Skip to content
FSS Knowledge Base

Jeff Delaney video on AI Web Site & App Stack

Section titled “Jeff Delaney video on AI Web Site & App Stack”

Subject: Idea: The “Modern AI-Dev Web Site/App Stack” (Obsidian + Keystatic + Astro/SvelteKit)

Hey Jeff,

Huge fan of the channel, for years, for good reasons.

I’ve been refining an architecture for my business’ monorepo that completely strips out the bloat of modern SaaS wireframing/CMS tools, and I haven’t seen anyone covering this specific workflow yet. I think it would make a killer video.

The goal was a 100% local, zero-vendor-lock-in, AI-augmented workflow for building content-heavy sites (Astro). Later realized the same structure works for dev of apps/PWAs (SvelteKit).

Here is the stack:

1. Phase 1: Strategy & Flow -> Obsidian Instead of using Relume or Figma, the entire site/app structure and 20% core messaging (delivering 80% of the value) is mapped out in an Obsidian vault. Because it’s pure local Markdown, I use local AI (Claude Code / Gemini CLI) to analyze the vault, strategically enhance high-level business objectives, and expand the copy. The [[internal links]] act as a text-based, clickable UX prototype. Obsidian’s Canvas and Graph View provide visual structure as a bonus.

2. Phase 2: Schema Enforcement -> Keystatic The Phase 1 Markdown is pulled into Keystatic (running locally in the monorepo) to flesh out the remaining content, enforcing Astro/Svelte component schemas via MDX without writing raw code.

3. Phase 3: AI Localization -> GitHub Actions Because the source of truth is highly structured English MDX, an automated AI pipeline reads the files, and if needed, translates them into a /country/language/ routing structure. No clunky side-by-side human CMS translation UIs needed.

4. Phase 4: Render -> Astro / SvelteKit Astro and SvelteKit ingest the localized MDX and render the UI.

AI effective at every stage

  • This stack and workflow allows any AI tool (with agents and Skills) to help with every stage. Especially powerful for tools that operate locally (Claude Code, Gemini CLI, …).
  • Very efficient for AI context (tokens) since only dealing with text (in AI native markdown) instead of web HTML/CSS/JS.
  • Very effective for human context limits, focusing on what matters most (80/20).

The result is a workflow where strategy, content, translation, and code are perfectly separated, entirely Git-backed, and natively integrated with AI CLIs.

I know you appreciate setups that cut the BS and focus on developer velocity. If this sounds like something the Fireship audience would find valuable, feel free to steal the architecture for a video. Happy to share my repo or elaborate if it helps!

Cheers,

[Your Name]
[Link to your project/socials]