The idea: one workspace that knows your business
One folder that holds who you are, who you sell to and what you look like, so every marketing task starts from your business instead of from nothing.
The mistake I see people make with AI and marketing is treating every task as a fresh conversation. You explain your company again, paste your brand colours again, describe your audience again, and you get a competent generic answer again.
This is the opposite. You build one folder that knows what your business is, what it looks like and who it sells to, and every task after that starts from there. The setup is ten minutes and it is the whole difference.
The example I used is a real estate company in dubai, but nothing here is specific to property. Swap the website and the reference screenshots and the same five things run.
The five things I actually ran
One prompt each. The whole set cost about three dollars of API credit.
- An on-brand Instagram reel, generated end to end. Not a script, the finished video, in the company's visual style. Which means you can produce as many as you want, for the feed or for ads.
- Competitor reels, scraped and turned into concepts. Ten trending reels in the niche, about seven million views between them, analysed for what they were doing, then new scripts written for us. The deeper version of this is The reels system.
- A trending podcast, turned into five posts. It found a two-hour real estate podcast with over 100,000 views, pulled the ideas out, and wrote five posts with visuals, ready for LinkedIn or Instagram. The prompt was three sentences.
- 100 outreach leads with personalised openers. Name, niche, LinkedIn URL and a first message written for that specific person. The serious version of this is The b2b outreach system.
- Competitor ads, analysed. The one I would start with. More on that below.
Why the ads one is the best first task
Facebook's ad library is public, and it shows how long each ad has been running. An ad that has been live for four years is an ad that is making money. Nobody publishes their conversion rates, but everybody publishes this, and it is the same information wearing a disguise. One prompt gets you every competitor's longest-running ads, what they have in common, and new concepts built on the pattern.
The two APIs that do the heavy lifting
- Apify
- Everything that involves reading other people's data: Instagram, facebook ads, Google trends, YouTube, LinkedIn. Never try to scrape these yourself.
- An image and video model
- I use kie.ai, one key for several generation models, so the workspace can output the finished asset rather than a description of it.
- Claude itself
- Your $20 plan. The thinking, the writing and the code are all included.
That is the entire stack. Two keys in a .env file. The reason it is so small is that Claude Code writes whatever glue is needed between them, per task, rather than you assembling a pipeline in advance.
Setting the project up so it knows who you are
1Make a folder and start Claude Code in it
Make an empty folder named after the business, open it, and run
claude. That is the whole project. There is no scaffolding to download and nothing to configure yet.2Give it the website and let it learn the business
The first prompt is not a task, it is an induction. Give it the URL and ask it to work out what the company sells, to whom, and what the marketing is trying to achieve, then write that down.
here is our website: ellingtonproperties.ae read it and work out what this business is, who it sells to, and how it talks. write what you learn into CLAUDE.md. the point of this project is to get more leads through marketing.
Everything after this is better because of it. Without it you re-brief the model on every single task, and it shows in the output.
3Show it what the brand looks like
This is the step that separates output you can post from output you delete. Make a
reference/folder and drop in four or five screenshots of your own Instagram grid.Not a brand guideline document. Screenshots. The model is looking at the images, and "this is the vibe" communicates faster through four pictures than through a page of adjectives about your brand.
4Connect Apify for anything that involves other people's data
Apify is a marketplace of scrapers: instagram posts, facebook ads, google trends, youtube, linkedin. get an API token from settings and put it in
.env.The mistake that wastes an afternoon
Never ask Claude Code to scrape a site directly. It will try, it will half work, and you will spend an hour debugging something that was never going to hold. These platforms fight scraping. Apify maintains scrapers that keep working. Use them.
5Connect an image and video model
Claude writes the concept and the script. It does not draw. For the visuals I use kie.ai, which is one key for several image and video models, so the workspace can produce the finished asset rather than a description of one.
APIFY_API_KEY=... KIE_AI_API_KEY=...
6Ask for the output as a web page, not a chat reply
The last line of every prompt I write is some version of "output all of it as a single HTML page in our style".
A wall of terminal text is technically the same information and practically useless. A page you can scroll, with the videos embedded and the numbers laid out, is something you will actually read and can send to someone else.
The prompts, in full
These are close to what I actually said, dictated rather than typed. They are long, and that is the point: the context is the work.
Competitor ad analysis:
do a competitor analysis on facebook ads. scrape the top ads from real estate brokers in dubai. focus on legitimate companies, and on ads that have been running for a long time, because that is the signal. look at what the hooks have in common. then generate five new creatives for us. output all of it as one html page, in our style.
Podcast to posts:
find a recent, well performing podcast episode about the dubai real estate market. pull out the five most interesting ideas in it, and write a social post for each one, with a visual. output as an html page in our style.
The shape is always the same: A source, a judgement, an output format. Where to look, what counts as good, and what you want back.
What it costs
- Claude plan
- Pro, $20 a month
- All five tasks above
- About $3 of API credit, in total
- Apify
- $5 of free credit a month, then pay as you go
- Image and video generation
- Pay per generation, cents per image
- What i spend
- A few hundred a month on Apify, but that is daily client work, not trying this out
The cheap part is the thinking. The cost is scraping and generating pixels.
Where it falls over
- Direct scraping. Covered above and worth repeating, because it is the single most common way to lose an afternoon.
- A project with no context. Skip the website and the screenshots and everything it makes is generic. It will still be fluent, which makes it worse, because fluent generic is harder to notice.
- Asking for five things in one prompt. One task per run. The results are better and you can see which part went wrong.
- Trusting the first ad analysis. Ad longevity is a strong signal, not proof. A big company can run a bad ad for years out of inertia.
- Publishing without reading. Everything here is a first draft with good bones. Your name goes on it.
If you have not set Claude Code up yet, do that first on The setup guide. What it all costs in total is on The pricing page.
