Anatomy of Newsbeat: How My Daily AI Brief Actually Gets Built
Build it with AI · Episode 1
Every morning, a fresh edition of Newsbeat is waiting for me.
The lead story. Today's brief. The AI events coming up near me.
I don't touch any of it.
In Do it with AI, I showed you what I built. Build it with AI is the sister series. It goes a level deeper into how real AI use cases actually get built.
So for the very first episode, I'm taking Newsbeat apart, piece by piece, to see what makes it tick.
What Newsbeat Is
Newsbeat is my daily AI news briefing. It covers three beats:
- Agentic AI
- Model provider updates
- AI and agent events in the Northeast
The output is a website. A lead story up top, today's brief, the rest of the stories, and a panel of upcoming events. Open any story and it shows the full details in a dialog, without leaving the page.
The Anatomy: One Loop, Five Pieces
From thirty thousand feet, Newsbeat is one loop that runs every single morning.
- The scheduled task. It all kicks off here. It's just a prompt covering agentic AI, model providers and what's happening in the Northeast, plus a time: 6 AM.
- The research. At six, an agent goes to work. It searches, reads the actual sources, and writes up notes, with links.
- The Space. All of that lands in a Space, page by page: the daily brief, the stories and the events.
- Dewy. Dewy, my always-on agent (and the Dewanly mascot), picks up that Space.
- The website. Dewy updates the site with today's stories and events.
Research in. Site out. Every morning.
Now let's build each piece.
How I Built It
Here's the build in the order I did it in the episode. The Space, the pages, the scheduled tasks and the site all live in ChatGPT.
Step 1: Create a Space
The Space is where all the data is stored.
I created a new Space and called it Daily News.
Step 2: Write the Editorial Guide
Inside the Space, I opened Show pages, created a new page and called it Editorial Guide.
Why a separate page and not a longer prompt? I want to maintain it outside of the task itself. When my standards change, I edit one page. The task stays the same.
The guide covers four things: what to cover, source standards, what happens in the daily edition, and branding. It's further down in full.
Step 3: Create the Research Task From Chat
I created the scheduled task straight from chat. I used @ to point at the guide and the Space, so it doesn't have to guess:
Create a new scheduled task for 6:00am EST that researches the web for the latest updates on Agentic AI, the latest from model providers and any upcoming events. Follow the guidance from @Editorial Guide and make sure to put the content in the @Daily News space.I switched the model to GPT-6.1 Sol before sending.
For schedules that follow daylight saving time, select America/New_York (Eastern Time).
It came back with a task called News Beat Daily, repeating daily at 6:00 AM ET.
Step 4: Run It Now
Don't wait until tomorrow morning to find out if it works.
I went to my scheduled tasks, opened News Beat Daily and ran it right away.
A few minutes later there was a new page in the Space, titled October 27th, 2026, under Edition Archive. All three sections, primary-source links, event logistics and clearly labeled caveats.
Step 5: Build the Site From a Design Mockup
Same chat. I used Claude design to mock up how I wanted the site to look. I dragged those designs in and asked:
@Sites create a new site for my daily news. Use @October 27th, 2026 for the content.About nine minutes later, I had the site. Upcoming events, the lead story, today's brief and all the other stories. Each story opens in a dialog with the full details.
Step 6: Schedule the Site Refresh
The first build is a one-off. To keep the site current, it needs its own scheduled task.
I went to Sites, opened the site, and went to its settings. There's a Scheduled tab where you can create a task that refreshes the site:
Create a new task that runs at 7:00am EST that updates the News Beat site [@your site] based on the latest document from @Daily News. The site should always show the latest news and events from today's dated document under Edition Archive.Daily, 7:00 AM ET. That's it. The site now refreshes itself on that schedule.
Why 7 and not 6? It gives the research task enough time to finish first.
The Editorial Guide
I promised in the video I'd share this. Here it is.
The scheduled task tells the agent when to work. The Editorial Guide tells it what good looks like. Without it, you get a summary of whatever the search engine returned. With it, you get something you can trust.
A few things it does that I wouldn't skip:
- Source standards. Use original announcements, docs, papers and organizer pages. Read the source behind a search result before treating it as verified. Link every item.
- No invented facts. Never make up availability, registration status, pricing, performance claims, dates or citations. Keep the source's caveats. Label analysis separately from facts.
- Dedupe. No repeating the same announcement across categories or editions. If a category has nothing meaningful today, say so instead of padding it.
- Daily edition rules. Use the date in my time zone, one dated page per edition under Edition Archive, and don't overwrite my edits or past editions.
Copy it, then make it yours. Swap in your topics, your region, your sources and your brand. Some lines are specific to my setup, like my separate ChatGPT tasks and the site access note.
Editorial Guide
Coverage
1. Northeast AI and agent events: Connecticut first, then Boston, New York, and the wider Northeast. Record event date, local time, city, format, organizer, and registration link when verified.
2. Agentic AI: meaningful developments in agents, tools, frameworks, practical workflows, research, and adoption. Include LangChain/LangGraph primary documentation, engineering posts and Harrison Chase's public work, alongside other original sources on orchestration, context, memory, harnesses and evaluation. Check practitioner and framework developments separately from major-provider launches; follow linked original evidence, label vendor claims and coverage/access limits, and do not promise exhaustive X coverage.
3. Model provider updates: relevant announcements from OpenAI, Anthropic, xAI, and other model providers.
Source standards
Use original provider announcements, official documentation, research papers, and organizer event pages wherever available. Read the source behind a search result before treating its claims as verified. Link each item to its supporting source. Distinguish publication dates, update dates, and event dates.
Prioritize meaningful verified developments from the past 24 hours. Clearly label older background and upcoming events with their original dates. Within each workflow, avoid duplicate announcements across categories and editions; the dewy workflow may compare its previous completed Page. A category with no meaningful verified update should say so rather than be padded.
Summaries must preserve source caveats. Label analysis separately from reported facts. Never invent availability, event registration status, pricing, performance claims, dates, or citations. Preserve full source evidence and useful context in the daily Page; cards may use shorter summaries.
Daily edition
For the Space and Site workflow, use the edition date in America/New_York and actual research start, cutoff and completion times. Create or reuse exactly one primary Page directly under Edition Archive, titled only with the date, for example October 11th, 2026. The same Page holds its reader and full native editable article content, source links, evidence and caveats. Preserve owner renames, moves and edits, stable article order and identifiers within the Space and Site, completed editions and visible corrections. The separate ChatGPT plain and interactive tasks research and deliver independently without reading or writing these Pages.
Branding and access
The brand is News Beat. Apply the owner's supplied News Beat visual references: light blue-white surfaces, dark navy text, cobalt blue, coral for events, and amber for model providers. Keep the Space and dated Pages owner-private. The companion Site remains public at the owner's explicit October 16 request.What I Learned Building It
- Keep the rules out of the task. The task is short. The guide is long. Each one is easier to change on its own.
- Run it now. Testing the task right away beats waiting for tomorrow's 6 AM run to find a problem.
- Stagger the schedules. Research at 6, site at 7. The site should only read an edition that's finished.
- Tag, don't describe. Use @ to point at the exact guide, Space, page and site. Less guessing.
- Set up the refresh from the site itself. My first try was from the main chat, and I got stuck telling it which site. Creating it from the site's Scheduled tab was cleaner.
- Design first. A quick mockup gave the site builder something concrete to aim at.
Research In, Site Out
Newsbeat looks like one thing. It's actually a handful of small, boring pieces: a prompt with a time, a research run, a Space, a guide and a site refresh.
None of them is hard on its own. Wired together, they give me a fresh briefing every morning without me touching it.
Start with the Space, the guide and the 6 AM task. Add the site when you're ready.
Watch the full episode: https://youtu.be/rn-TzFwE4B4. Subscribe for the next build, and tell me in the comments what I should take apart next. More at mocherif.com.