Is this a slash command or a skill?

Same thing now. Anthropic merged custom commands into skills, and their documentation says it outright: a file at .claude/commands/deploy.md and a skill at .claude/skills/deploy/SKILL.md both create /deploy. Skills are the current form and the one worth building on, so that is what the builder writes. The directory name becomes the command you type.

Does this work if my site is not built with Next.js?

The builder reads your repository before it proposes anything, then asks you to confirm the publish path rather than trusting what it found. That is deliberate. We have run it end to end on a Next.js site, which is the honest limit of what we can claim. On any other stack you are leaning on that confirmation step, which is where a wrong guess gets caught before it hardens into every page you publish after it.

What does it cost per video?

Nothing beyond the Claude plan that Claude Code already requires. Transcription runs on your own hardware through faster-whisper, an MIT-licensed reimplementation of OpenAI's open-source Whisper model. There is no per-minute charge and nothing to upload.

Does my video have to show my screen?

No. This page came from a video with no screen recording in it at all: one still image sitting behind a talking head for the full ninety-eight seconds. Across forty-nine sampled frames, the steadiest region of the picture moved three tenths of one grey level out of 255. Watch the reel and you can see for yourself. The skill measures that before it writes a word, because a person gesturing in front of a frozen screenshot reads as motion and does not mean the screen changed. When nothing moves it writes an explainer, rather than narrating clicks nobody performed.

Can it publish without me watching?

No, and that is the design. It stops to confirm your publish path, tells you the transcription model size and lets you choose, asks you for the video, and stops again before the deploy. The work between those points it does on its own. An earlier version wrote pages straight from the transcript, and what came out was confidently wrong in fluent prose, with nothing on the page marking it as invented.

Where does the skill file end up?

At .claude/skills/your-name/SKILL.md in your repository, on a branch, so you can review it and merge it like any other change. Put it under your home directory at ~/.claude/skills/ instead if you want it on every project and only for you. Every frontmatter field is optional. Write a description anyway, so Claude knows when the skill applies.

Claude Code blueprint

Someone asked for the written version. Build the skill that writes it.

You post a video. Someone comments asking for the written version, and now you owe a stranger a page that does not exist yet. This is the blueprint for a Claude Code skill that writes it against your own repository: transcription running on your hardware at no cost per video, a round of questions before any copy exists, your build gates, your deploy command. One line starts the build.

Watch it in under two minutesOn Instagram. The page below stands on its own if you would rather read.

Line that starts it: 1
Per video after setup: $0
Where your audio stays: Local
Questions before it writes: 4

Where this leaves you

A skill committed to your own repository, run by typing its name with a slash in front of it. The directory name becomes the command.
Transcription on your own hardware, nothing uploaded and no charge per video, with the voice filter left on so it stops writing when you stop talking.
A round of questions before any copy exists. That is the step that keeps the page from asserting things you never said.
Your type check, your lint, your build, wired in as gates. The deploy step stops and asks before anything ships.

Do this once

The builder reads your repository before it proposes anything. Nothing here touches your live site.

You need Claude Code on a paid Anthropic plan, the repository that publishes your site, and ffmpeg plus Python 3.10 or newer for the transcription step. The builder checks for the last two and offers to install what is missing.

01
Get Claude Code onto your machine. It is a separate product from the Claude app and from claude.ai chat, and Anthropic's documentation is blunt about the plan: the free Claude.ai plan does not include Claude Code access. You need Pro, Max, Team, Enterprise or a Console account.
curl -fsSL https://claude.ai/install.sh | bash
Windows PowerShell instead: irm https://claude.ai/install.ps1 | iex
02
Open it at the repository that publishes your site. Not a scratch folder. The builder reads your real files to work out where pages live, and it cannot do that anywhere else.
claude
Commit or stash whatever is in progress first. The builder stops on a dirty tree rather than mixing its work into yours.
03
Paste one line. That is the easy button. It points Claude Code at the builder prompt and hands over.
Read https://www.graniteai.co/prompts/video-guide-builder.md and do exactly what it says.
Open that link yourself first if you would rather read what you are pasting. It is a plain text file.
04
Answer its questions about where your pages go. It works out your framework, the directory a page lives in, whether an index needs a new entry, which branch ships to the live site, and what your gates are called. Then it stops and asks you to confirm, with its own recommendation first. This step decides where every future page lands, so give it the two minutes.
05
Let it set up transcription locally. It checks for ffmpeg and Python, tells you the model download size before it downloads anything, and lets you pick a larger or smaller model depending on whether you have a GPU. Your video never leaves your machine.
No API key and no per-minute charge. faster-whisper is an MIT-licensed reimplementation of OpenAI's open-source Whisper, and it runs on hardware you already own.
06
Hand it a real video and watch the whole run. It stops before the deploy and shows you the page locally. After that the command is yours, and every later video is the same command and the same short round of questions.

The builder never pushes to your production branch. Wiring up your deploy means writing the command into the skill, never running it. The install commands and the plan requirement above were checked against Anthropic's documentation on 1 September 2026.

Questions people ask

Is this a slash command or a skill?
Same thing now. Anthropic merged custom commands into skills, and their documentation says it outright: a file at .claude/commands/deploy.md and a skill at .claude/skills/deploy/SKILL.md both create /deploy. Skills are the current form and the one worth building on, so that is what the builder writes. The directory name becomes the command you type.
Does this work if my site is not built with Next.js?
The builder reads your repository before it proposes anything, then asks you to confirm the publish path rather than trusting what it found. That is deliberate. We have run it end to end on a Next.js site, which is the honest limit of what we can claim. On any other stack you are leaning on that confirmation step, which is where a wrong guess gets caught before it hardens into every page you publish after it.
What does it cost per video?
Nothing beyond the Claude plan that Claude Code already requires. Transcription runs on your own hardware through faster-whisper, an MIT-licensed reimplementation of OpenAI's open-source Whisper model. There is no per-minute charge and nothing to upload.
Does my video have to show my screen?
No. This page came from a video with no screen recording in it at all: one still image sitting behind a talking head for the full ninety-eight seconds. Across forty-nine sampled frames, the steadiest region of the picture moved three tenths of one grey level out of 255. Watch the reel and you can see for yourself. The skill measures that before it writes a word, because a person gesturing in front of a frozen screenshot reads as motion and does not mean the screen changed. When nothing moves it writes an explainer, rather than narrating clicks nobody performed.
Can it publish without me watching?
No, and that is the design. It stops to confirm your publish path, tells you the transcription model size and lets you choose, asks you for the video, and stops again before the deploy. The work between those points it does on its own. An earlier version wrote pages straight from the transcript, and what came out was confidently wrong in fluent prose, with nothing on the page marking it as invented.
Where does the skill file end up?
At .claude/skills/your-name/SKILL.md in your repository, on a branch, so you can review it and merge it like any other change. Put it under your home directory at ~/.claude/skills/ instead if you want it on every project and only for you. Every frontmatter field is optional. Write a description anyway, so Claude knows when the skill applies.

Or hand us the video

The video guides under Tools on this site came out of the workflow this blueprint describes: a video in, one pass, a published page out. If you would rather have the output than the build, that is the part we do.

Send us the video See every tool