The Claude Content Studio

What does a creative agency actually do that software cannot?

Strip a campaign back and there are five jobs. Somebody looks at what competitors are running. Somebody researches the market properly. Somebody keeps the brand sounding like itself. Somebody writes the thing. Somebody makes the video. Those five jobs are why a creative retainer runs several thousand a month, and most businesses pay between two and a half and ten thousand a month for a marketing agency in 2026.

Six free repositories now cover a real share of that, and they run inside one window.

Be clear about what this is before you build it. It is not a machine you feed a brief into and collect a finished ad from. It is an assembly line with you standing at the controls, moving work from one station to the next and approving as you go. That distinction is the difference between people who get value from this and people who install six things and give up.

What you are actually building

Six pieces, and it helps to think of them as staff.

The researcher. gpt-researcher investigates a market properly, checks multiple sources and returns a report with citations rather than confident guesses. Around twenty nine thousand stars, Apache licensed, actively maintained.

The brand keeper. hyperfx-ai/marketing-skills contains a brand-context skill that stores your positioning, audience, customer language, proof points and voice rules in one file. Every other skill reads it first. MIT licensed.

The writer. ComposioHQ/awesome-claude-skills is mostly a directory of a thousand-plus skills, and the part you want is a specific skill bundled inside it called content-research-writer, which turns research into a sourced draft with a proper hook.

The browser. microsoft/playwright-mcp gives Claude control of a web browser so it can look at pages a human would look at. Thirty five thousand stars, Apache licensed, from Microsoft.

The producer. remotion-dev/skills teaches Claude how to build video with Remotion, which makes videos out of code rather than out of a timeline. Claude writes the composition, adds footage, audio and captions, previews it and renders the file.

The coordinator. arturseo-geo/content-pipeline-skill runs a project through research, writing, editing and a quality check in sequence.

That is six repositories running inside Claude Code, which is the seventh thing you need and the only one that is not free.

Before you start: what these words mean

Six terms and nothing else in this document is difficult.

Claude Code. Claude running on your computer instead of in a browser tab. It can open files, run programs and control other software. It needs a paid Claude plan, starting at twenty dollars a month, and there is no free version of it.

The terminal. A plain window where you type instructions instead of clicking. Everything here is a single line you copy and paste. You never need to understand it.

A repository, or repo. A folder of files somebody published for anyone to download. All six above are repos.

A skill. A written playbook the AI reads before working. Plain English, not code. Installing a skill hands Claude a document it can consult.

An MCP, or connector. Software that lets Claude reach something outside itself, like a browser or another program. The name does not matter. What matters is that a connector is the difference between Claude describing something and Claude doing it.

An API key. A long password that lets one program use another one. You will create two or three, paste them into a settings file, and never think about them again.

The order that actually works

This is the section that saves you the most time.

Do not install all six on day one. Five of the six are useful on their own, and the two hardest installs are the two you need least. Build it in three phases, and stop whenever it is doing enough.

Phase one, about an hour, free. Claude Code, the brand context file, and the Remotion video skills. At the end of this you can describe a video and get a rendered file back. This alone is the majority of the value.

Phase two, another hour. Add the writer skill and the coordinator. Now a brief moves through research, writing and editing with structure rather than one long prompt.

Phase three, half a day, and genuinely optional. Add gpt-researcher and the browser. These are the two that involve installing Python, creating search API keys and dealing with websites that do not want to be read by software.

Anyone telling you the whole thing takes thirty minutes is describing phase one, on a machine that already has the developer tools installed. For someone starting from a normal laptop, phase one is an hour and the full build is a day.

Phase one, step by step

Step one. Install Claude Code. Anthropic's setup page covers it and you need a paid plan.

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.