5 Claude Code Skills Every Creator Needs

Every skill in the video, free. Clone the repo, drop the folder in, done.

Repo: https://github.com/austinpeechatt/creator-skills

---

Install (2 minutes)

Paste this in your terminal:

git clone https://github.com/austinpeechatt/creator-skills.git
mkdir -p ~/.claude/skills
cp -R creator-skills/skills/* ~/.claude/skills/

Restart Claude Code. Type / and you’ll see them.

Want them in one project only instead of everywhere? Copy into .claude/skills/ inside that project.

---

The six skills

They’re one pipeline, in this order:

idea → diagram → record → edit → clip → package → publish

1. notion-brain — capture

Ramble at it. It writes a structured idea into your Notion database — title, premise, angle, notes. Six weeks later you open it and the thinking is already there.

Ideas die when they don’t have a home. This is the home.

Needs: a Notion integration token and a database. Important gotcha — Notion integrations only see pages you explicitly share with them. Create the integration, then open your database and share it with that integration. Most setup failures are this.

2. excalidraw-diagram — visualize

Prompt it, get editable Excalidraw diagrams. The default is eight concepts in one file, not one.

That sounds wasteful. It isn’t. You delete six every time — and recognising the right diagram takes about twenty seconds, while describing the one you want takes three rounds of re-prompting. Elimination beats iteration.

Needs: Python 3, and Playwright if you want PNG renders.

3. video-editor — cut + graphics

Transcribes your take, finds every um, uh, false start and dead gap, and cuts them. Then places motion graphics on exact phrases.

The part that matters: it knows which scenes are talking head and which are screen recording. Graphics go left and right of you, never over your face, and never on top of a screen share where they’d compete with what you’re showing.

Needs: Python 3, FFmpeg, Playwright, and an ElevenLabs key.

On transcription — a local Whisper build (brew install whisper-cpp) is fine for a quick transcript. This skill uses ElevenLabs Scribe instead because Scribe returns word-level timestamps. That granularity is the whole thing: it’s what lets it cut filler precisely and land a graphic on the right word.

4. finish-shorts — clip

Point it at a finished long-form. It scans for the short-worthy moments, plans graphics and cutaways for each, and renders them in parallel.

Same face rule as above, rotated for vertical: on a full-frame face beat, captions pin to the top. On a split-card beat, they sit at the split. Never across your face.

Needs: same stack as video-editor.

5. youtube-thumbnail — package

Four different directions per run, plus a comparison grid so you can pick fast.

Two things that actually move quality:

• Feed it references. Your own high-performing thumbnails first. That single change did more than any prompt tweak.
• Keep the person description short. One or two sentences. Long descriptions of facial features push the model toward a generic, stylised face.

Be realistic: image models still can’t reproduce a specific person. Treat the generated face as a layout placeholder and composite your real photo in afterwards. Judge these on composition, type and colour.

Needs: a Google AI Studio key. Renders run about 4 cents each.

6. buffer-scheduler — publish

Schedules shorts to YouTube, Instagram and TikTok, and health-checks the queue.

The thing that will bite you: Buffer re-fetches your media at publish time, not when you schedule. A post set for next Tuesday downloads that video next Tuesday. If the URL is dead by then, the post fails and sits in error until you happen to look.

So host your media somewhere permanent you control — a static site on Netlify, Cloudflare Pages, S3. Temporary file hosts expire and silently break scheduled posts. That’s exactly why health_check.js exists.

Needs: Node 18+, a Buffer account, permanent media hosting.

---

The part that actually matters

None of the above is why the output doesn’t sound like AI.

Underneath all six sits an Obsidian vault — how I talk, what I’ve already covered, which hooks are burned, my voice fingerprint. Every skill reads from it. That’s the difference between automation and slop.

Same tools, no vault: “In this video we’ll dive deep into the transformative world of AI-powered content creation.” Technically fluent. Nobody wrote it.

You’re not automating the content. You’re giving the automation a memory of you to work from. Build the vault first — everything else compounds on top of it.

---

If something breaks

Open an issue on the repo and I’ll take a look. Each skill has a .env.example showing exactly which keys it needs — copy it to .env and fill it in.