talks.soubiran.dev is the static archive for the talks I give at conferences and meetups. It publishes each Slidev deck at a stable URL along with its source, thumbnails, PDF, and optional recording resources. The root redirects readers to the talk index on soubiran.dev.

The archive also acts as a content source. My personal website builds its talk pages from the published metadata, while mcp.soubiran.dev uses a separate catalog and the available transcripts.

Why each appearance has its own workspace

I keep the talks in a public pnpm monorepo. Each dated directory contains an independent Slidev project for one event appearance. When I give a talk again, I copy it into a new directory instead of making the old event URL point to changing slides.

This keeps event-specific edits and resources attached to the version that the audience saw. The workspace still shares dependency versions, root commands, and one local Slidev theme, so a visual or tooling change does not need to be maintained in every deck.

The pattern fits a collection whose entries need stable histories and URLs but use the same presentation system. The source repository remains the place to inspect a deck or reuse its code.

How decks become a catalog

Each talk splits its metadata between two sources. Its package.json records the event, location, description, and optional recording or article. The frontmatter in slides.md supplies the title, language, and topics. A generation script reads both files and gives every appearance a date-based ID and public resource URLs.

The root build compiles every Slidev workspace into one dist directory, then adds redirects and two JSON files:

  • meta.json contains the display data, resource links, and statistics used when soubiran.dev builds its talk index and transcript pages.
  • talks.json is a versioned content catalog with stable IDs, topics, event details, and links. mcp.soubiran.dev validates this file before it lists talks or retrieves a transcript.

Thumbnails and PDFs follow a separate authoring step. The thumbnail scripts copy light and dark images into each deck’s public directory, so the next Slidev build includes them. PDF exports and audio files are copied to dedicated asset hosts. Generated redirects keep /pdf, /audio, /recording, /transcript, /article, and /src links consistent even though those resources come from different places.

How publication stays ordered

GitHub Actions builds the complete archive, lints the repository, and tests the shared theme. Cloudflare Builds runs the production build and deploys dist through Workers Static Assets.

A push to main also sends an authenticated request to redeploy.soubiran.dev. This Cloudflare Worker starts a Cloudflare Workflow with the Talks Worker name and the deploy hook for soubiran.dev. The Workflow checks Cloudflare until the latest successful Talks build is the version running in production, then calls the deploy hook. The new personal website build can now fetch the matching meta.json. Without this ordering, it could fetch the previous catalog and need another deployment before a new talk appeared.

Decks, catalogs, and transcripts are public. The redeployment endpoint is not. The workflow sends Cloudflare Access service credentials from GitHub Actions secrets, which keeps deployment control separate from the public archive.

Ecosystem