Last Wednesday on One Shot Show, our guest opened a blank folder, typed one sentence into Claude Code, and 45 minutes later had a working RSS reader running on the web.
There was no starter template. No code written by hand. There was a description of what he wanted, a lot of back‑and‑forth, and a few moments where he had to slow Claude down and tell it to keep going.
This was Season 2, Episode 4. Dheeraj Sharma co-hosted with me, and our guest was Mark Miller (Senior Storyteller). Mark reached out a while back after I mentioned I wanted to feature readers building real AI workflows. The thing that caught my attention was an RSS reader he built for himself to track the blogs and Substack posts he cares about. He logs into it every morning.
So I asked him the obvious question: could you show people how to build this from scratch, live, at the minimum useful version?
He said yes. And what the session actually taught was bigger than RSS.
The real lesson was this: The gap between something you can imagine and something you can actually use has shrunk to almost nothing.
Mark said it plainly near the end:
“This thing can do whatever you imagined. That’s my bottom line to you.”
I want to show you why a guy who has been coding since 1996 believes that, and what he did differently to make it true.
Why This One Is Worth Your Time
Most people reading this have an idea backlog. A tool you wish existed. A small app that would fix one annoying part of your week. Or an app that collects the 100+ newsletters you actually read into one clean feed, so they stop dying unread in your inbox.
The reason those ideas stay ideas is the build. Building used to be the wall. You needed to know how to code, or hire someone, or learn a no-code tool well enough to fight through it.
That wall is what Mark watched come down, and he is convinced this AI moment is the early internet all over again. “It’s the same thing that happened when the internet came out,” he said on the stream. “It scared everybody to death.”
So when he heard Boris, the creator of Claude Code, say on Lenny’s podcast that he had not touched a line of code since November, Mark’s reaction, from someone who wrote Perl for a living, was, and I am quoting him, “bullshit. There’s just no way that can happen.”
So he turned the podcast off, went to his computer, and half an hour later had a functioning app pushed to the web. His words: “in my mind I was going, this is impossible.”
That is where it started for him.
That is the shift worth paying attention to. Not the RSS reader. The fact that the building part has gotten small enough that the only real question left is what you want to make.
The One Habit That Made the Build Work
Here is the part I did not expect from someone with Mark’s background.
His coding experience was the thing slowing him down at first. When he started with Claude, his instinct was to define everything. Tell Claude what to do and exactly how to do it. He kept hitting walls. Then he realized the move was the opposite of how he had worked for a decade.
“Don’t tell Claude how to do something,” he said. “Tell Claude what you want done, and then follow his instructions.”
You could watch this play out the whole session. His very first prompt was not a spec. It was this: “I want to create an RSS reader that’s available for public viewing. Ask me questions.”
That last part matters. When you tell Claude to ask you questions, it walks you through the decisions instead of guessing. It asked Mark about AI features, visual style, where to deploy, which feeds to include. For each one, Mark mostly took the recommended option. Instead of steering the architecture, he was answering an interview about his own idea.
Within a few minutes a working reader showed up in the preview window. Plain, a little lame looking, but real. Running locally on his machine. From there the session was just refinement, all in plain English:
Pull in a featured image for each article, and if a source has not set one, use the author’s headshot instead.
Make it one card per row, image taking up about 30 percent, total width 850 pixels.
Add a left-hand menu listing the sources so I can filter.
Put a search bar under the menu so I can search the text of any article.
The headlines are too small, fix them.
At one point he stopped typing and just talked to it through the microphone. Same result. The app got better with every instruction, and you could see each change appear on screen as it happened. That is the thing about building this way. You are not imagining what it might look like. You are watching it become the thing you described.
None of those instructions required knowing what Next.js or Tailwind CSS is. Claude offered those, Mark accepted them, and they never came up again. As he put it, the job is not to know how to program. The job is to know enough to tell Claude what you want as the output.
Steal the Whole Build as One Prompt
Mark built this the slow way, one request at a time, and honestly that is how I would tell a beginner to learn. The back and forth is where you figure out what you actually want. But some of you just want to paste one thing and get a working reader.
So I took every request he made during the session and stitched it into a single prompt. Open Claude Code on an empty folder, swap in your own feeds, and let it run:
Build me a public-facing RSS feed reader, then run it locally and give me the URL so I can see it in my browser. If anything is unclear, ask me questions before you start. Otherwise build it, start a local server, and tell me the local address.
Feeds to include (swap in your own):
- https://aimaker.substack.com/feed
- [your second feed]
- [your third feed]
Some sources I will hand you as a page instead of a feed. When I do, find the correct RSS feed yourself. For example, find the feed for this Substack: https://substack.com/@acommunityarchitect, and the RSS feed for this YouTube channel or video: [paste a YouTube URL].
Layout:
- One article card per row. Each card is 850 pixels wide and centered on the page.
- Inside each card, show the article image on the left taking up about 30% of the width, with the title, source, date, and a short description beside it.
- Pull in the featured image for each article. If an article has no image, show a clean placeholder instead.
- Make the headlines large enough to read at a glance.
Navigation:
- A left-hand menu attached to the left side of the content area, listing each source as a filter.
- Pin the menu so it stays visible while I scroll.
- Put an orange bracket next to the active source in the menu.
- A search bar directly beneath the menu that searches the text of any article.
Background:
- A gradient that starts as deep royal blue on the left and fades into soft pink on the right, with the cards kept clean and readable on top of it.
Keep this first version simple, with no accounts and no database. I just want to run it locally, see my feeds, and click through to read.It will still ask you a question or two, and it will still show you a plain first version before it looks like anything. That part does not go away, and you would not want it to.
You can also execute this in one shot using Goals, where the agent will build it end to end—from planning and self‑verification to deploying it to Vercel.
It Was Never Really About RSS
Somewhere in the middle of the build, Dheeraj said the thing that reframed the whole episode. We were watching an RSS reader take shape, but the RSS reader was never really the point. The point was that the process underneath it now works for almost anything. You describe an app, you get an app, and you never have to know the code running it.
Mark gave the example that made it land for me. He pays for read.ai to transcribe his meetings. One day he opened Claude and typed “build me a substitute for read.ai.” Claude just said okay. Sit with that for a second. If you are paying for a tool you like, you can often describe it to Claude and get your own version of it. You can even point Claude at the app’s webpage so it sees how the thing is put together before it starts.
The real unlock is bigger than newsletters. The gap between “I wish this existed” and “here it is running on my screen” has mostly closed. Mark put a bow on it near the end: “This thing can do whatever you imagined. That’s my bottom line to you.”
But here is what gets interesting once building stops being the hard part. The hard part moves.
Dheeraj raised it live, and a viewer named Laurie backed him up. If you can pull in every feed you care about, you are one step from rebuilding the exact problem you were trying to escape. Dheeraj described his Gmail inbox as a graveyard of newsletters he never opens. Laurie said that back when she had a mountain of feeds in Feedly, deciding what to actually read got brutal. This explains why, if you build an RSS reader carelessly, you’ll just end up in the same position again, with too many newsletters you won’t read.
So the question shifts from “can I build it” to “what do I actually want this thing to let through.” Mark’s answer is deliberately small: show only the 10 most recent items from each source, and stop there. Dheeraj pushed it further, imagining an agent that reads what matters to you and surfaces the five things worth your time. Either way, the machine builds it in minutes. You are the one deciding what it is for and what it keeps out.
That is the part that does not get easier. When anyone can build anything, the scarce skill is deciding what is worth building and what is worth your attention.
Mark’s Rules of Thumb
Across the session Mark kept dropping one-liners that were really just years of trial and error compressed into a sentence. Here are the ones worth taping above your desk:
Check your model before you start. Sonnet 4.6 will do pretty much what you want on your first builds. Opus with the giant context window is a token hog, and using it as your default is how you go broke without noticing.
For anything bigger than a small app, start in a chat first. Talk through what you actually want, let it help you shape an outline and a plan, then take that plan into Claude Code to build. The plan is the work. The build is the easy part.
Push it when you want the result. Ask it to teach you when you want the skill. “You do it for me” gets the job done. “Walk me through this step-by-step” turns the same moment into a lesson.
Treat the first build as a minimum viable version, full stop. It works, and that is all it does. No security, no sanity check, no engineering review until you ask for one.
Be polite to it. Costs you nothing, and the back and forth just goes better when you are not barking at it.
What I’d Tell You to Do With This
My advice is simple: pick the smallest real thing you have wanted to exist. Not a business. A small app that fixes one annoying part of your week. Then:
Open Claude Code on a desktop plan, point it at an empty folder, and start with one sentence that ends in “ask me questions.”
Take the recommended answers when you do not know. You will learn the terms by being curious as they come up, not by studying first.
Keep it on your own machine to start. Local is safe and fast. Deploying to the web is a later problem, and a different skill.
When Claude tells you to do a technical step, try “can you do it for me?” before you assume it is yours.
Before you trust it with anything real, ask it to review its own work as a release engineer and a security reviewer. Read what it says.
You do not need Mark’s 30 years. He spent most of the session arguing that the experience was almost beside the point now. What you need is one idea small enough to finish, and the willingness to describe it and keep going.
The gap between thinking it and using it is the smallest it has ever been. The only thing left on your side of that gap is deciding what to build.
See you in the next one!
Watch the Full Session
This was Season 2, Episode 4 of One Shot Show.
We go live every Wednesday at 10:00 AM EST on Substack.
Timestamp:
(01:46) What we’re building today: your own RSS reader
(02:30) How Mark and I connected through the newsletter
(06:01) Mark’s background: coding since 1996, and why this feels like the early internet
(08:31) The story behind his first build: calling BS on “no code since November”
(09:09) His orchestra deep-dive site, built before he ever touched Claude Code
(10:04) A look at the full RSS reader he uses every morning
(11:00) “If you can think it, Claude can help you build it”
(14:39) The desktop setup: three windows, one blank folder
(15:41) The first prompt: “create an RSS reader, ask me questions”
(18:40) Watching Claude build it, and the “allow, allow, allow” lesson
(20:51) The core habit: tell Claude what you want, not how
(22:39) The first working version shows up in the preview
(24:34) Audience question: how do you keep a feed from becoming a doom scroll?
(25:49) Laurie’s point on Feedly, and using an agent to rank what to read
(27:47) Dheeraj: the real takeaway is not RSS, it’s building any app
(28:52) Why “Claude Code” might scare the people who’d benefit most
(30:23) Adding a YouTube feed, and the feed scraper for sites without RSS
(31:29) Does politeness actually change Claude’s output?
(33:46) The featured-image fallback, using a headshot when one is missing
(36:13) Refining the layout, and the model that quietly burned $20 in an hour
(38:21) Adding a source menu and a search bar
(42:14) Talking to Claude by voice instead of typing
(44:40) Deploying to Vercel live, and what you actually need to understand first
(52:54) Personas: the release engineer, security, and why the first build isn’t safe
(55:40) “Build me a substitute for read.ai,” and it just said okay
(57:49) The microphone button, RSI, and brainstorming out loud
(58:47) Wrap-up and next week
Resources Mentioned
Claude Code: the tool Mark used for the entire build, run from the desktop app. Mark noted you need a paid Claude plan to access the desktop version. The whole session happened here.
Vercel: free website hosting. Claude can push a project to it directly. Mark deployed the reader live during the show.
Next.js: the web framework Claude chose for the build. Mark never had to think about it.
Tailwind CSS: the styling Claude used. Same story, it just handled it.
Sonnet 4.6: the model Mark ran for this. His warning: an upgrade had quietly defaulted him to Opus 4.8 at 1 million context, and he burned through $20 in an hour. For a build like this, Sonnet was plenty.
Wispr Flow: the voice dictation tool I use to talk to Claude instead of typing. Mark mentioned getting repetitive strain before he found the microphone button.
Lenny’s Podcast: where Mark heard Boris, the creator of Claude Code, say he hadn’t written code since November. That doubt is what kicked off his first build.
read.ai: the meeting transcription tool Mark uses. He also asked Claude to build him a substitute for it, and it just started.
Feedly: the RSS reader Laurie brought up in the chat, and the source of the “millions of feeds, brutal to manage” problem an agent could help filter.
YouTube and Anthropic RSS feeds: every YouTube channel has its own RSS feed Claude can find. For sites like Anthropic’s research pages that lack one, Mark built a scraper into his full version.
