The argument in one line.
Slide decks built as HTML pages are interactive, deployable websites that AI models design better than any PowerPoint template, and a non-technical person with a coding agent and a skill can ship one today without writing a line of code.
Read if. Skip if.
- You make presentations regularly and find PowerPoint friction exhausting.
- You want a low-stakes first Claude Code or Codex project with an immediately useful output.
- You build software products and want to understand the design-for-agents shift.
- You have deep expertise and want to package your workflow so others or their agents can run it.
- You are already deep in agent infrastructure development.
- You need an offline-capable or regulated-environment slide format.
The full version, fast.
HTML slides are strictly better than PowerPoint: interactive media embeds, laser-pointer cursors, live links, deployable at a real URL. A Claude Code skill called Frontend Slides generates them from a plain content outline without manual coding. The creator built it through 10-20 rounds of iterative AI feedback, then extracted the workflow into a skill that went viral with 1.4M views and hit 22K GitHub stars. Three lessons distilled: skills will replace small-utility apps because agents absorb the interface layer; output personalization happens automatically at the agent layer without complaint floods; and HTML is genuinely the model's native output language because models trained on more web pages than any other visual medium.
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.

01 · Intro: 22K stars, 0 lines of code
Credibility hook: 22K GitHub stars, 1.4M launch views, built entirely through natural language prompting.

02 · What HTML slides can actually do
Live demo of a finished deck: clickable media, laser-pointer cursor, Easter egg reveals, live embeds, all impossible in PowerPoint.

03 · Demo: install, draft, design, build
Step-by-step walkthrough: GitHub install, paste an outline, answer 4 onboarding questions, pick from 3 cover previews, watch Claude Code self-correct in real time.

04 · Inserting media: the file-path tip
Practical tip: you cannot drag an image into Claude Code. You send the file path. Mac path-bar trick shown.

05 · Sharing and deploying with Vercel
Deploy the HTML deck as a real live URL via Vercel: just tell Claude Code to do it.

06 · Origin story: from one prompt to viral
Started from seeing an HTML deck at a talk. 10-20 feedback rounds to shape Claude output. Packaged into a skill. Went viral on X with 1.4M views. Added 30+ templates after users complained of generic output.

07 · Three lessons: skills, agents, HTML
Lesson 1: skills replace apps, agents as distribution layer. Lesson 2: one skill, different output for everyone, agents absorb customization. Lesson 3: HTML is the model's native output language.

08 · How to make your own skill
Four-step framework: Test, Iterate, Package, Share. The skill is the last step, not the first.

09 · Everything is code: start with slides
Closing argument: slides are the perfect vibe-coding entry project. No blank page. You already know what you want. Start here.
Lines worth screenshotting.
- A skill built with zero code can hit 22,000 GitHub stars: the barrier to shipping open-source tools has collapsed.
- HTML slides are a website, not a file: you deploy them at a URL, embed live media, and add cursor effects PowerPoint can never touch.
- The secret to avoiding AI design slop is showing three cover previews before generating the full deck, locking style direction before burning tokens on the wrong aesthetic.
- Onboarding a skill with four questions (purpose, length, content, density) is what separates a useful tool from an expensive random-output generator.
- Skills eliminate complaint floods because agents absorb user customization: users tell their agent to change things instead of filing tickets with the developer.
- You do not start with a skill. You build the thing first, iterate until it hits your taste, then tell the AI to freeze the workflow into a skill.
- HTML is the model's native output language because its training data is saturated with web pages, so asking it to design in HTML unlocks its full front-end capability.
- Apps rot in phone drawers; skills live inside the tool users already open every day. Skills win on distribution, not features.
- If Markdown is the model's native input language, HTML is its native output language: both are text, but only HTML is visually beautiful to human readers.
- The file-path tip is the one practical thing most demos skip: you cannot drag an image into Claude Code, you send the file path.
- Templates built from opinionated reference images prevent generic output at scale: the skill now ships 30+ auto-selected by mood and content.
- Slides are the ideal first vibe-coding project: no scary blank page, you already know the subject, and the output is immediately usable in real life.
Four steps that turn any workflow into a reusable skill
The skill comes last, not first: you build the output you want, beat it into shape against your own taste through rounds of feedback, then freeze the workflow so anyone's agent can run it.
- Build the actual output you want before thinking about packaging: the skill is what you extract from an iterative build process, not what you plan before it starts.
- Front-load the model with structured context (purpose, density, length, content) before generating any artifact or it produces generic, mediocre output every time.
- Show design-direction options as previews before building the full artifact: locking style early prevents wasted tokens and saves users from results they would reject anyway.
- Skills eliminate complaint floods because agents absorb customization at the output layer: users tell their agent to change things instead of filing feedback with the creator.
- HTML is the right output format for any visual knowledge work: models trained on more web pages than any other designed medium, so HTML unlocks their full design capability.
- Slides are the lowest-risk first vibe-coding project because there is no blank-page problem: you already know the subject, the format is clear, and the output is immediately useful.
- The four-step build loop (Test, Iterate, Package, Share) applies to any repeatable workflow you have taste about: the only requirement is caring enough to give it 10-20 rounds of feedback.
Terms worth knowing.
- Skill
- A reusable instruction set that wraps expertise, taste, and workflow into something a coding agent can run on command, installed via a GitHub link rather than an app store.
- Frontend Slides
- An open-source Claude Code skill that takes a content outline and generates a complete, interactive HTML presentation deployable as a real website.
- Vibe coding
- Building software through natural language prompts to an AI coding agent without writing code manually, iterating on output through description and feedback.
- Design slop
- Generic, aesthetically mediocre AI output produced when the model lacks enough opinionated context about style direction before generating.
- Onboarding gate
- A set of structured questions a skill asks a user before generating anything, used to give the model enough context to produce non-generic output.
Things they pointed at.
Lines you could clip.
“You're not making a PowerPoint. You're making a website.”
“Instead of you going to find a software, the software comes to you.”
“You don't start with a skill. You end with a skill.”
“If Markdown is the model's native language on the input side, then HTML might be its native language on the output side.”
“The biggest winners of vibe coding are not programmers. They're everyone else.”
Word for word.
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
The bait, then the rug-pull.
She launched it a few months ago and it kind of exploded: 22,000 GitHub stars, 1.4 million launch views, zero lines of code she wrote herself. Zara Zhang's Frontend Slides skill is a proof-of-concept for something bigger than decks: that anyone with taste and a coding agent can ship open-source tools, and that the most beautiful output format AI has ever had might not be Markdown.
Named ideas worth stealing.
Skill-building 4-step loop
- 01 Test
- 02 Iterate
- 03 Package
- 04 Share
Build the output you want first, tweak it against your own taste, then freeze the workflow into a skill. The skill is always the last step.
Onboarding context gate (4 questions)
- Purpose
- Length
- Content
- Density
Front-load a model with these four questions before generating or it produces slop. Works for any AI tool that needs user intent upfront.
3-cover design lock
Show three cover-page design previews before building the full artifact. Locks style direction before spending tokens on the wrong aesthetic.
How they asked for the click.
“The link to install it is in the description. Go make something beautiful.”
Soft and generous: GitHub link plus encouragement, no upsell. Paired with VIBE CODE YOUR SLIDES slide on screen showing github.com/zarazhangrui/frontend-slides.







































































