The Claude Content Studio

Step two. Make a folder for this work. Call it anything. Open the terminal, move into that folder, type claude and press enter.

Step three. Install the brand skill:

npx skills add hyperfx-ai/marketing-skills --skill brand-context

Installing just this one skill matters, because most of the others in that repo are built to run against a paid connector at forty nine dollars a month, and brand-context is the one that works entirely on its own.

Step four. Fill it in. Ask Claude:

Use the brand-context skill and interview me. Ask one question
at a time and write my answers into the brand context file.

Answer properly. This file is read by every stage that follows, and it is the difference between output that sounds like your business and output that sounds like nobody.

The questions that matter most are the specific ones. Not "who is your customer" but who they were before they found you, what they nearly bought instead, the objection that kills most of your deals, and the exact phrases your best customers use when they describe what you do. Twenty minutes here changes everything downstream.

Step five. Install the video skills:

npx skills add remotion-dev/skills

Step six. Make your first video:

Create a Remotion project. Build a fifteen second vertical video
with a headline, a subheading and a background colour from my
brand context. Show me a preview before rendering.

Claude will pull down what it needs on first run, including the video encoder and a headless browser. That download takes a few minutes once and never again.

Step seven. Preview, then render:

Open the studio so I can watch it, then render it to a file.

If a video file appears in your folder, phase one is done and you already have the most useful part.

Phase two, the writing pipeline

Step one. Install the writer:

npx skills add ComposioHQ/awesome-claude-skills --skill content-research-writer

Name that specific skill rather than installing the whole repository. Most of what is in there is a directory of links to other people's skills, which is useful for browsing and useless as an installed thing.

Step two. Install the coordinator:

npx skills add arturseo-geo/content-pipeline-skill

An honest note on this one. It has around twelve stars and one maintainer, and it describes itself as six specialist agents covering research, writing, editing, search optimisation, analytics and a final quality check. It works, and it is the least proven thing in this document. If it breaks or goes unmaintained, a written checklist does the same job with more effort and no dependency.

Step three. Run a brief through it:

Use the content pipeline to turn this brief into a script.
Read my brand context first. Brief: [YOUR BRIEF]. Stop and
show me the outline before you write the full draft.

The stop-and-show instruction is doing real work. Left alone it will produce a finished draft off a misunderstanding, and reading an outline takes thirty seconds.

Phase three, research and competitor work

This phase is genuinely optional and it is where the friction lives.

The researcher

gpt-researcher is a separate program written in Python rather than a skill, so it needs Python 3.11 or later on your machine, plus its companion connector, plus two API keys. One for a language model and one for search, usually Tavily, which gives you a thousand free credits a month with no card.

Ask Claude to walk you through it:

Help me install the gpt-researcher MCP server from
assafelovic/gptr-mcp. Check my Python version first, tell me
exactly which API keys I need, and walk me through one at a time.

The honest alternative. Claude has web search built in, and the content-research-writer skill from phase two produces sourced drafts on its own. For most people that covers it. gpt-researcher goes deeper and is worth it if research is a large part of your work, and it is the single hardest install here.

Competitor ads, and the part to be careful about

The appealing version of this is pointing a browser at the Meta Ad Library and collecting what competitors are running. In practice that is the least reliable step in the whole build.

Ad libraries defend themselves with rate limits, IP blocks and CAPTCHAs, the page structure changes often enough to break automation, and scraping the site rather than using the official route may breach the platform's terms, particularly if you are logged in. The official Meta API is the clean route and it comes with its own problems: it requires government ID verification taking one to three working days, and for ordinary commercial ads in most markets it returns very little.

So treat competitor research as human work with AI assistance rather than an automated stage. Open the ad library yourself, screenshot what is running, and hand the screenshots to Claude:

Here are twelve competitor ads. Break down the hook, the offer,
the format and the call to action in each one, then tell me what
angle nobody in this set is using.

That takes fifteen minutes, produces a better answer than scraping would, and carries no terms of service risk at all. The browser connector is still worth installing for reading normal web pages, which it does well:

claude mcp add playwright npx @playwright/mcp@latest

Running the whole thing

Once it is built, a project looks like this.

You bring the brief and the competitor screenshots. Claude reads your brand context, researches the market with sources, writes an outline you approve, produces a draft, edits it, and hands you a script. You approve the script. Claude builds the video composition, adds captions and audio, previews it, and renders the file.

Five stages, three approval points, one window.

What it does not do is run unattended. Nothing here automatically triggers the next stage, and the value comes from the stages existing and being consistent, not from you being absent. Anyone selling this as a machine that produces finished ads while you sleep is selling you something that does not exist.

What this actually costs

A solo operator runs this for the price of a Claude plan. Twenty dollars a month, free search credits, Remotion free, nothing else. That is the honest floor.

A five person business pays the Claude plans, plus Remotion seats for anyone touching it, plus search if you use deep research. Realistically a few hundred a month all in, against a creative retainer that starts several thousand.

The licence rule that catches people out

This deserves its own section because it is the one thing here that can quietly put you out of compliance.

Remotion, the video engine, is not free for everyone. Its licence states that individuals, companies with three or fewer employees and non-profits may use it commercially for free, and that for-profit companies with four or more employees need a paid Company Licence.

The threshold is headcount, not usage. A solo freelancer selling videos made with it pays nothing. A four person company using it internally has to buy licences. Current pricing is twenty five dollars per seat per month for normal use, with a separate usage-based tier for anyone rendering at volume through an automated pipeline.

Nothing enforces this in the software. You will not be blocked and nobody will email you. It is on you, and the moment your company hits four people it becomes a real obligation.

The three artifacts you need

One, a completed brand context file. Section four. Every stage reads it and the whole system is capped by its quality.

Two, a competitor screenshot folder. Twelve to twenty current ads from your category, refreshed monthly. This replaces the fragile scraping step with something better and takes fifteen minutes.

Three, a written creative brief template, 200 to 400 words, and almost nobody has one. One page with fixed headings: what the video is for, who it is aimed at, the single idea it must land, what it must never say, the call to action, and the format and length. Fill it in for every project before you touch Claude. The reason this matters more here than anywhere else is that the pipeline will confidently execute a bad brief all the way through to a rendered video, and you will have spent an hour discovering that the brief was wrong. Write the template once and reuse it forever.

If something goes wrong

Honest limits

The pieces do not hand off to each other automatically. Six separate tools with different requirements live in one window, and you are the thing connecting them. That is still a genuine improvement over five browser tabs and a video editor, and it is not the autonomous studio the framing sometimes suggests.

The coordinator skill is the weakest link by adoption, with roughly twelve stars and a single maintainer. Everything else here is backed by Microsoft, a well-known open source project, or a company with a commercial product behind it.

Competitor ad scraping is the step most likely to fail, and the manual alternative in section six is better anyway.

There is a copyright consideration worth knowing. In March 2026 the Supreme Court declined to hear a case on whether AI can be an author, leaving standing a ruling that work created entirely by AI with no meaningful human involvement gets no copyright protection in the United States. You can still use and sell such work. You may not be able to stop others copying it. Human creative choices layered on top do attract protection, which is another reason the approval points matter.

Disclosure rules for AI-generated advertising are tightening across Google, Meta and TikTok, and EU obligations on marking AI content applied from August 2026. Check the current requirement on whichever platform you are publishing to.

And the honest bottom line. This produces a competent first draft of everything, fast and cheaply. It does not have taste, it does not know your market, and it will render a beautiful video of a bad idea without hesitating once.

What this changes

The expensive part of creative work was never the execution. It was that execution took so long that testing an idea properly was unaffordable, so people made one thing carefully and hoped.

What this setup changes is the cost of a version. When research, script and render collapse into an afternoon, you stop protecting your first idea and start running five. The constraint moves from how much you can produce to how well you can judge what came back, which is the part no repository in this list can do for you.