Hundreds of things happen in artificial intelligence every day, scattered across news sites, laboratory blogs, forums, podcasts and leaderboards. The AI Daily Signal reads it all and sends one transmission: the developments that mattered, grouped into themes, explained in plain English with every piece of jargon defined.

How it works

An automated pipeline gathers the day's coverage from news outlets, official laboratory announcements, Reddit, Hacker News and benchmark trackers. AI clusters the stories, measures how strongly each theme registered across sources — that's the signal meter you see on every story — and drafts the summaries. A human editor reviews and approves every transmission before it is published. Nothing goes out unread.

What's coming

A weekly roundup, a YouTube edition of the daily transmission, and practical training built on what we learn every day — starting with a course on task design: how to brief AI systems that can work autonomously for hours, where defining the task well matters far more than wording the perfect prompt.

How this site was built

There is no development team behind The AI Daily Signal. The site was designed, built and is operated day to day by one person directing AI tools — and this page exists to show what that means in practice.

The division of labour:

Claude (claude.ai) acts as architect and operations engineer. Through live connections to the infrastructure (the Model Context Protocol, MCP), it designs the systems and then runs them in conversation: deploying server functions, applying database changes, reading logs, diagnosing failures and fixing them — often while the morning edition is mid-flight.

Claude Code writes the website itself: every page, component and translation in this codebase, working from written briefs, verifying its own work against live data, and pushing to production.

Supabase is the machine room: a Postgres database, eight scheduled server functions, authentication, and storage for the podcast audio. Vercel hosts the site, deploying automatically from GitHub on every change.

Each morning, unattended, the pipeline gathers several hundred items from news feeds and podcast transcripts. The Anthropic API drafts the day's briefing, newsletter and podcast script — through a relay design in which the database itself dispatches the long AI calls, because serverless functions are not allowed to wait that long. ElevenLabs records the approved script as the day's episode, chapter marks included. The edition is machine-translated into ten languages. The data desks fill themselves: model benchmarks from Artificial Analysis at dawn, end-of-day market prices, and a curated global picture of AI infrastructure.

One rule is absolute: nothing publishes without human review. The machine drafts; the editor reads, corrects and presses publish. The marginal running cost is a few pounds a month.

Built with AI: the projects

The same approach, applied to other problems. This list grows.