So when TopView launched a connector for Claude, I wanted to see if one chat could do all of it.

New here? Grab a coffee, you’re in the right place. Been here a while? Thanks for sticking around.

Subscribe now

If you know someone who still plans ads across 5 tabs, send this to them.

Share

By the end of this post you’ll have TopView connected to Claude, my prompt, and the 3 things I’d check before spending any credits.

If you’re new to this

TopView is an AI video and image studio. You give it a product, it makes ads.

What’s new is that it now plugs into Claude through MCP. MCP is basically a connector that lets Claude use another app from your chat.

So you don’t open TopView at all. You upload a product photo in Claude, type “Launch a campaign” and Claude calls TopView for you.

What happens after you send the prompt

First it does the research. It looks at TikTok, YouTube, Amazon and Shopee for your type of product and pulls out who buys it, which hooks work and where competitors are weak.

Then it writes the plan. You get an audience insight, a message map, a campaign plan and a content brief. All of it shows up in the chat, so you can push back before anything gets made.

Then it makes the creatives. Product visuals, UGC-style videos and a few versions of each ad, using models like Seedance, Kling and Nano Banana.

And everything lands in one campaign workspace. Each ad sits next to the research behind it, which is pretty handy when a client asks why you picked that hook.

What you need

• A Claude plan where you can add custom connectors. The desktop app and claude.ai both work.

• A paid TopView account. There’s no free plan, and every paid plan I checked includes MCP access. Prices change with promos all the time, so check [topview.ai/pricing](https://www.topview.ai/pricing) on the day you sign up. On 8 October the Plus plan was $36 a month billed yearly, with 100 credits a month.

• And credits. Write down your balance before the first run, you’ll want that number later.

Connect TopView to Claude

The easy way is to let Claude do it. Open a new chat and paste this:

Help me install this MCP locally and provide the login authorization guide: https://mcp.topview.ai/mcp

Claude walks you through it and gives you a login link. Sign in with your TopView account and that’s it.

If you’d rather do it by hand:

1. In Claude, open Settings, then Connectors.

2. Click Add custom connector.

3. Name it `TopView` and paste `https://mcp.topview.ai/mcp` as the URL.

4. Click Connect and sign in with your TopView account.

To check it worked, start a new chat and paste this:

What TopView tools can you use?

It should list the TopView tools. If Claude answers from its own knowledge, the connector isn’t on yet.

My prompt

Start with a clean product photo. Plain background, label readable. Upload it and paste this:

Here’s a photo of [product], a [one-line description]. Launch a campaign for it. Audience: [who, age, where they discover products]. Show me the research, the strategy and every creative you make, and tell me how many credits each step used.

The last sentence is the one I’d never skip. Claude reports the credits after every step, so you see the cost while it’s happening and not at the end.

You can also skip the photo and paste an Amazon link to the product. TopView’s own example does exactly that, and what comes back is a campaign plan pack: the strategy, the content brief and the creative direction, saved in the workspace.

What it costs

Every generation uses TopView credits, and video costs far more than images. On TopView’s pricing page, 50 credits covered about 350 GPT images or only 33 seconds of Seedance video. So start small. Ask for 2 videos and 2 images on your first run, and check your balance before and after.

Before you spend credits

Unlimited doesn’t work through Claude. Some TopView plans have unlimited generations, but only on topview.ai. Through Claude, every generation costs credits.

Check every label. AI still misspells product labels, so zoom into every image before you post it.

And if Claude answers without using TopView, start your message with `Use TopView to` and send it again.

Recap, in the order you’d run it

1. Get a paid TopView plan and write down your credit balance.

2. Add the TopView connector in Claude and sign in.

3. Ask Claude which TopView tools it can use.

4. Upload a clean product photo.

5. Paste the prompt, with the credits sentence.

6. Check every creative and every label before you post.

If this saved you an afternoon, forward it to one person who runs ads for a product. That’s how this newsletter grows.

Share

See you,

Aljaz

Claude wrote every scene. My laptop rendered the MP4 in about a minute.

The tool is free, and it has 57,800 stars on GitHub.

New here? Grab a coffee, you’re in the right place. Been here a while? Thanks for sticking around.

Subscribe now

If you know someone paying for Premiere just to make reels, send this to them. It might save them $23 a month.

Share

By the end of this post, you’ll have HyperFrames installed, my exact first prompt, and the 4 fixes that took my first video from flat to something I’d post.

If you’re new to this

Most video tools work on a timeline. You drag clips, add text, move keyframes.

HyperFrames skips the timeline. Every scene is a small web page: text, images, colours and motion, written in code.

You don’t write the code, Claude does. You describe the video, Claude builds the pages, and HyperFrames records them into an MP4, frame by frame, on your own computer.

It’s made by HeyGen, the AI avatar company. They released it as open source (Apache-2.0), so it’s free to use, including for client work.

What you need

Three things:

• An AI coding tool. I use Claude Code. Cursor, Codex and Gemini CLI work too.

• Node.js 22 or newer.

• FFmpeg, the free tool that writes the video file.

If you’ve never installed Node or FFmpeg, ask Claude Code to do it for you.

Paste this:

Check if Node.js 22+ and FFmpeg are installed on my computer. If not, install them and show me each step.

It checks your machine and installs what’s missing.

Install HyperFrames

In Claude Code, run this:

claude plugin marketplace add heygen-com/hyperframes
claude plugin install hyperframes@hyperframes

It adds HyperFrames as a plugin, with all its skills.

In Cursor, Codex or Gemini CLI, run this instead:

npx skills add heygen-com/hyperframes

Pick the Core Skills group when it asks. That’s all you need.

To check it worked, start a new chat and type `/hyperframes`. It should ask what your video is about.

Your first video

Here’s the prompt I’d start with. Paste this:

```

/hyperframes Make a [30] second vertical video (1080x1920) for Instagram Reels.

Topic: [what the video is about].

Message: [the one thing viewers should remember].

Script: [paste your script, or ask it to write one].

Look: [your brand colours as hex codes] and [your fonts].

Use real screenshots of [links] instead of invented UI.

Cut to a new shot every 1 to 2 seconds. Captions 2 to 3 words at a time, one keyword highlighted.

Show me the plan first, then build it.

It asks a few questions, writes a short brief, then shows you a scene-by-scene plan. You approve it before anything gets built.

That plan step matters. Changing a scene in the plan takes seconds. Changing it after the build takes minutes.

The 4 fixes that made it look real

My first version looked flat. Every shot used the same layout: one colour, one card, one headline. It looked like a template, because it was one.

So I studied creators who get thousands of comments on reels like this. Then I changed four things.

1. Real screenshots, not drawn interfaces.

My first draft drew fake app windows. The creators show the real thing: the actual GitHub page, the actual pricing page, the actual app.

HyperFrames can capture a website for you. Paste this:

Capture [link] and use the real screenshots in the video. Zoom into [the part that matters] and highlight it.

It saves real screenshots of the page, then zooms and highlights them in the scene.

2. A new shot every 1 to 2 seconds.

The reel I studied had 30 cuts in 39 seconds. Mine had 8 cuts in 42 seconds. That was the whole difference in pace.

One shot per spoken phrase is a good rule.

3. Show what the tool makes.

The best b-roll is the tool’s own output. For HyperFrames, that’s videos made with HyperFrames. Their website has 10 of them.

4. Change the layout, not only the content.

Mix four layouts: split screen, full-screen screenshot, full-screen text card, full-screen face. Same video, but it never looks like a slideshow.

Change any scene by chat

Once it’s built, HyperFrames opens a preview in your browser.

You can click any text or clip and edit it there. Or stay in the chat. Paste this:

In scene 3, make the title orange and make the zoom slower. Keep everything else.

It rebuilds only that scene.

Render it

When you’re happy, ask Claude to render it, or run:

npx hyperframes render

On my Mac, a 42-second vertical video rendered in about 1 minute. An 8-second test took 31 seconds.

The file lands in your project folder. No upload, no watermark, no credits.

What it actually costs

HyperFrames: $0.

Your AI coding tool: whatever plan you already pay for. I suggest using Opus 5.5 model for this.

A voiceover is optional. HyperFrames can use a free voice on your computer, or your HeyGen account. With HeyGen, a 26-second voiceover in my cloned voice cost 2 credits. Check your balance before and after your first one.

Recap, in the order you’d run it

1. Install Node.js 22+ and FFmpeg.

2. Install the HyperFrames plugin in Claude Code.

3. Paste the first-video prompt and approve the plan.

4. Ask for real screenshots, a cut every 1 to 2 seconds and mixed layouts.

5. Fix scenes by chat or in the preview.

6. Render with `npx hyperframes render`.

If this saved you an afternoon, forward it to one person on your team who edits videos. That’s how this newsletter grows.

See you soon,

Aljaz

Sites in ChatGPT is in beta since July. Grok Bot came out in August, Claude Projects changed on 17 September and MiniMax H3 is from July. All four were going around my feed as this week’s news, some of them with pretty big numbers, so I stopped trusting the feed and went through every launch on the company’s own page instead.

What was left were 10 real launches. And to be honest, the thing that surprised me most is that half of them are about the same thing, which is your ads.

New here? Grab a coffee, you’re in the right place. Been here a while? Thanks for sticking around.

Subscribe now

If someone on your team runs paid social and didn’t have time to follow the news this week, send them this.

Share

Before we start

For each launch you get what it does, who can use it, what it costs and what I would do with it for a client. Everything is checked against the company’s own launch post.

A lot of it is US first. I’m in Slovenia, so where something is not available in Europe I say so, because it changes what you can actually do this week.

Higgsfield Ads Studio

This one I’d try first.

Higgsfield launched it on 1 October and pushed it hard on Instagram on the 5th. You give it your website once, it pulls in your logo, colours, fonts, tone of voice, audience and selling points, and that becomes your brand kit (you can also fill it in by hand). Then it makes static ads for one product, up to 20 in one run, and every one is built on a different selling angle.

It costs 4 credits per ad, which is about $0.20. So 20 ads, let’s say around $4.

It’s only static images, for video they send you to their Marketing Studio. You can also run it inside ChatGPT with the Higgsfield extension.

I wouldn’t use it for final creative. I would use it to find angles, basically run 20, see which 3 you’d never think of yourself, and then make those properly with the team.

Meta Muse for small business

On 29 September Meta gave Muse, their AI agent, a set of skills for businesses.

It can now read your Instagram professional analytics, your Facebook Page and your Meta ad accounts, and it connects to Shopify, Klaviyo, Stripe, QuickBooks, Notion, Slack, Canva, Figma and a dozen more. So you can ask it what’s working and it drafts next week’s posts and ads from that. Canva also added a Muse connector the same day, which makes designs from your brand assets.

Meta says it’s “free for most of what people need”, with paid plans for more, but there are no prices yet. They also say nothing gets published, sent or spent without your approval, which is good because this one sits right next to your ad budget.

The catch is that it’s US and Canada only, in the Muse app. So for me in Slovenia, not yet.

TikTok

TikTok announced a whole bundle at Advertising Week in New York on 5 October. Some of it is live, some of it is “coming”, so be careful which part you read.

What’s live for US advertisers is the TikTok Ad Network, which shows your ads in nearly 400,000 other apps and runs from the same Ads Manager you already use. Search ads now run on Smart+, so it picks keywords and creative for you, and Smart+ Creative Selection puts your own assets, TikTok’s AI creatives and creator content in one approval flow.

The part I find most interesting is actually the TikTok for Business MCP, which is now a connector in Claude, Perplexity, Replit, Manus and Snowflake. In normal words, you can ask Claude about your TikTok campaigns. TikTok says advertisers using it grew more than 200% from July to September.

Buy Direct (checkout inside TikTok), a Shopping Assistant and a Lead Agent for DMs and lead forms are all announced without a date. I wouldn’t plan anything around those yet.

ChatGPT ads

OpenAI posted an ads update on 5 October and it has two parts.

First is tracking. You can now send conversion data through Hightouch, Tealium and LiveRamp, and see results in AppsFlyer, Triple Whale, Northbeam and others. For incrementality tests there’s Haus, Measured and WorkMagic, and DoubleVerify and IAS are piloting brand safety.

They shared partner numbers too. WeightWatchers paid 15.3% less per acquisition than their paid search benchmark, and 93% of Portland Leather’s visitors from these ads were new to the brand.

Second is a new image ad format that shows up while someone is generating an image in ChatGPT, labelled and kept separate from their picture. Testing starts in the US later this month with a first group of advertisers.

This one I’d watch, not jump in. Ads only show to Free and Go users, and only in the US, UK, Canada, Australia and New Zealand, and there’s no public price or minimum spend yet. But ChatGPT has 1.2 billion people a week, so of course I’m keeping an eye on it.

HyperFrames Studio

Not an ads launch, but probably the one I’ll use the most myself.

HeyGen launched it on 5 October. It’s a free desktop video editor where Claude Code or Codex builds the video and you direct it, so you change things by asking and not by dragging clips around on a timeline. It’s open source, it runs on Mac and Linux (Windows is “coming soon”), it has 57,500 stars on GitHub and it installs as a Claude Code plugin.

I already make my reels in code with a setup I built myself, so this is the one I’m curious to put next to it. For motion graphics and product intros I’d try it first.

GPT-6.1 Sol

OpenAI released it on 29 September. They say it gets close to GPT-6 Astra for about a fifth of the price.

It’s in ChatGPT Work and Codex on Plus and above, not in regular chat, which confused a lot of people this week. It’s also rolling out starting with Pro, so maybe you don’t have it yet.

On the API it’s $2 per million tokens in and $10 out, Astra is $10 and $50. We use Astra mostly for the hard image, video and coding stuff because it’s the expensive one, so if you’re hitting limits in Work mode on long jobs, Sol is what I would try first.

The 4 I’d skip for now

Gemini 4 Argon launched on 30 September and the benchmarks look strong, but right now only a small group of cyber defenders can use it. Paid API customers and Google AI Ultra are next, no date. So bookmark it and come back.

The Figma MCP drama went viral on 30 September, when someone from Figma confirmed that only approved tools can connect to their MCP server. It’s like that since April, and Claude, Claude Code, ChatGPT and Codex are all on the list, so if that’s what you use, nothing changes. Only thing to know is that writing to the canvas is free in the beta and Figma says it will be paid later.

Midjourney’s 2 October update is mostly bug fixes on the alpha site. Collaboration tools are coming next week, I’d wait for those.

Reflection’s Beam is a huge open model from 5 October, made for coding and agents, weights later in October. Great for developers, nothing for your marketing this week.

If you have one hour

I would go in this order. Run Higgsfield Ads Studio on one product, 20 ads for about $4, and keep the 3 angles you didn’t expect. If you’re in the US or Canada, connect Muse to your Instagram and ad account, read-only first, and ask it what’s working. If you advertise on TikTok in the US, open the Ad Network and Smart+ Search in Ads Manager. If you’re on ChatGPT Work, switch your next long job to Sol. And if you use Claude Code, install HyperFrames Studio and try one 15-second product intro.

ChatGPT’s image ads I’d just put in the calendar for later this month.

If someone on your team runs paid social and didn’t have time to follow the news this week, send them this.

Share

See you next week.

Aljaz

Wednesday morning I opened Instagram with my coffee and it was dots, dots, dots. Somebody had an “I tested it” video out less than an hour after the launch, and it was on 85,000 views before I’d finished my cup. OpenAI’s own video had 274,000.

Fun videos. Cute little plush characters. But nobody told me what it costs, what it gets to do inside my accounts, or what happens to my stuff if I change my mind later.

So I read the launch post, the product page and the whole help center, down to the small print.

New here? Grab a coffee, you’re in the right place. Been here a while? Thanks for sticking around.

Subscribe now

Got a colleague asking if this is worth paying for? Send them this.

Share

So what is a dot?

It’s ChatGPT that keeps going after you close the tab.

Normal ChatGPT waits for you. You type, it replies, it sits there. A dot gets a goal from you and keeps working on it on its own time, between your chats. And it can write to you first, when it needs you to decide something or when it spots something you’d want to know about.

Took me a second to get used to that idea. ChatGPT messaging me, not the other way round.

It runs on GPT-6 Astra. It has its own computer in the cloud with its own browser, and it connects to more than 4,000 apps (Google Calendar, Drive, GitHub, Slack and so on).

You give it a name. You pick a character for it, or a pet if you prefer. OpenAI’s launch page has one example I keep coming back to: a tester’s dot saw that he never invoiced a publication, made the invoice, and sent it once he said ok.

Out of the whole launch, that’s the example that stuck with me.

Where you talk to it

Setup happens on a computer, in the ChatGPT desktop app or in your browser. No creating one on the phone.

After that it goes wherever you go: the ChatGPT app, Slack, Microsoft Teams. Same conversation in all three, it doesn’t lose track. You can also call it and just talk.

What it can’t do at launch: call you, or have its own email address. Texting is “coming next”, says OpenAI. You can connect your personal email though, and it’ll use that for your tasks.

It has its own computer

If you remember one thing from this post, make it this one.

Your dot doesn’t live on your laptop. It gets its own computer in the cloud, and you can open that computer whenever you want and see what it’s up to. Like sitting next to an assistant and watching their screen.

Your laptop stays out of it. Access to it is switched off at the start. If you switch it on, the dot can get to your files and use your browser when its own browser gets blocked on some site.

Its profile has three tabs: In progress, Scheduled, Completed. Scheduled is for things it repeats. OpenAI’s own example is checking your calendar every morning and telling you what’s on.

What it costs

Your first dot comes with ChatGPT Pro or Business Premium. You don’t pay anything on top. Pro starts at $100 a month (for me in Slovenia it shows €103).

In a lot of headlines you’ll see dots next to a $500 plan. Not the same thing. Pro now has three usage tiers, and the $500 one is for faster answers and more usage. You get a dot on all three. The cheapest Pro is enough.

Three smaller things from the help center:

Talking to your dot doesn’t count against your ChatGPT limits. The tasks it starts in Codex or ChatGPT Work do.

In the first month after launch you get more room for the heavy jobs.

Later on you’ll be able to buy more dots, or pay for one to work faster. No prices for that yet.

10 jobs I’d give a dot first

I made this list from what OpenAI says a dot can do: jobs on a schedule, your connected apps, and reading around in the background. It follows my day, from the first coffee to the evening.

Morning first.

Calendar at 7:00. What’s on today and what I need to prepare.

Email overnight. It goes through what came in during the night and gives me the five emails I have to answer today.

Trends. I already get a page like this every morning from a system I built with Claude Code. This is the job I’m most curious about, because a dot would do it without me building anything.

During the day.

My to-do list. It does the bits it can do by itself, and asks me about the rest.

A deck that stays current. In OpenAI’s demo, the dot updates a leadership deck when new numbers come in, and marks what’s still open.

Before a call. Every email, note and file I have with that person, on one page.

Content, which is most of my week.

One interview, a week of posts. A new transcript comes in, the dot picks moments for clips, writes show notes and drafts posts. I approve or change them. This one comes from OpenAI’s own creator example, and it’s the closest to my work.

Fridays: competitors. What five competitors posted, launched or changed on their pricing pages that week. One page.

And the boring life stuff.

Invoices I forgot. It finds work I never billed and prepares the invoice.

A trip. Dates and a budget from me, three options back from the dot, I choose.

What I wouldn’t let it do

Anything that works while I sleep can also get it wrong while I sleep. So I read the safety part slower than the features.

You decide which apps it gets. When it looks around in the background without being asked, its tools are read-only, so it can’t send or change anything. With Custom Rules you can let it do something, make it ask first, or block it. Anything touching your accounts or sharing your info gets checked before it happens, and some things, like changing a password, it always leaves to you. It can log in to sites with your saved passwords, and the model never sees them.

My rules would be boring. Ask me before you send, book or pay for anything. First week, read-only. More access after I’ve seen what you did with that.

One thing I didn’t expect. If you disconnect an app, the dot keeps what it already learned from it. You only get rid of that by deleting the dot. So I’d connect apps one at a time, slowly.

On a personal plan you choose if your dot’s chats are used for training. For Business, Enterprise and Edu that’s off by default.

Is anyone else doing this?

Yes. Meta has Muse (travel bookings, email), Google has Gemini Spark. What OpenAI has that the others push less: the dot’s own computer, and the fact that it sits in Slack and Teams, where lots of people spend their whole workday anyway.

If you get one this week

You need Pro or Business Premium. Set it up on your computer and give it a name. Connect two or three apps, not twenty; calendar and email are enough for a start. Write your Custom Rules before the first job. Then give it one job on a schedule, the 7:00 calendar one for example, and after a few days open its computer and look at what it did.

OpenAI is rolling dots out slowly, so yours might take a while to show up.

I’ll write about how three of these jobs go, the parts that break included. Subscribe and you’ll get it.

Subscribe now

Got a colleague asking if this is worth paying for? Send them this.

Share

See you next week.

Aljaz

I didn’t read 2,211 posts. I read one page, and it took about as long as my first coffee.

New here? Grab a coffee, you’re in the right place. Been here a while? Thanks for sticking around.

Subscribe now

If someone on your team starts their day scrolling for ideas, forward this to them.

Share

Today I’ll show you what that page is, where the posts come from and how it picks the few worth your time. At the end there’s a full setup guide if you want your own.

Why I built it

I post about AI every day, and I only post about tools I’ve actually tested. That means I need to know early what people are talking about, so I have time to try it before I write about it.

For months that meant opening Instagram, YouTube, GitHub and Substack every morning and scrolling until something looked interesting. It took close to an hour, and I still missed things that blew up on accounts I don’t check every day.

So on 18 Sep I asked Claude Code to build me something that does the scrolling for me. The first email came the next morning, and today’s was the eleventh.

What it reads

Every hour, at seven minutes past, my laptop runs a check. Each source has its own pace, so it isn’t asking YouTube the same question 24 times a day.

It reads new videos from the YouTube channels I follow, plus a search for AI marketing topics. On GitHub it looks for repos that gained the most stars in a day, and new releases from tools I use, like n8n and Claude Code. It reads the Hacker News front page, Product Hunt’s AI launches and the Technology and Business leaderboards on Substack.

It also reads my own Instagram and Threads, through my own login, plus 22 hashtags. And it watches the release notes pages of tools like OpenAI and fal.ai, so I know the day something changes.

The accounts come from one spreadsheet with 217 rows. Each row has the platform, the account, a tier from 1 to 3 and one line on why I follow it. That last column ended up being the most useful one, because when an account shows up on the page, it reminds me what I’m looking for.

Everything lands in one database file on my laptop. In 11 days it has saved 6,044 posts, and more than half of them (3,294) come from Instagram. Every time the radar sees a post again, it updates the views, likes and comments, so the numbers are never more than an hour old.

How it picks what’s worth reading

Most of those 6,044 posts are normal, so the radar uses two rules to find the ones that aren’t.

The first rule is a fixed bar. A YouTube video needs 100,000 views, a GitHub repo needs 1,000 new stars in a day and an Instagram post needs 500 comments. A Hacker News story needs 300 points, and a Substack post needs 500 likes.

The second rule is the one I trust more. For every account on my list, the radar remembers what a normal post gets over the last 14 days. If a new post gets three times that, it goes near the top, even if the account is small.

This morning an uncover.ai post had 76 comments. That’s nowhere near the 500 bar, but the account usually gets about 6, so it was 12.7 times its normal and it made the page. The first rule would never have caught it.

Older posts count for less, so by day seven a post is worth half. Then the page is written in four parts: launches from the tools I use, posts over the bar, the accounts on my list and new names on the Substack leaderboards. At 7:00 it lands in my inbox.

What was on it this morning

Five things made me stop.

• opusjake, who’s on my list, posted how to make Claude spin up 100 sub-agents so you don’t have to keep re-prompting. It has 7,795 comments in under two days.

• VoiceStudio, an open-source voice app on GitHub, gained 3,221 stars in a day.

• Higgsfield’s own account posted a thank-you to its community, and it has 5,201 comments. It sat at the top of Launches, so I’m going to check what they shipped this week.

• A Growth Memo post on attribution in AI search had 36 likes. That sounds small, but it’s 5.1 times what that newsletter usually gets, and that’s why it sat on the page next to posts with thousands.

• And OpenAI put out a video on GPT-6 Astra in practice, plus an explainer on rogue AI agents. Both were in Launches.

I’m testing the sub-agents setup first, and if it holds up it’ll be next week’s post.

The whole page takes me about 10 minutes. The hour I used to spend scrolling now goes into testing the one thing worth testing.

How I know it’s working

The last part of every email is a short health list. Each source gets one line with “ok”, when it last ran and how many posts it found, so I can see at a glance that nothing is missing.

Instagram and Threads get a bit more care. The radar only reads my Following feed, one page at a time with random pauses, and it stops for the day the moment Instagram says “try again later”.

What it costs

Nothing to run. YouTube’s API, GitHub’s API, RSS and public pages are all free. YouTube gives you 10,000 units a day, and the radar uses about 1,300 of them.

I did look at a paid scraping service and decided against it. I’d rather have fewer sources I can check myself than pay for numbers I can’t.

Building it used my Claude subscription. The first version took one day, and since then I’ve added sources and small fixes as I needed them.

Build your own

I wrote the whole setup down as a step-by-step guide. It’s 7 prompts you paste into Claude Code in order, with a check after each one, so you know a step works before you start the next.

It covers the list, the database, the free sources, both rules, the 7:00 email and the schedule. There’s also an optional part for Instagram and Threads, with the rules I follow to keep my account safe.

The guide is free for subscribers. Subscribe below and you’ll find the link in your welcome email.

Subscribe now

Already subscribed? Reply to any Coffee First email with RADAR if you didn’t get it and I’ll send it to you.

If you only remember one thing

Start with the sources you already check by hand, and write one line on why you follow each account. Add the two rules, get one page at the same time every morning, and put a health line at the bottom so you know everything ran.

If someone on your team starts every day scrolling for ideas, forward this to them.

See you next week.

Aljaz

The same week, LADbible reported customers publicly calling out local businesses for their ChatGPT adverts.

So I made six flyers in ChatGPT and looked at what gives them away.

New here? Grab a coffee, you're in the right place. Been here a while? Thanks for sticking around.

Subscribe now

Know someone who makes the flyers at your company? Forward this to them.

Share

By the end you’ll have one prompt, the five signs to check for, and a 30-second check before you post.

How I made the six flyers

I picked six jobs a small business does every month: a summer sale, a free webinar, a café menu, a job ad, a jazz night in the park and a granola launch.

For each one, I wrote the exact lines of text and asked ChatGPT for the kind of flyer you see on every Facebook event page.

The good news first. Every word I gave it was spelled right: every date, time, price and the café’s wifi password.

What went wrong

ChatGPT added text I never wrote to all six posters.

Each one got a handwritten slogan in a corner. The boutique got “More Fashion Brighter Days”. The café got “Good Coffee Brighter Days”. The jazz night got “Good Music Brighter Together”. Six different businesses, and the same word on every poster.

It also wrote on everything it drew. The café’s chalkboard says “Good food, good coffee, good people”. A mug on the job ad says “Good ideas, better tomorrow”. The webinar poster got a row of promises at the bottom: grow your audience, generate more leads.

The granola poster is where it becomes a real problem. ChatGPT added three round badges: “High fibre”, “Plant power” and “Perfect for any time”. I never gave it any of those.

In the EU, “high fibre” is a regulated nutrition claim, and a product has to meet a set amount of fibre before the label can say it.

A designer would never add a claim the client didn’t give them. ChatGPT fills empty space with the kind of copy it has seen on thousands of templates.

The five signs people spot

Once you’ve seen these, you’ll see them on every noticeboard.

A slogan nobody wrote. A small handwritten line in a corner, usually with a heart.

A brush stroke behind every headline. Every important line sits on a painted streak of colour, so none of them stands out.

An icon for every detail. A calendar next to the date, a clock next to the time, a pin next to the address, each in its own circle.

Everything at the same volume. Five fonts, four colours and nothing that clearly comes first.

Perfect people in perfect light. Smiling teams around a laptop, a model with shopping bags, a speaker with his chin on his hands.

The prompt

The fix has two parts. The first part protects your text. The second part gives ChatGPT something to design around.

Upload your poster, then paste this with your own lines in the TEXT section:

The TEXT section does most of the work. Quotes tell ChatGPT these exact characters matter, and “nothing else” gives it permission to leave space empty. Without that line it keeps adding copy.

The IDEA section is what makes it look designed. “Make it cleaner” gets you a tidier version of the same template. One sentence about a picture gets you a poster.

What an idea looks like

Here’s the one sentence I gave ChatGPT for each flyer, and what came back.

The sale

The discount is the picture: “50%” set enormous in a serif, cropped by the edge, with one linen dress on a hanger casting a hard shadow across it.

Every word right and no slogans.

The webinar

The long title is the design: set huge, flush left over six lines, with “2026” as the only word in colour.

Every word right and nothing added.

The café menu

A printed menu card for a small European café, with one lime-tree leaf drawn in a single line. Lipa means lime tree.

All 10 lines right and nothing added. It grouped the drinks and the food on its own, without inventing headings.

The job ad.

“Your seat is free”: one empty chair pulled out from a clean desk, shot from above, with the job title as the hero.

No slogans and no fake mug, but it cut two lines: “Young dynamic team!!!” and “Good salary”. It also kept small icons next to the three facts.

The jazz night

A 1950s record sleeve: a duotone close-up of a saxophone and tall condensed capitals over blocks of colour.

Every date, time and place right. It dropped “food • drinks • bring your friends!!!” and put “Riverside Park” at the top as a small logo.

The granola launch

The product is the hero: one pouch on a plain oat-coloured backdrop, with a single column of oats falling behind it.

Every word right and the three invented claims are gone. The exclamation marks changed, and the price still sits in a round badge.

None of the six redesigns added a slogan, and every date, time and price came out right.

About the “codes”

You’ve probably seen posts promising 200 ChatGPT codes for posters: /swiss, /editorial, /retro70s.

They aren’t commands. ChatGPT has no list of codes, and the slash does nothing. They work because ChatGPT knows what a Swiss poster or an editorial layout looks like, so the word acts as a style hint.

A code alone gives ChatGPT a look without any rules. I wrote each style as one full line instead, so it knows what the look means. Six of them:

Paste the whole line into the STYLE section.

The full set of 24 for posters and, plus a ready IDEA for six common flyers, is in the Secret stack along with other free resources that are available only for my susbcribers. If you’d like to get access, please subscribe below and you’ll receive a link to the Secret Stack.

Subscribe now

Four follow-ups for the same chat

When one word comes out wrong, don’t start over:

When a line goes missing:

When extra text creeps back in:

When you need it for Stories:

The 30-second check before you post

Read the poster out loud against your TEXT list. Check every date and time, every price with its currency sign and comma, and every name and address. Then look for any line that isn’t on your list, including text on signs, mugs and packaging. If you sell food, drinks or cosmetics, check twice for claims you didn’t write.

Recap

1. Write your flyer’s text as a list, one line per quote.

2. Write one sentence about the picture the poster is built around.

3. Upload your old flyer to ChatGPT and paste the prompt with your TEXT and IDEA.

4. Pick a style line or leave STYLE empty.

5. Do the 30-second check, then fix single words in the same chat.

If this saves someone at your company an evening in Canva, forward it to them.

See you next week.

Aljaz

I’ve seen people copying the whole Minecraft game with it in just a couple of hours. GTA 6 and similar stuff. Mostly pretty fun stuff, but nothing useful.

So I started thinking, what can I use Astra for in marketing? Which clients, what can it help me with?

It’s the biggest and the best model so far, so I focused on the video. The most time we spend on with my team is the video. We’re struggling the most with the video.

Can we do a promo video with it? A trailer? A hero video for a SaaS client? The whole social media ad? Can we do a TV ad with it?

So I experimented with it a bit more. That’s why I’m posting it now and not three weeks ago when it came out.

New here? Grab a coffee, you're in the right place. Been here a while? Thanks for sticking around.

Subscribe now

First, what Astra is

Astra is the biggest and newest model of OpenAI. It’s the most capable model, and of course the expensive one. So we’re not using it for everything. We’re using it mostly for hard image tasks, video tasks, or really hard coding tasks.

We use it with Higgsfield a lot, because in my opinion it’s the most capable tool out there, with all the AI models in it. We use it for generating images with GPT Image 2.5, for videos with Seedance, and so on.

Basically, you’re using Astra to prompt Higgsfield better, so the output is better than if you just do it in Higgsfield.

Higgsfield connects to ChatGPT as a plugin, and the videos use your Higgsfield credits, not your ChatGPT plan.

The setup

I set up the project the same way I do the rest of them:

1. In ChatGPT, create a new project (not a chat) and name it after the brand.

2. Upload the brand files: logo, colours, three to five product photos from different angles, and three to five images of the look you want.

3. Open Plugins and connect Higgsfield.

4. Start a chat inside the project and switch the model to GPT-6 Astra.

The five videos

These are the five videos I generated, but that’s not even close to what the model is capable of.

1. UGC ads

To be honest, I played with it a little. I didn’t spend a lot of time doing this.

The first time it came back pretty bad, because I did not use much prompting. I just wanted to see what it could do on its own. Then I started prompting it a bit to get the result I wanted: good enough for me to post it on social media.

If I prompted it better, focused on the influencer better, and developed the whole new persona first, then probably the outcome would be even better.

Then image input was just a can of Poppi on a white background:

Then I used this prompt:

First, it gave me 3 different videos, then put them together in a UGC video:

2. Launch map

I was struggling here to find the business I could test this on. Why the Apple Store? I didn’t want to do something related to real estate, but that’s probably where you can use it most. If you’re selling a listing, you can show, with a video, pretty quickly and in a good way, how far it is from school, from the hospital, from grocery stores, and so on.

I chose the Apple Store because it’s maybe what marketing brands are also doing.

I uploaded a photo of a storefront and used this prompt:

Here’s the final result:

Type the real times in yourself. If you leave them out, the model makes up times that look right.

3. Walkthrough

This one was focused on real estate listings.

The problem I had was finding the listing that has all the pictures: the entrance, the living room, the kitchen, the terrace, bedroom, bathroom, and so on. I spent, let’s say, 10 to 15 minutes looking for the right listing.

Then I put all the rooms I wanted in the video into the model and played around with a prompt:

Here’s the result:

If I spend more time, the video would probably be even better.

4. Product build

This one is pretty good, and that’s why I love the Higgsfield plus Astra combination. When I take a picture of a product, a Jordan sneaker in this case, the model is so capable that it does not change the product at all.

Of course, I use my prompt at the end of each session when I’m doing images and videos:

Here’s the full prompt you can reuse:

You should get a similar result:

5. Explainer

The hair transplant took me the most, because I didn’t know what process I could show.

I would show this to a clinic. It’s especially useful for new patients who don’t know what the procedure looks like, what’s happening during it, or how long it takes. It explains, at a typical patient level, what’s happening and what they can expect. Plus, the doctor can be there, explaining the video himself.

Reuse this prompt on your case:

My input was only this photo, which was really not the best. If I had used a better process explanation photo, the result would be even better.

And this was the final result:

Where it went wrong

Mostly with the last example, the hair transplant.

You really need to put in the right picture, the right process, the right illustration. Same as with all the other tasks: if you don’t put in good input, the output’s gonna be bad.

What it costs

Not sure, to be honest. I ran out of credits in a week because I was testing so much, so it wouldn’t be right to say how many credits I actually spent.

It was definitely worth it. Each time a new model comes out, I try to test it in as many ways as possible: on everyday tasks, on clients’ tasks, and so on. So I can see what it’s capable of, and what I can do better, quicker, and of course cheaper with it.

That’s why I always run out of credits when a new model comes out.

Which one to try first

It depends on the industry you’re in and what you’re trying to say to the clients.

For most brands, the UGC ads will be most useful. For real estate agencies, the walkthrough or launch map. For brands in general, the Jordan build. The hair transplant for more complex services or products.

Even IKEA could put the paper instructions in a video.

The full prompts

The full prompts are in the Secret Stack. It’s free when you subscribe, and the link is in your welcome email.

Subscribe now

See you next week,

Aljaz

You know the drill: “Comment CODES and I’ll send you 200 more.”

Every time I scrolled, another one. What amazed me was that one of them had 30,000+ comments.

So I went looking for the codes myself.

If you’re new here, welcome! If you’ve been reading for a while, thank you.

Subscribe now

if someone on your team makes the product shots or the ads, forward this to them. It’ll save them an afternoon.

I found six public lists. Between them, they promise over 1,100 codes.

Most of them overlap. A fair few are named after film studios and video games (which you don’t want anywhere near an ad).

Once I took those out, around 600 were left. So I tested them all.

Honestly? You don’t need 600 of them. You need about 20, and one prompt with it.

By the end of this, you’ll have both. Plus a template you can reuse for pretty much anything.

First, what these codes actually are

A code is one word with a slash in front. Something like /producthero. Or /flatlay. Or /staticad.

You type it into ChatGPT, usually with a photo and it tells ChatGPT what kind of image you want.

Here’s the part I didn’t expect: they aren’t a ChatGPT feature.

There’s no menu. There’s no list from OpenAI. The slash does nothing (really).

ChatGPT reads the word as a hint. It knows what a flat lay looks like, so it makes one.

That changed how I think about them.

If the slash does nothing, plain words work too. You can stack them. You can even invent your own. And that’s where the fun beggins.

So the codes are a shortcut. One word instead of a sentence.

And a shortcut is only useful if you know where it’s taking you.

[IMAGE: the same product photo run with /packshot, /flatlay and /producthero, side by side]

The line that makes every code work

Most people type the code and nothing else.

You get a nice image. Of something that isn’t quite your product.

ChatGPT redraws what you upload. It moves labels. It smooths faces. It rewrites headlines (apparently it knows better).

To fix this, you need to use one line. Add it to every request with a photo or text in it:

It tells ChatGPT what it’s not allowed to touch.

Without it, a code is a bit of a coin flip. With it, you get your product in a new picture.

The five-part request

Every request I’d write has the same five parts.

Only the first two are required. I’d still use all five.

Paste this:

I’ll explain:

1. Main code: Just one. It decides what you get.

2. Add-ons: Up to three, for light, angle, or framing. I’d stop at three. After that, they start fighting each other.

3. Details: This is optional, but if you want to keep it aligned with your brand, give it a little more context.

4. Size: 1:1 for listings, 4:5 for Instagram, 9:16 for Stories, 16:9 for YouTube. If you don’t say it, ChatGPT guesses.

5. Keep: The line above. Don’t skip it.

ChatGPT Codes

1. Product shots

These four cover most of what a shop or a launch needs.

• /packshot: Product on plain white, ready for a listing.

• /producthero: The product as the star: dramatic light, glossy reflection.

• /lifestyleshot: The product being used in a real setting.

• /flatlay: Shot from directly above, with props around it.

Let’s test one. Paste this:

You get a hero shot in warm evening light, camera low, looking up at the bottle.

2. Ads

Four codes, four kinds of ad.

• /staticad: A single-image ad with a headline and a button.

• /storyad: A 9:16 ad for Stories.

• /adset: Several versions of one ad with different hooks, to test.

• /billboard: Your ad on a real roadside billboard.

Test time:

Put the headline in quotation marks. I’d do it every time. It keeps the words from drifting.

Then run /adset on the same product. Five ads, five angles. I’d put a small budget behind each and keep the winner.

And /billboard is the one for clients. They get a billboard mockup faster than any deck.

3. Headshots

• /corporateheadshot: Suit, neutral backdrop, soft light.

• /founderportrait: A confident portrait for an About page or a press kit.

Paste this, with a selfie uploaded:

Without that second line, you get someone who looks a bit like you (and a bit like everyone else on LinkedIn).

4. Explaining your product

People buy what they understand. These three show how a product works.

• /explodedview: Every part pulled apart in 3D.

• /cutaway: The product sliced open to show the inside.

• /insidethebox: What the customer sees when they open the package.

Paste this:

Every part, pulled apart and labelled. It works as a carousel slide or on a product page.

5. Brand

Three codes, three messages, one chat.

• /logoconcepts: Several logo directions for a brand name.

• /brandboard: One sheet with logo, colours, fonts and imagery.

• /brandmerch: Your logo on tees, mugs and totes.

Send these one after the other, in the same chat:

Same chat, so it remembers what you picked. Logo, then the brand sheet, then the merch.

6. The add-ons

These go after any code. They’re the quickest way to make a result look less generic.

• /goldenhour: Warm, low evening sun.

• /softlight: Even, flattering light, no hard shadows.

• /lowangle: Camera low, looking up. Makes things look bigger.

• /negativespace: Lots of empty space, room for your text.

Pick one for light and one for angle. That’s usually enough.

What to do when it goes wrong

Don’t start over. Fix it in the same chat.

1. The text came out wrong. Send this:

2. It looks generic. Add one real detail: your brand colours, where your customer uses it, the time of day.

3. The code seems ignored. Write it out: “a product hero shot” instead of /producthero. Same thing, just clearer to ChatGPT.

4. A video code does nothing. Video codes need a video tool connected to ChatGPT, like the Higgsfield plugin. Without one, you get a still image.

Before you post anything

• Read every word, price, date, and name out loud against your original.

• Zoom in on hands, logos and small text. That’s where the mistakes hide.

• Use your own photos. Never someone else’s face without asking.

The whole thing, in the order you’d use it

1. Pick one main code for the job.

2. Add up to three add-ons.

3. Say the size.

4. Add the Keep line.

5. Check every word before you post.

That’s the system. The code is just the first line of it.

The full list

I put all 600+ codes in the Secret Stack, sorted for marketers.

The 110 useful ones have a line on what each does. The rest are grouped, so you can find a style in seconds. I’ll keep adding to it as I find codes that work.

It’s free. Subscribe, and the link is in your welcome email.

Subscribe now

If this saved you an afternoon, forward it to one colleague who makes your ads.

See you next week.

- Aljaz

I write Coffee First, a newsletter about AI that marketers can actually use.

Every day I find the tools, codes, prompts, and build systems and guides worth your time. Then I show you exactly how to use them, in a few minutes, before your coffee gets cold.

The Secret Stack is where all of it lives. Subscribe and you’ll get access to all of them. It’s free now and will later be available only to paid subscribers.

Subscribe now

Why it exists

Every time I post, I say “comment a word and I’ll DM you the link.”

You comment and I send the link. A month later, you need it again and can’t find the DM.

So I put everything in one place. One link, and it never goes out of date.

What you get

→ Every resource from my posts, with the exact steps to use it.

→ New additions every day. You don’t need to do anything to get them.

→ Free. Subscribing is free too.

How to get it

The link isn’t on this page. It’s in your inbox.

1. Subscribe below. It’s free.

Subscribe now

2. Look for an email from Aljaz Peklaj with this subject: Your Secret Stack is here

3. Click the link and bookmark it.

Can’t see it? Check Promotions and Spam. Gmail likes to hide it there (apparently),

Came from a DM?

If you commented a word on one of my posts, the thing I sent you lives in the Secret Stack too.

Same link. Everything else sits right next to it.

Still nothing?

Reply to any Coffee First email, or leave a comment under this post.

I’ll send it to you myself.

Two things to do once you’re in

→ Bookmark the Secret Stack. You’ll be back.

→ Add me to your contacts. Then every new email lands in Primary, not Promotions. It takes 20 seconds.

My emails come from [SENDER ADDRESS]. Here’s how to add it:

Gmail on a computer

1. Open my welcome email.

2. Hover over my name at the top.

3. Click the person icon with the + (Add to contacts).

Gmail on your phone

1. Open my welcome email.

2. Tap my picture next to my name.

3. Tap Add to contacts.

Found it in Promotions? Drag it to the Primary tab. Gmail asks if you want to do this for future messages. Click Yes. On your phone: tap the three dots, then Move to, then Primary.

Apple Mail

1. Open my email and tap my name.

2. Tap Create New Contact (or Add to VIPs).

Outlook

1. Open my email and click my name.

2. Click Add to contacts.

Or just reply to any email with the word “coffee”. A reply tells your inbox you want to hear from me.

If a colleague keeps asking which AI tools you use, send them this page. It’ll save you both 10 minutes.

Share

I wanted to find repos that do a real marketing job: audit a landing page, rewrite an email, read a competitor’s pricing, help with SEO.

I’ll be honest. I didn’t find only seven, GitHub is full of them. So I had a little trouble choosing which one I should really publish.

I chose them based on the following criteria:

• number of total stars + how many stars they are still receiving daily. This tells me that repo is popular and people are actively downloading and using it.

• when it was published. I removed all the repos that are more than 1.5 years old. Something that might have worked 2 years ago does not necessarily work now.

• The quality of the documentation and repo itself. I went through documention of each one and tested the shortlisted ones.

At the end I chose 7 of them that might help you and your marketing team. That doesn’t mean the others are not good, so I’ll probably be doing the next round soon.

By the end of this, you’ll have all seven installed, and you’ll have run one prompt in each.

If you’re new here, welcome! Subscribe if you want weekly tips on how to use AI for marketing. If you’ve been reading a while, thank you.

Subscribe now

First, what a “repo” actually is

A repo is a folder on GitHub with code in it. Each one does its job, and you can import them easily into your Claude Code or ChatGPT Codex or others.

The ones below are a specific kind: agent skills.

A skill is a folder of instructions that teaches Claude Code (or ChatGPT’s Codex, or any agent that reads skills) how to do a job. You install it once, then you call it with a slash command, like /cro.

Two of the seven aren’t just skills but complete tools you run on your machine (deer-flow and OpenMontage). The installation takes longer than the others, but they also get more done.

To get started, you’ll need to install Claude Code first. If you don’t have it, run this:

Then type `claude` in your terminal.

You can also use Visual Studio Code if you don’t like the terminal.

If you use Claude desktop app, just switch from Chat to Code.

All done? Let’s go.

1. marketingskills

Corey Haines literally wrote down what a good growth marketer does, as skills.

CRO audits, copywriting, SEO, analytics, launch plans, pricing pages. 50,895 stars, 7,709 forks, MIT.

Imagine having another virtual team of marketing on your hand to help you with your product launch, campaign, market research.

To install it, run this:

Let’s say, for example, you run an eCommerce store, and you’d like to check how it’s performing and where you’re loosing money.

Paste this prompt and adjust it to your actuall website:

Run /cro on my landing page at [URL] and list the five biggest conversion leaks, ranked by how much money each one is costing me.

It will read the page and give you five leaks with actual fixes for each. The first time I saw the output it read like a pro CRO consultant would write it. Except, it didn’t send me an invoice with it.

Pro tip: the skill works better if you give it more context; for example, tell it who the page is for. Add one line to the prompt:

The buyer is a [role] at a [company size] company.

2. humanizer

Humanizer is a skill that humanizes your writing and removes the signs of AI from text. You know what I’m talking about: the em dashes, the “it’s not X, it’s Y”, the rule-of-three adjectives, the closing line that restates the opening. Everything that’s cringe about AI.

To install it, run this:

The --global add-on means it works in every project, not just the one you’re in. You want that for this one so you can use it across the whole Claude.

Let’s use it on one example. You were just about to publish a new post on your social media, and you used AI to help you write the caption. You want it to make it sound more human.

Paste this prompt:

/humanizer 

Rewrite this caption so it reads like I typed it on my phone. Keep every fact. Don’t add anything. 

[paste your caption]

It will return the same caption, but shorter and less AI slops. Try it on captions, emails, website copy, and others text before you publish it.

(But there’s a twist: It won’t make bad copy good. It makes good copy sound like a person.)

3. last30days

Last30days is an agent skill that researches any topic across Reddit, X, YouTube, Hacker News, and the web for the last 30 days. Then scores what it finds by upvotes and likes instead of by what an editor thinks.

To install it, run this:

Let’s say I want to scale you want to scale your B2B SaaS and you’re wondering what’s currently working on different platforms, what competitors are posting and what’s getting traction.

Paste this prompt:

/last30days AI video ads for B2B SaaS

Research what has worked on Reddit, X, and YouTube during the last 30 days.

Focus specifically on AI-generated or AI-assisted video ads promoting B2B SaaS products. Include product demos, customer stories, founder-led ads, problem–solution videos, feature launches, comparison ads, educational ads, retargeting creatives, and short-form demand-generation videos.

Exclude consumer products, ecommerce ads, entertainment content, generic AI-video tutorials, and posts that do not relate clearly to B2B SaaS marketing.

Find the five most relevant high-engagement posts or videos from each platform. For every example, include:

- Publication date
- Creator or company
- Direct link
- Visible engagement metrics
- Intended B2B audience
- Product category
- Funnel stage: awareness, consideration, conversion, or retargeting
- Opening hook
- Main pain point
- Core promise
- Creative format
- Approximate length
- Call to action
- AI tools or production methods mentioned
- Why people engaged, based on the replies or comments
- What a B2B SaaS marketer could adapt without copying the creative

Rank examples within each platform. Do not compare raw engagement figures across Reddit, X, and YouTube as though they were equivalent.

Separate the findings into:

1. Actual B2B SaaS video ads
2. Organic B2B SaaS videos that could be adapted into ads
3. Discussions or case studies about producing AI video ads

Then synthesize:
1. Five hooks appearing repeatedly
2. Five creative formats currently working
3. The pain points receiving the strongest audience response
4. Three approaches that appear oversaturated
5. Three underused creative opportunities
6. Differences between founder-led, customer-led, product-led, and faceless creatives
7. Patterns in video length, pacing, captions, demonstrations, proof, and calls to action
8. Which ideas appear best suited to LinkedIn, YouTube, and paid social
9. Ten AI video-ad concepts a B2B SaaS company could test this week
10. One recommended concept with:
- Target audience
- Funnel stage
- Hook
- 30–45 second script
- Shot-by-shot outline
- On-screen text
- Proof element
- Call to action
- Three alternative opening hooks

Prioritize examples with observable evidence. Clearly distinguish facts, creator claims, audience reactions, and your own interpretation.

Do not invent engagement figures, campaign performance, conversion rates, ROAS, or information hidden behind inaccessible pages. If an example has strong public engagement but no demonstrated advertising performance, state that clearly.

Afterward, you’ll get a sourced list with links. Think of it like the research intern that never sleeps, and it’s the one I run every time we launch a new client or product.

4. firecrawl

Give Firecrawl a website, and it turns the pages into clean text your AI agent can actually use. That could be one pricing page, 200 customer reviews, or an entire blog. Useful when you’re launching a new product or website and you want to combine the best elements of your competitors’ websites.

To install it, you’ll first need to get a free API key from Firecrawl. The free tier is enough to try this example.

Then run this prompt:

npx skills add firecrawl/skills

Then paste this:

Read every page under [competitor.com]/pricing. Create a table showing what they charge for, what each tier includes, and what they don’t mention.

Instead of spending ten minutes jumping between tabs, you get one clean comparison table. Pay attention to what the competitor leaves unsaid. That’s often where the most useful information is.

5. DESIGN.md

Claude can build a decent landing page. Getting it to use the same colours, fonts, and spacing twice is the harder part.

Awesome Design MD is a collection of DESIGN.md files from real brands. Each file spells out how the brand should look and sound, so Claude has a set of rules to work from.

There’s nothing to install. Find a DESIGN.md you like, copy it into your project folder, then change the colours and fonts to match your brand.

You can build anything with it, but let’s say you want to build a new landing page for your product. Paste this prompt:

Read DESIGN.md before you start.

Build a landing page for [product] using the colours, fonts, spacing, components and writing style defined in that file.

Include:
- A hero with one main call to action
- The problem the product solves
- Three main benefits
- A section showing how the product works
- Social proof
- A final call to action

Make the page work well on desktop and mobile. Use clear HTML and make buttons, links and text easy to use and read.

Do not add colours, fonts, effects or design patterns that aren’t in DESIGN.md.

Put everything in one index.html file with the CSS and JavaScript included. Don’t use a framework.

The result should already feel like it belongs to your brand.

Before I started doing this, every page Claude made for me used a slightly different shade of blue or violet. You know. Typical vibe-coded-looking stuff.

You can also write your own DESIGN.md and reuse it for every project. Mine is basically four hex codes, two fonts and a few spacing rules. That’s been enough to keep everything consistent.

#6. deer-flow

DeerFlow is a super-agent from ByteDance that researches, cross-checks and writes. It’s great for research jobs that would normally leave you with 20 open tabs and messy notes.

Give it a question, and it breaks the work into smaller searches, checks what it finds, and pulls everything into one report with sources.

This is a full application, so be patient as the setup takes around 15 minutes.

To install it, run this:

Then follow the setup instructions in the README. You’ll need an API key for a language model and another for search. Tavily has a free tier, and the README explains where both keys go.

If you are just launching a newsletter and you want to scan others in your niche to see what they are doing, paste this:

Research the 10 largest AI marketing newsletters.

For each newsletter, find:
- Its audience size and the date that figure was reported
- Free and paid subscription options
- Other ways it makes money
- Publishing frequency
- Main content format
- The channels or tactics that drove its growth
- Sources for every number and factual claim

If two sources disagree, show both figures and explain which one appears more reliable. Do not estimate missing numbers.

Finish with a comparison table and five lessons a new AI marketing newsletter could apply.

DeerFlow plans the research, sends different parts of it to separate agents, compares the sources, and writes the final report.

It’s useful before a launch, when you need to understand the market properly but don’t want to lose half a day collecting numbers by hand.

7. OpenMontage

Think of OpenMontage as agentic video production. It takes a script and turns it into a complete video, including the shots, voice-over, music, captions, and final edit.

It comes with 12 production workflows and more than 700 skills.

Again, this is a full application, so allow around 15 minutes for the initial setup.

To install it run:

Follow the README to finish the setup and add your API keys. OpenMontage uses paid image and video models, so every generated clip has a cost. A short ad can cost less than a dollar with cheaper models, but the total depends on which models you choose and how many times you regenerate the clips.

If you want to create a video ad for your product, paste this:

Create a 20-second vertical video ad for [product].

Audience: [target customer]
Problem: [problem they have]
Main benefit: [most important product benefit]
Proof: [result, testimonial or product evidence]
Call to action: [desired next step]

Requirements:
- 9:16 format at 1080 × 1920
- Exactly three shots
- Show the product within the first three seconds
- Use the colours, fonts and visual rules from DESIGN.md
- Write a short voice-over that sounds natural when spoken
- Add burned-in captions that match the voice-over
- Keep each caption to two lines or fewer
- Use subtle background music that does not compete with the voice
- End with the product name and call to action
- Do not invent customer results, product features or statistics

Before generating the video, show me:
1. The three-shot plan
2. The voice-over
3. The on-screen text
4. The estimated duration of each shot

Then create the first cut.

The first version probably won’t be the one you publish, but it gives you something real to work with. You can see which shots work, where the script drags, and what needs another pass.

That alone can save hours compared with starting from an empty timeline.

Step-by-step guide on how to run this:

1. Install Claude Code.

2. Add Marketing Skills, Humanizer and Last 30 Days with npx skills add.

3. Create a Firecrawl account, copy your API key and add the Firecrawl skill.

4. Put one DESIGN.md in your project folder and change the colours and fonts to match your brand.

5. Run /cro on your own landing page. This is the quickest way to see whether the setup is actually useful.

6. Set up deer-flow and OpenMontage on a day you’ll have more time to test it.

You don’t need to install everything at once. Start with the landing-page audit, use what it gives you, and add the other tools when you have a job for them.

If these repos were useful, send this to someone who might use it too.

Share

Hi, I’m Aljaz and I share the AI stuff that you can actually use before your coffee gets cold.

Reply to this email and tell me which one you ran first. I read every reply.

- Aljaz