. The skill teaches you the platform as it drafts.

youtube-script. Spoken scripts for YouTube longs, shorts, talking-head clips. Timed beats with B-roll cues.

youtube-description. Transcribes a finished cut (ElevenLabs Scribe), writes a 2-sentence description, your fixed CTA, and chapter timestamps.

youtube-thumbnail. Two-face interview/podcast thumbnails. rembg cutouts, one highlight word, 1280×720 dark + light. Encodes verified 2026 CTR research.

youtube-publisher. Uploads a video or a whole numbered series via the YouTube Data API. Titles, descriptions, thumbnails, ordered playlist. Resumable, idempotent.

long-form. Newsletter issue, blog, Substack. Outline first, prose second.

newsletter-writer. Lifecycle + onboarding emails and broadcasts. Same voice as your posts, rendered to your email template, optional push to Resend.

lead-magnet-creator. Builds free assets in 10 formats. Notion, PDF, Sheet, GitHub starter, GPT, web tool, video. The skill that built this vault.

repurpose. One master piece (webinar, transcript, post) into N channel variants. Hub-and-spoke.

researcher. 3 modes. What's hot in your space. What's working for you. What customers said in interviews.

graphics-designer. On-brand graphics. HTML to PNG via headless Chrome.

launch-video. 30-60s motion graphics. Remotion + ElevenLabs.

video-use. Edit any video by chat. Cuts on word boundaries, grades, burns subtitles. Vendored from browser-use/video-use.

shorts-edit (+ shorts-cut, shorts-audio, shorts-motion, shorts-qa). Raw talking-head recordings → postable vertical shorts. Drops the retakes and the dead air, cleans the audio, burns word-by-word captions, adds cue-anchored motion graphics — in one video encode — then proves it with a 7-check QA gate. Built on 14 OBS recordings, 57 minutes of raw. Every rule in it exists because a defect got through review looking fine. See Editing shorts.

Editing shorts

The five shorts-* skills are one pipeline. Put your raw recordings in raw/, then:

/shorts-edit   cut these OBS takes into vertical shorts

<project>/
  raw/     source recordings (untouched)
  clips/   <name>_final.mp4   <- deliverables
  edit/    every decision, as a file you can inspect and re-run

shorts-edit orchestrates; shorts-cut picks the last complete attempt of each beat and kills dead air; shorts-audio fixes the two audio traps that cost hours; shorts-motion draws cue-anchored graphics; shorts-qa is the gate.

The gate is the point. Every bug in this pipeline's history was invisible to eyeballs and obvious to a measurement — a caption verifier that sampled long words reported perfect while short words were silently wrong; a tier grade was revealed 5.7s before it was spoken and every frame looked fine in isolation. So nothing ships until qa.py passes. Not "looks good" — passes.

Graphics render against skills/shorts-edit/assets/BRAND.md, which ships conformant with a neutral palette. Edit it to your brand (or point SHORTS_BRAND_MD at the one you already keep) and brand_lint.py fails until the renderers follow — including the rule no type checker can catch: exactly one accent colour per graphic.

The Obsidian side

Skills produce content. Your Obsidian vault tracks it — no external tool, no sync, every piece a markdown file you own.

Open this folder in Obsidian (it ships pre-configured) and content/Content.base gives you a Kanban board grouped by status (Idea → Drafting → Review → Published → Backlog), plus table and card views. It's a native Obsidian Base — no community plugin needed. Same schema the Notion DB had, now as frontmatter on files you own:

title · status · pillar · format · channel · author · publish · drive

New drafts land in content/posts/. Want fast semantic + keyword search over your whole back catalogue (and for agents, via mcp__qmd__*)? Run ./scripts/qmd-setup.sh once, then qmd query "...".

The content graph (your ideas compound)

Tags and pillars slice your posts. They don't connect them. So the vault also ships a knowledge-graph layer that mirrors the LLM-wiki pattern onto your content:

content/concepts/   one note per durable idea (your thesis, in pieces)
content/entities/   recurring people, customers, competitors
content/mocs/        narrative arcs that thread posts into a story

Each concept note lists the posts that express it. Obsidian's backlinks make that bidirectional automatically, so every post shows its concepts in the backlinks pane without you editing the post. Open the graph view and watch it thicken.

The point is the loop: the writer skills pull the relevant concept as grounding before you draft (so you build on your sharpest prior framing instead of a blank page) and deposit new ideas back after (so the canon grows underneath you). /researcher gets a fourth mode that mines the graph for the next unwritten beat. Start empty, publish, and in a few weeks the graph is your content thesis.

How it works + templates: content/CONTENT-WIKI.md. The vault ships one worked example-* note per type to copy.

Need help?

Built by us at Oxygen. Stuck? hello@oxygen-agent.com

MIT. Use it however you want.