Most people's first attempt at an AI writing assistant fails in the same way. They open a chat, type "write me a script about X," get back something competent and anonymous, and conclude the tool is overrated.
The tool is fine. The setup is missing. A writer with no reader, no house style, no examples of your work and no rules is exactly what you asked for, and it will hand you an average of everything ever written on the subject.
What follows is the whole build, in the order it happens, with the actual lines that go into it. It takes about fifteen minutes and it runs on a free account. Once it exists you never build it again.
Before you start: what a free account actually gives you
The build in the video runs on a free Claude account, and I checked that against Anthropic's own documentation rather than assuming it. The support page on Projects states plainly: "Projects are available to all users, including those with free Claude accounts."
Two constraints come with it that the video does not mention and that you should know before you begin.
You can have five projects. The page says "Free users can create a maximum of five projects." That is enough for a Writer and a few experiments, and it is not enough to build a large crew of specialists on the free tier. Plan the five deliberately rather than filling them with tests.
The knowledge base behaves differently. The page states that "Enhanced project knowledge with RAG is only available to users with paid Claude plans (Pro, Max, Team, or Enterprise)," and that when enabled it expands capacity "by up to 10x while maintaining response quality."
That second one matters more than it sounds, and it is why the advice later in this article about keeping files short is not a stylistic preference. On a free account you are working inside the smaller container, so what you put in it has to be dense and deliberate. Everything below is written for that constraint, which means it also works on a paid plan with room to spare.
Alternatives, since this build is not Claude-specific. The same structure works in ChatGPT's custom GPTs or Projects, in Gemini's Gems, and in most assistant tools that let you save standing instructions plus reference files. The names change, the two boxes do not. I built this one in Claude because that is what I run daily, not because the others cannot do it.
Step one: create the project and understand the three parts
In the sidebar, open Projects, then New project in the top right. Name it something obvious. "My Script Writer" is a better name than anything clever, because in three months you want to know what it is at a glance. If it asks what you are working on, one line is enough: writing short video scripts in my voice.
Inside, three things matter and only three.
The three parts, and what each is for
- The chat, in the middle, where you actually ask for work. This is the part everyone already understands.
- Project instructions, which is the job description. Standing rules that apply to every single request, forever.
- Project knowledge, which is where you upload your own examples so it learns how you sound.
Hold one sentence in your head and the rest of the build is obvious: instructions are the job, knowledge is the taste. Almost every badly built agent has these confused, with style examples pasted into the instructions and rules buried in an uploaded file where they get treated as reference material rather than law.
Step two: write the job description
Open the project instructions box. You are writing a job description for someone who will never ask a clarifying question unless you tell them they are allowed to.
Start with who it is. "You are my script writer. You write short spoken video scripts for my channel." Plain, no flourish. The role sets the whole frame and a vague role produces vague output.
Then name the reader, and this is the line most people skip. A writer is nothing without a reader. Ours says it writes for a beginner creator who wants to grow on YouTube but feels overwhelmed and short on time.
Notice how specific that is. Not "creators." Not "people who want to grow." A person with a stated situation, a stated want, and a stated obstacle. The moment it knows who it is talking to, everything it writes gets sharper, because half of what makes writing feel generic is that it was aimed at nobody.
The prompt pack pushes this further, and the instruction is worth reading even if you write your own: name the one person, their level, what they want, and what they already tried that did not work. Something like a complete beginner who wants a first income online, is skeptical because they have been burned by get-rich-quick videos, and has about an hour a day. That is a reader. "Everyone" is not.
Then set the style, specifically. Two lines do most of the work here.
Write speech, not poetry. Full sentences a real person would actually say out loud.
Open every script on the payoff in the first two lines. Never a warm up, and never start with "in this video."
The first one kills the essay register that AI defaults to. The second one kills the throat-clearing that kills retention.
Then give it the skeleton, because a good opening followed by mush is the standard failure mode. Structure every script as a hook, then quick context, then the meat, then the payoff, then a soft call to action. With that in place every draft comes out built rather than assembled.
Then the line that changes the daily experience of using it. Ask for three hook options at the top, then the script. That single instruction is the difference between being handed one guess and being handed choices, and choosing is a much better use of your judgement than rewriting.
Step three: the guardrails, which is the part everyone skips
Underneath what you have written, start a new section and call it something absolute. "Rules. Never break these."
- Never invent a statistic or a fact. And if you use a number, flag it so I can verify it. This is the one that stops confident fabrication, which is the failure that actually costs you credibility.
- Ban the robot tells. No em dashes. No filler words like delve, unlock, game changer, or "in today's world." Naming them individually is what works; asking for "a human tone" does not.
- Keep every line short enough to say in one breath. A physical test beats an abstract one. It is either sayable or it is not.
- If my brief is missing the topic, the angle, the length, or the goal, ask me up to two quick questions before you write.
- You draft, I decide. Never tell me a script is finished, and always give me options rather than one answer.
- If you break any rule, catch it and rewrite before you show me. This is the line that makes it police itself, and you will need it, because it will slip.
Rule four is the most valuable sentence in this entire article, so I want to be direct about why.
Most people let it guess, and then wonder why the script is off. A thin brief produces a confident draft built on four assumptions you never made, and you spend the next twenty minutes correcting a document that was aimed slightly wrong from the first word. Permission to ask two questions converts that into a ten second exchange before any writing happens. It is the single biggest jump in quality you will get from one line, and it costs nothing.
Rule one deserves a note too. Flagging rather than refusing is deliberate. A writer that will not use numbers writes weak scripts. A writer that uses a number and marks it for checking writes strong scripts you can verify. The prompt pack formalises this with a bracketed verify tag so the flags are searchable in a draft.
Step four: teach it your taste
Move to project knowledge and click add content. This is where the thing stops being a well-configured stranger.
Add two or three of your best pieces of work. Not your favourites. The ones that actually landed. If you have transcripts of your best-performing videos, those are the strongest material you own, because they are proof of what works for your audience rather than proof of what you enjoy writing.
Then the file that matters most: one short page describing your voice. Not a vague paragraph about being "conversational but authoritative." A checklist.
What goes on the voice page
- Who you talk to, in the same specific terms as the reader line above.
- Your tone, named plainly rather than aspirationally.
- That you write in the first person, if you do.
- The words and phrases you love to use. Actual examples, lifted from your own work.
- The words you ban. This is usually the most useful list on the page.
- Your energy level. Calm and considered and relentlessly upbeat are different instruments.
- Two lines you would never say. These define the edges, and edges are what a model cannot infer.
That last item is the one people leave out and it does the most work. Positive description tells it where to aim. Two sentences you would never say tell it where the wall is, and a model given no wall will eventually walk through it.
Now the snag, and it is the one from the free-plan section above. Do not upload one giant file. Keep each one tight, roughly one to three pages. It pulls a much stronger signal from a few sharp pages than from one long messy document, and on a free account you are working inside the smaller knowledge container anyway. Short beats long here for two independent reasons, which is unusual and worth trusting.
Step five: run it, and watch the rules fire
The test worth running is a deliberately bad brief. Type something thin, like "write me a script about side hustles," and send it.
If rule four is in place it will not write anything. It will ask you two questions, typically the angle you want and the length and where it is going. That is the moment that saves you from ten bad drafts.
Answer in one line. Something like: the angle is why most people quit in week three, make it sixty seconds, it is for YouTube.
What comes back should show every rule firing at once. Three hook options at the top, because you asked for them. An opening on the payoff. The structure you set. Language aimed at your specific reader. No em dashes and no filler. And any number in the draft flagged for you to verify.
Then do the part that is actually your job. Read the three hooks and keep the strongest. Find the flattest line in the middle and say so: this line is weak, make it punchier and more specific. Cut the ending: drop the last fifteen seconds and end on the idea rather than a summary.
You are not sitting back hoping it lands the whole thing first try. You are steering it line by line, and that takes about three minutes.
The habit that makes it permanently better
Here is the thing that separates a good agent from one that keeps improving, and it takes ten seconds each time.
Whenever you correct it on something, paste that correction back into the instructions. Ours kept opening hooks with a rhetorical question, which I do not like, so a line went in saying not to. That mistake is now gone permanently rather than being re-corrected every week.
Every time you do this the writer gets a little better and stays better. You are not just using it, you are training it, and after a month of small additions the instructions become a document that genuinely encodes how you work. That document is the asset. The chat is just where you spend it.
What you get free, and what comes after
Everything above is the free build, and I want to be exact about the boundary because vagueness here is how people end up annoyed.
Free, and staying free: the full video lesson, and the complete prompt pack. The pack is not a teaser. It runs to five sections: the complete project instructions as one copy-paste block, the voice document template to fill in, a hook and retention framework with five hook types and worked examples, separate structural skeletons for shorts and for long-form, and a list of the AI tells that give a machine-written draft away. It is at ideasrepay.com/academy, it needs no account, and you keep it.
What is not free is the rest of the crew. One writer is not a team. A working studio also needs a Researcher that validates a topic before you spend a day on it, a Voice, an Editor, a Designer for packaging, and a Supervisor whose entire job is catching what is wrong before anything ships. Those seats, the prompts that create them, and the pipeline that makes them hand work to each other cleanly are the paid walkthrough. I am not going to pretend otherwise or dangle them as coming soon.
Build seat one first regardless. If you do not use the Writer daily for two weeks, no amount of additional seats will change anything, and you will have found that out for free.
Come and build the rest of the crew
Start with the free lesson. Our Build Your AI Team course walks the Writer build end to end on a free account, and the prompt pack is yours to keep with no account required. Our free Claude prompting course sharpens the underlying skill this whole build depends on, which is describing a job well enough that someone else can do it.
When one seat is genuinely running and you want the studio around it, that is at ideasrepay.com, and the walkthrough is The One-Person Studio.
It is the build rather than an article about the build. Seventeen steps across five phases, click by click. You get the role prompts for the remaining seats to paste straight in, so the crew exists in an afternoon instead of a month of trial and error. You get the three-way niche fit test that sets your earnings ceiling before you make anything, the full production run from a blank idea to a published video through the four quality gates you never hand over, the packaging and quality control checklists, the disclosure rules in plain language, and the route from your first upload to monetization and diversified revenue. The Crew Build Kit and the Studio Operating Kit come as downloads, and the whole thing follows one operator from month zero to month fourteen so you can see the shape of a year rather than guessing at it.
One payment of $99 opens that walkthrough and every other one we have published, plus every one we publish after it, across all three verticals: online businesses, YouTube and content, and offline work in the real world. Downloads, templates and scripts are included in every blueprint, there are no renewals and no upsells, and you can email us while you build. That is launch pricing, and it moves to $199 after the first 500 members. There is a community building alongside you and I mentor through it personally.
If you want the wider context first, How to Prompt Claude is the underlying method in five moves, and AI YouTube Automation Gets Channels Terminated is the honest picture of the business this crew is built for.
Go and build the Writer. Not later, now, while the steps are in front of you. It is fifteen minutes and you never do it again.