How to Build a Claude Notion Graphics Skill

A Claude Notion graphics skill can add finished animated b-roll to an existing video, matching the rhythm of the spoken words. The workflow shown in this video uses Claude Code, a skill folder with a skill.md instruction file, and a video plus transcript upload to produce five animated clips in about 15 minutes.

What does a Claude Notion graphics skill actually do?

A Claude Notion graphics skill is a folder of instructions and scripts that teaches Claude, Anthropic AI assistant, to read a video and its transcript and then generate animated b-roll clips that sync with the spoken words. The video, published on September 25, 2026 by the channel's creator, demonstrates the skill on a 33-second talking-head introduction and reports five finished clips from a single run.

The skill shown in the video is b-roll specific. It overlays or fills the screen with animated graphics while the original audio keeps playing. The creator says two more skills are in progress: one for product launches and one for faceless explainer videos, planned for release over the following week as of the September 2026 recording.

The claimed result is specific: graphics appear exactly when certain words are spoken, the animation style stays consistent, and the creator says he made no follow-up prompts. That is speaker-reported experience from the transcript, not an independently verified benchmark.

How the b-roll skill works inside Claude Code

The skill runs inside Claude Code, Anthropic agentic coding tool that works in the terminal and a desktop app. A file called skill.md is the core instruction set. According to the creator, it tells Claude how to read the video and transcription, generate the graphics, and keep them synchronized with the rhythm of the words.

The signature technique is a shape-morphing method: one object changes continuously into different forms, and each shape fuses into the next. The creator credits this continuity for the polished, engaging Notion in the intro demos. Supporting folders hold scripts for generating the graphics, reading the video, rendering the output, and examples he collected along the way.

The skill also reads the composition of the source footage. When the creator cut himself into a side third of the frame, the skill placed graphics in the empty space. When the frame is full, it can cover the whole screen. That placement awareness is embedded in the skill instructions, not prompted at runtime.

Step-by-step: running the skill on your own video

The creator's workflow follows the same order every time. He recommends working inside a persistent project folder so reusable assets and lessons accumulate over time.

1. Create a new project folder in the Claude Code desktop app, such as a YouTube demo folder, and enter code working mode.
2. Open the skill's GitHub repository, copy its URL, and ask Claude to read the skill and import it into the workspace, letting it set up any AGENTS.md or Claude.md files it needs.
3. Upload your raw video and its timestamped transcript, then ask Claude to run the b-roll skill on that media.
4. Choose a clip density and a color palette. Medium produced about three clips; intense produced four to five on the 33-second test video. First-time users can upload brand colors and style assets here.
5. Review the clips. The default skill output is downloadable files; the creator asked Claude to embed the clips in an HTML page so he could preview them inside the local workspace.
6. Approve the plan or re-prompt, then keep, remove, or reorder the individual clips in your editor.

On the demonstrated run, generating all assets took about 15 minutes, based on the creator's own estimate in the video. He chose intense density deliberately, preferring to cut clips later rather than ask for more.

Effort level, density, and what the 2026 test showed

The creator used the default medium effort setting on Claude 5.5, the model version named in the video published September 25, 2026, and says he has not tested the skill at low or high effort. Everything in the intro demonstration was produced at medium.

The video also contains an unrelated segment describing an Age of Empires II replication test across reasoning effort levels, where builds at different effort levels differed in cost, time, and token consumption, and a bonus build used a master prompt assembled from notes. Those cost and token figures belong to that separate test and should not be read as numbers for the Notion graphics workflow.

Density is the main cost lever you control. Three clips at medium density versus four to five at intense density changed how much b-roll the 33-second clip received. Starting intense and trimming is the creator's stated preference.

Transcription: uploading a transcript vs. a local model

The skill does not yet generate its own transcript, so you supply one with the video. The creator mentions two paths: send the audio to the ElevenLabs API, or run a local transcription model on your own machine.

He says he wants to wire a local model into the skill but has not solved that step yet, so the demonstrated workflow uploads a transcript file manually. For privacy-sensitive footage, a local transcription model keeps the audio on your machine, though the video and graphics generation still happen in your Claude session.

The timestamped transcript matters because the skill uses it to time graphics to specific words. Without it, the sync that makes the b-roll feel deliberate would be guesswork.

What the skill cannot do yet

Scope is limited by design. This skill only inserts b-roll-style Notion graphics into a video you already edited. The creator did the initial cuts and the full-frame-to-side-panel transition in his own video editor before uploading.

Transcription is manual. There is no built-in automatic transcription path as of the September 2026 video, only the API or local-model options described above.

Broader automation is a roadmap, not a shipped feature. The creator's long-term goal is a pipeline where Claude cuts a raw talking-head video, makes it engaging, produces the graphics, and does what a human editor would do. He describes this as work in progress, with results to be shown once production-ready. Treat any claim about that future pipeline as a plan, not a current capability.

Finally, the quality claims in the video are one creator's experience on one 33-second clip. Results will vary with footage, prompt, and model version, and the skill's own output defaults to downloadable files rather than the HTML preview he requested.

FAQ: common questions about this workflow

• Does the skill work without a transcript? No. The demonstrated workflow requires you to upload a timestamped transcript with the video. The creator says automatic transcription is a known gap he plans to solve with a local model, but it was not working when the video was recorded.

• Which model and effort level were used? The video shows Claude 5.5 at the default medium effort. The creator states he has not tested the skill at low or high effort.

• Do I need to edit my video first? Yes, for now. You should make your cuts and any layout changes, such as moving yourself to a corner of the frame, before uploading. The skill handles only the Notion graphics layer on top.

• How long does one run take? The creator estimated about 15 minutes to generate all assets for a 33-second clip with five b-roll clips at intense density. That is his report from one session, not a published benchmark.

• Can I customize the visual style? Yes. The skill accepts a default palette, and the creator recommends uploading your brand colors, style, and other assets during the first session so the graphics match your identity.

Turn your own videos into written articles

The workflow above turns a finished video into animated graphics; the same video often holds explanations and lessons that never make it into text. If you publish on YouTube and want that knowledge in article form, Skala Blog follows the same logic in reverse: paste a YouTube URL, get a transcription, and generate a readable article from what you already said.

This piece itself started as a video transcript in Portuguese, shared in connection with Dev Doido do canal do youtube and the tooling at crazystack.com.br. If you have interviews, tutorials, or opinions sitting in your upload archive, a transcription-and-article pass is the simplest way to make them searchable.

Source video