Hey all, happy Friday,

We drove out to Long Island this weekend to visit my dad and stepmom. Took James to a playground on Saturday (the same one where I learned to ride a bike, which feels like the kind of detail I include to make you think I had an idyllic childhood. I still have a scar from falling on one of those fateful training-wheelless days). Naturally, I considered putting my 13-month-old on a bicycle. Sadly, there was no bike to be had. He was, however, very interested in a spider net climbing toy that he immediately got himself tangled in. For a minute there, I thought we were going to have to cut him out or leave him behind. We got him out. But it was a tough decision.

Moreover, James has developed an absurd fascination with the outdoors. All he wants to do is stare out the window. Outside the city, this mostly meant watching Linus chase squirrels (who, for the record, are significantly less battle-hardened than the Prospect Park squirrels. They are not pulling knives. They are not fighting back. They just... waddle away).

(I have also been to Barry's twice this week and did legs today, so I will not be walking tomorrow. This is relevant to absolutely nothing, but I need someone to know I gymed.)

I also bought a Mac Studio on Tuesday and set up an agent in our Slack named Skippy, but that is a story for another newsletter.

Anywhoooo…

The tool is Claude Cowork.

If you have been following the last three issues (one, two, three), you built the system: the formats, the brand cards, the competitive analysis.

Cowork is what actually executes it all. You connect it to your Claude Project, point it at your browser, and tell it what to do. It makes a plan. It does the work. You watch (or go eat lunch).

I am going to walk you through the exact workflow. By the end, you will have the same system.

Three things to have ready. If you followed Issues 2 and 3, you already have two of them. If you are brand new here, Cowork still works without a Claude Project. You just give it your brand docs directly. The project makes it better, not mandatory.

1. Your Claude Project. This is the homebase. If you followed Issue 2, you already have a project with your brand research brief, Brand Spec Card, Visual Style Card, and the 15 format prompts. If you followed Issue 3, it also has the Format Classification Guide. That project is where everything lives. Every conversation inside it inherits all of that context automatically. If you have not set up a project yet, go do Issue 2 first. This newsletter is the layer on top.

Drop in any additional files that give Claude more context: brand guidelines PDFs, founder interview transcripts, ingredient docs, product spec sheets. The more it knows, the better the briefs.

2. Claude Desktop with Cowork. Download the app at claude.com/product/cowork (Mac or Windows). Open it, click "Cowork" in the sidebar, and link it to your Claude Project. You need a paid plan: Pro ($20/mo), Max ($100-200/mo), Team, or Enterprise. If you are already paying for Claude, you already have access.

3. Claude in Chrome. Go to Settings > Connectors in Claude Desktop. Install "Claude in Chrome." This is what lets Cowork navigate real websites in your browser. One thing: you need to be logged into your analytics platform already. Cowork navigates your browser, it does not log in for you.

Now let's get into it.

Five steps. Each one, I am going to show you the exact instruction I typed into Cowork, and then what actually happened when it ran. Every instruction is copy-pasteable (they are all in the companion doc if you want to grab them without reading my commentary). The whole thing took about 90 minutes the first time. Now it runs on its own every Monday.

Before you write a single brief, you need to know what is already working and what is not.

I use Atria in this example, but this works with Motion, Triple Whale, Northbeam, or even raw Meta Ads Manager. You just swap the URL. H

ere is the exact instruction I gave it:

This is Template 1 in the companion doc. Swap the URL for whatever analytics platform you use. [LINK TO DOC]

Cowork opens Chrome. Navigates to Atria. Clicks through the UI. Switches to the right brand account. Scrolls through reports. Reads the numbers off the screen.

You might need to redirect it once or twice ("that is the wrong tab, go to System Reports > Top Performing Creatives"). It learns fast. The more specific you are in the initial instruction, the less you need to course-correct.

If you ran Issue 3's competitive analysis workflow, you already have benchmark data. Tell Cowork to pull it.

Alternatively, as a better solution, I built a tool called AdLib (adlib.getskipper.ai) that does this.

It is free. Consider yourself #blessed.

It pulls directly from the Facebook Ad Library (not from your ad accounts), sorts every active ad by estimated impressions (so the top of the list is what is actually spending), and lets you compare your brand vs. competitors across format, persona, angle, and emotion. I vibe-coded the whole thing.

Note: AdLib does not currently work with health and wellness brands due to Meta's API restrictions on that category. If that is you, the companion doc has an alt version using Facebook Ad Library directly

You can use it, use Facebook Ad Library directly, use Foreplay, or whatever you prefer. The point is getting competitive data, not the specific tool.

If you are using my tool, run a comparison before telling Cowork to review the data.

Which gives you this:

And then once that is loaded, tell Cowork:

Template 2 in the companion doc. There is also an alt version if you want to use Facebook Ad Library directly instead of AdLib. [LINK TO DOC]

If you do not have competitive benchmark data and do not want to set it up, skip this step. The performance data from Step 3 is usually enough. The high-ROAS/low-spend signals tell you most of what you need.

This is the step where everything comes together. Cowork now has your performance data (Step 1) and your competitive benchmarks (Step 2). It knows which angles are returning 2.6x on barely any spend. It knows which formats your category runs that you have never tested. It knows which personas you are ignoring.

Now you tell it to do something with all of that. And this is the part I did not expect to work as well as it did: you also tell it to search for what is happening in culture right now and factor that into the briefs. Sports events, wellness trends, seasonal moments. It scores everything by relevance and sorts the output so the most urgent briefs are at the top.

One thing to understand about the instruction below: it tells Cowork to match the copy to the format structurally. If the brief is a bullet-point ad, it writes actual bullets. If it is a testimonial, it writes a quote with a star rating. If it is Us vs Them, it writes a comparison grid. Most people write one block of copy and paste it into every format. That is why their ads all look the same. The format IS the creative decision.

Here is the exact instruction I gave Cowork:

Template 3. This is the big one. [LINK TO DOC]

Cowork cross-references everything. The high-ROAS/low-spend angles. The missing formats. The untapped personas. The emotion gaps. It searches the web for what is happening in culture right now. Then it picks the briefs that fill the most gaps simultaneously, scores them by urgency, and writes copy that is structurally matched to each ad format.

If you want to go further, you can have Cowork study real ad references and generate Nano Banana 2 images. If you manage a design team, this step replaces the part where you spend 45 minutes building a mood board in Figma. Cowork builds it from 600+ real ads and translates it into NB2 prompts automatically.

Here is the instruction I used:

Cowork navigates to each Foreplay board. Studies dozens of real ads per format. Identifies the patterns (green checkmarks for bullet-points, split comparisons for Us vs Them, bold percentage overlays for statistics). Then calls the fal.ai API with NB2 prompts.

This is the part that made me stop and rethink the entire way I approach creative briefing.

Once you have the workflow dialed (Steps 1-3), you can schedule it to run automatically. Cowork has a /schedule feature that turns any task into a recurring job. Daily, weekly, on specific days. You write the instructions once, pick a cadence, and Cowork runs it on its own.

Here is what I scheduled:

Claude saves the instructions, confirms the schedule, and adds it to your "Scheduled" tab in the sidebar. Every Monday, it spins up a new session, runs the full workflow, and the briefs are sitting there when you open your laptop.

You review. You give feedback. You approve or iterate. But the research, the data pulling, the gap analysis, and the first draft of copy? That already happened while you were not looking.

One important caveat: scheduled tasks only run while your computer is awake and the Claude Desktop app is open. If your laptop is closed Monday morning, Cowork skips the run and picks it up when you open it. Not a dealbreaker, but worth knowing.

Most people will not need this. The markdown briefs from Steps 1 through 5 are the workflow. But after a few rounds of feedback ("the copy is too long," "we need seasonal hooks," "sort everything by relevance"), I asked Cowork to turn the briefs into something I could actually work from. It built a live dashboard.

Editable copy fields. Relevance scores. Urgency tags. Nano Banana prompts. Visual references. Batch management. Connected to a database. Deployed to a URL.

15 briefs. Scored and sorted. The March Madness headline sits at the top (score: 80, tagged time-sensitive). Click into it and you see the angle (Confidence & Ease, ROAS 2.63), the persona (Everyday Unwinder), both headline alternatives, format-matched body copy, and a Nano Banana prompt ready to run directly in the platform.

You assign briefs to batches and each batch caps at 4 statics + 2 motion. When a batch is full, it is ready to produce.

I did not write a single line of code. I did not design the UI. I gave Cowork feedback in plain English and it rebuilt the whole thing each round. The dashboard template is in the companion doc (Template 6) if you'd like to try it, but most people will not need it. The markdown briefs from Steps 1 through 5 are the workflow. This is just what happens when you keep going.

I need to be honest about this part because the downsides matter.

It is a research preview. Sometimes it clicks the wrong tab. Sometimes the Chrome navigation is slower than doing it yourself. Sometimes it generates a 50KB React app that will not render and you spend 30 minutes debugging deployment.

The first run takes patience. I said 90 minutes, and that is true for elapsed time. But about 20 of those minutes were me redirecting Cowork when it clicked the wrong tab or scrolled past the data I needed. By the third week, the redirects dropped to almost zero because the scheduled task instructions were dialed in. The first run is the roughest. It gets smoother fast.

Browser automation is good, not great. Simple navigation (go here, pull data, read this page) is reliable. Complex multi-step workflows across authenticated platforms occasionally get lost. If your analytics platform has two-factor popups or unusual navigation, expect to hand-hold more on the first run.

Platform UI changes break things. If Atria redesigns their dashboard next month, the scheduled task might navigate to the wrong place. You will need to update the instructions. This is the same problem browser automation has always had. It is not unique to Cowork, but it is real.

It cannot replace your judgment. It can pull data, find patterns, and write copy. But whether "Your bracket is busted. Your morning does not have to be." is a good March Madness headline? That is still you. First draft quality is about 70%. The iteration loop (Step 5 feedback, then the weekly improvements) gets it to 90%. The last 10% is always human.

Limits burn faster. Cowork consumes Claude usage faster than regular chat. If you are on Pro ($20/mo), you will hit limits during a session like this. Max ($100-200/mo) is more realistic for heavy use.

It runs on your machine. If your laptop sleeps, the session pauses. Scheduled tasks skip and resume when you open the app. This is not a cloud service (yet).

But. Even with all of that. The briefing process that used to take me a full day now runs itself every Monday morning. I review and iterate for about 30 minutes. The system gets smarter each week because the project context accumulates. That is not a productivity hack. That is a fundamentally different way to operate.

[LINK TO COWORK TASK TEMPLATES: the exact instructions I gave Cowork, cleaned up as reusable templates you can paste directly]

I am going to go attempt to stand up from this chair, which, after two Barry's leg days in one week, is genuinely uncertain. James will be at the window when I get home. Staring outside. Waiting for his brother to chase something. Probably a squirrel. Probably a soft one.

If you try this workflow, please reply and let me know what you scheduled. Not what you asked Claude to do once. What you set up to run every week. I am building a database of recurring Cowork workflows and I genuinely want to see what people automate.

Have a great weekend,

Will