How I built this

Claude (Anthropic)

Used to plan features, audit the live site, research facts and sources, draft the copy and turn ideas into exact instructions for Lovable to build. It also runs two scheduled research jobs, described below.

Claude on a schedule

Two jobs run in Claude without me opening it. Neither one publishes anything to this site. Each hands me a report that I read first, and I can change the written rules they follow.

Daily AI announcements digest

Every morning at 8:52 a.m. Eastern, Claude checks what the frontier AI labs, their CEOs, and a handful of research groups and investors announced in the last 26 hours, and sends me a short digest as a phone notification.

It checks official feeds and pages first: OpenAI, Anthropic, Google DeepMind, Meta, xAI, Microsoft, Mistral, DeepSeek, Alibaba’s Qwen and Moonshot AI, plus a16z, METR, Epoch AI, Artificial Analysis, Stanford HAI, Hugging Face and the Allen Institute for AI. Then it searches for coverage.

Every item carries a date and a link, and is marked “Official” (published on that organization’s own site) or “Reported” (found through a news outlet, which is named). Claude can’t read X directly, so a social post only counts if an outlet reported it or it sits on a public page. If nothing can be verified, the digest says so, and it always ends with a list of sources it couldn’t reach.

What it can get wrong: it can miss things or misread a page, and “Official” means it appeared on the organization’s own site, not that the claim is true. I treat the digest as a list of leads and open the source before I use anything. The public version of this is Lab Watch, which tracks announcements from the labs’ own sites.

Weekly live-events research

Every Monday morning, Claude researches free, dated events for the Live Events & Competitions collection: webinars, hackathons, capture-the-flag competitions and pitch competitions. It starts from the same sources each week (WiCyS, BrightTALK, SCORE, the SBA, Devpost, Major League Hacking and CTFtime) and adds others it finds, looking about eight to nine weeks ahead.

An event only makes the list if Claude opens the event’s own registration page and confirms a real date and that it’s free. Anything it can’t confirm, anything that charges and anything already closed is dropped, not guessed at. Every event in that collection has a date and goes stale, so it’s re-researched from scratch each week.

What I get back is a spreadsheet in the columns this library uses, plus a short note on how many events it found, what it left out and why, and which sources came up empty. It’s a draft: I read it before anything goes into the library. Dates can still change after Claude checks them, so the event’s own page is the final word.

This description was drafted by Claude from the settings of the two jobs, and I reviewed it.

Lovable

The tool that actually writes and deploys this site's code — every feature here, from the search page to this explainer, was built by describing it in plain English rather than writing code by hand.

shadcn/ui + Tailwind CSS v4

The component and styling layer that makes every button, card, form, dialog, badge and responsive layout work. Tailwind CSS v4 handles the colors, spacing and typography through utility classes; shadcn/ui supplies the accessible, reusable components built on top of it. Everything you click here is assembled from that stack.

Algolia

Powers the search bar and the search page — it handles typo tolerance and result ranking across 3,488 resources, so a misspelled search still finds the right thing.

Microlink

A paid service that takes a picture of a page. Most cards here show the preview image the site publishes itself, but plenty of sites publish none — or block automated visits altogether. For those, Microlink renders the page and returns a screenshot, which I store here rather than hotlinking, so the card shows the real thing instead of a placeholder.

Seedream

An AI image model I use for reference art, not finished art. I describe the mark I want — say, a hard hat for apprenticeships — in the site's terracotta, look at what comes back, and then the shape gets redrawn as a clean, sharp icon in the site's own code. Nothing it generates ships as-is, and no stock art is used anywhere on the site.

Creative Commons & public domain

The legal foundation for every sourced photo on the site. I only use images marked CC0, CC BY, CC BY-SA, or clearly public domain, and each profile shows a visible credit link to the original file and its license. I avoid anything labeled NC, ND, editorial use only, or 'all rights reserved' — if the license isn't clear, the photo doesn't go up.

Raindrop.io

My bookmark manager, and the first stop for every resource on this site. Anything I come across during the day gets saved and tagged there first; only what survives a second look gets researched properly and added to the library. The free plan covers what I need.

Jev (Typesafe)

Now live on the search page: when you type a phrase like "how to negotiate a higher salary", Jev reads the words and says which section of the library they point at, so results from that section are lifted to the top — and the page tells you which one it picked, with a one-click way to narrow to it. It only steps in when it's clearly sure, and if it's slow or unavailable, search carries on exactly as before. It also helps keep the keyword tags consistent: 3,523 entries, 3,770 different topic tags, about 7.3 tags each. I still make the final call on every tag and category.