Kel Wu

AI News Engine

Case study · Side project

A content pipeline I built for @productbykel. It turns the day’s AI news into Instagram reels and carousels, learns from how each post performs, and publishes nothing until I approve it.

  • Role Product, design and build
  • Timeline April 2026 to now
  • Built with Claude Code
AI News Engine dashboard

The problem

Posting AI news by hand doesn’t scale

I wanted @productbykel to cover AI news on a steady cadence for PMs and builders. Doing it by hand meant reading feeds, picking stories, writing copy, designing slides, recording voiceover and cutting video for every single post. The cadence fell apart whenever my day job got busy.

Full automation wasn’t the answer either. An account that posts a wrong fact about a model launch loses trust fast. I wanted the machine to do the legwork and keep me as the editor.

How it’s built

AI News Engine: from headline to published post

Claude does the research and drafting. Two human checkpoints decide what ships, and post performance feeds the next run.

  • AI step
  • Human checkpoint
  • Code
  • Service
  1. Ingest
    Code

    Every other morning a cron job pulls AI news and ranks stories by source coverage, PM relevance and freshness.

    • Vercel cron
    • NewsAPI
    • TLDR AI
  2. Pick the stories
    AI step

    A Claude agent reads the full articles, picks a lead story, suggests carousel stories and recommends a format.

    • Claude Sonnet 4.6
    • Performance signals
  3. Editor's cut
    Human checkpoint

    I choose the final three carousel stories. Claude pre-selects them and I can override any pick.

    • Dashboard
  4. Write the copy
    AI step

    Slide copy, the reel script and image prompts come back as structured JSON through forced tool calls.

    • Claude Sonnet 4.6
    • Forced tool use
    • Humanizer rules
  5. Produce the media
    Service

    Images, voiceover and video are generated, then rendered as a vertical reel and five square slides.

    • Flux 1.1 Pro
    • ElevenLabs
    • Remotion Lambda
  6. Review
    Human checkpoint

    I preview the voiceover, reel and slides. Nothing publishes without my approval.

    • Dashboard
  7. Publish and learn
    Code

    Approved posts go out on a schedule. A daily job pulls views and likes back so the next run knows what lands.

    • Instagram
    • YouTube
    • LinkedIn
App
Next.js on Vercel, password-gated
Data
Supabase Postgres and Storage
AI
Claude Sonnet 4.6 with forced tool use
Media
fal.ai Flux 1.1 Pro, ElevenLabs, Remotion Lambda on AWS
Channels
Instagram Graph API, YouTube Data API, LinkedIn API

Version two

Closing the loop

The first version made content. The second one learns from it. A daily job pulls views, likes and comments from Instagram and YouTube and matches each post back to the episode that produced it. From there the engine works out which formats and topics land, and passes those findings to Claude as performance signals the next time it picks stories. When the sample is too small to trust, it stays quiet instead of guessing.

I also added a news hub. It ranks stories by how many sources are covering them, how relevant they are to PMs and how fresh they are, and it merges near-duplicate headlines. Every story offers three paths: an automated carousel, an automated reel, or a filming brief with a hook, talking points and a call to action so I can shoot the reel myself.

Product decisions

The calls I made and why

Approval before publish

Claude drafts, I decide. Two gates keep a wrong story or a bad render off my accounts.

Remotion over a video API

I picked Remotion over Creatomate to own the templates and the render pipeline instead of renting them.

Swappable image model

All image generation lives in one route, so trying a new model is a one-file change.

One source of truth for voice

A shared set of humanizer rules goes into every prompt so captions sound like me, not like a model.

What broke

Running it in production

Timeouts

Long Claude calls hit Vercel’s function limits. I raised the limits on the heavy routes and collapsed a multi-step agent loop into a single pass.

A token that expired quietly

Instagram tokens last 60 days, and the sync failed silently for weeks once mine ran out. The engine now refreshes its own token every week.

Disappearing thumbnails

Instagram image links are signed and expire within days. Thumbnails now get copied to Supabase Storage once and served from there.

Open endpoints

The production URL sat outside Vercel’s protection, so anyone could have triggered a render or a post. Every page and API route now requires a login, and the cron jobs use their own secret.

Stack

  • Next.js
  • Vercel cron
  • Supabase
  • Claude Sonnet
  • Forced tool use
  • Flux 1.1 Pro
  • ElevenLabs
  • Remotion Lambda
  • Instagram Graph API
  • YouTube Data API
  • LinkedIn API
View on GitHub See the posts on Instagram Back to portfolio

The dashboard itself is private because it can publish to my accounts.

Ask me anything! ↓