← All plans
Exploratory

joshcoolman.com/news — Daily AI Video Digest

Context

The /ai-news command already works great locally: it calls ~/scripts/yt-ai-news.py to fetch recent AI videos from YouTube subscriptions + discovery searches, then Claude curates them into a digest. The goal is to automate that entirely so joshcoolman.com/news stays fresh without any manual step.

The sandbox app is Next.js App Router, file-based markdown content (no DB), deploys to Vercel on push. The pattern for adding a new content section is well-established (blog, docs, plans).

Auth

Locally — nothing new needed. ~/.config/yt-ai-news/tokens.json + existing env vars handle it.

In GitHub Actions — yt-ai-news.py currently bootstraps from a token file, not env vars. A small patch lets it accept YOUTUBE_REFRESH_TOKEN as an env var and write the token file itself before the normal refresh flow runs. Required GitHub secrets:

SecretValue
YOUTUBE_API_KEYexisting env var value
YOUTUBE_CLIENT_IDexisting env var value
YOUTUBE_CLIENT_SECRETexisting env var value
YOUTUBE_REFRESH_TOKENrefresh_token field from ~/.config/yt-ai-news/tokens.json
ANTHROPIC_API_KEYyour Anthropic key

Architecture

GitHub Actions (daily cron, 8am UTC)
  ├── python3 scripts/yt-ai-news.py     # fetch raw video JSON
  └── node scripts/generate-news.mjs   # curate via Claude SDK → save news/YYYY-MM-DD.md
       ↓
  git commit + push
       ↓
  Vercel auto-deploys
       ↓
  /news  →  latest digest + sidebar of past summaries

Files to Create / Modify

scripts/yt-ai-news.py — copy into repo + CI patch

Copy ~/scripts/yt-ai-news.py into the sandbox repo at scripts/yt-ai-news.py. Add ~10 lines near the top of get_access_token(): if YOUTUBE_REFRESH_TOKEN env var is set and no token file exists, write a synthetic tokens.json (expired access token + the refresh token) so the normal refresh flow picks it up. No behavior change when the token file already exists.

scripts/generate-news.mjs — new

Node.js script using @anthropic-ai/sdk. Reads raw JSON from stdin or $INPUT path, sends to claude-sonnet-4-6 with the curation prompt (same rules as the local /ai-news command — keep technical deep-dives/tutorials/announcements, skip clickbait/non-English/affiliate), requests markdown output, writes to news/YYYY-MM-DD.md with frontmatter:

---
title: "AI News — May 19, 2026"
date: 2026-05-19
videoCount: 11
---
[curated digest grouped by theme]

Use prompt caching on the system prompt (the curation rules are static — good cache candidate).

.github/workflows/daily-news.yml — new

on:
  schedule: [cron: '0 8 * * *']   # 8am UTC
  workflow_dispatch:               # manual trigger

jobs:
  generate:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: actions/setup-python@v5  # 3.11
      - uses: actions/setup-node@v4    # 20
      - run: pip install requests
      - run: npm ci
      - run: python3 scripts/yt-ai-news.py > /tmp/videos.json
        env: [YOUTUBE_* from secrets]
      - run: node scripts/generate-news.mjs
        env: ANTHROPIC_API_KEY, INPUT=/tmp/videos.json
      - run: |
          git config user.name "github-actions[bot]"
          git config user.email "github-actions[bot]@users.noreply.github.com"
          git add news/
          git diff --cached --quiet || git commit -m "news: $(date +%Y-%m-%d)" && git push

news/ directory — new

Markdown files: news/2026-05-19.md, news/2026-05-18.md, etc. Add .gitkeep initially. The first real file gets committed by the workflow.

lib/news/types.ts — new

export type NewsSummary = {
  slug: string        // "2026-05-19"
  date: string
  title: string
  videoCount: number
  content: string
}

lib/news/loadNews.ts — new

Follows lib/blog/loadBlog.ts pattern exactly (uses gray-matter, fs, React cache()):

  • getAllSummaries() — reads news/*.md, parses frontmatter, sorts date desc
  • getSummaryByDate(date: string) — reads single file

app/news/layout.tsx — new

Sidebar layout. Sidebar lists all past dates as nav links (/news/2026-05-19, etc.). Follows app/(docs) pattern. No entry in the main nav — just discoverable at /news.

app/news/page.tsx — new

Server component. Loads getAllSummaries()[0] (latest). Renders full digest content + sidebar. Metadata: title: "AI News", description: "Daily curated AI video digest".

app/news/[date]/page.tsx — new

Dynamic route. generateStaticParams() returns all date slugs. Renders single digest with MDXRemote (reuse the blog rendering chain: remark-gfm, rehype-slug).

app/sitemap.ts — update

Add /news to staticRoutes. Dynamically add all /news/[date] routes from getAllSummaries().

public/llms.txt — update

Add: news: daily curated AI video digest, auto-generated

Repo-Specific /ai-news Skill (optional)

Create .claude/skills/ai-news/SKILL.md in sandbox — identical curation rules as ~/.claude/commands/ai-news.md but references python3 scripts/yt-ai-news.py (repo-relative path). Useful when running Claude Code from inside the sandbox repo so you don't need the global command. Also opens the door to adding "save this digest to news/" as an extra step.

Verification

  1. node scripts/generate-news.mjs with a sample JSON → check news/YYYY-MM-DD.md created with correct frontmatter
  2. npm run dev → localhost:3000/news → latest digest renders, sidebar shows all past dates
  3. npm run build → no type errors, all /news/[date] routes statically generated
  4. Manual workflow_dispatch trigger on GitHub → Actions log passes, new commit appears on main
  5. Vercel preview URL shows /news updated with new content