Colophon

How this site is built.

This website is my personal playground to experiment with new technologies. Every page is built to be read by humans and AI agents. Here's the build, in detail.

Stack

Framework: Astro 6 , static-site generator, content collections with Zod schemas.
Hosting: Amazon S3 behind CloudFront. Static-only, no backend, no session state.
Deployment: GitHub Actions on every push to main. Build → S3 sync → CloudFront invalidation → IndexNow ping.
Search: Pagefind — fully client-side, no server, no analytics on queries.
Editor: VS Code with Claude Code. Hand-tuned.

Agent-readiness surfaces

Every page on this site also has a machine-readable copy underneath it. Each file below does a specific job for the AI agents reading it.

Discovery

How agents figure out what's available without crawling the entire website.

/robots.txt: Typical crawler permissions, but with a bonus: a Content-Signals section with specs to help AI agents use search, ai-input, ai-train on my website.
/ai-policy.txt: My policy for AI-use and citation-attribution. robots.txt points here when an agent needs more nuance.
/.well-known/agent-skills/index.json: The site-wide manifest (RFC agent-skills v0.2): every readable surface listed with a sha256 content hash, so agents can verify content hasn't changed before re-fetching.
/.well-known/llms.txt: RFC 8615 alias of /llms.txt. Covers crawlers that probe the standardised well-known path first.

Site-wide content

Three versions of the same content, built for different kinds of agents.

/llms.txt: Curated index. A short table of contents pointing AI tools to my essays, glossary, and well-known paths. Useful when an agent lands directly on an article instead of the homepage.
/llms-full.txt: Full corpus, every essay concatenated into one plain-text file, for agents that prefer one fetch over many. Kept in sync with each page's own llms.txt.
/corpus.json: Structured catalog of every essay (title, slug, pillar, publication date, URL) — for programmatic indexing without parsing raw HTML.

Standards-based feeds

The two feed formats most readers and crawlers already speak.

/feed.json: JSON Feed 1.1 — modern feed format for readers and agents that prefer JSON over XML.
/rss.xml: RSS 2.0 — legacy-compatible feed for traditional readers and crawlers that don't support JSON Feed.

/now snapshot

/now is the freshest page on the site, so it gets its own machine-readable bundle.

/now.json: JSON snapshot of the current /now page: where I am in the world, what I'm working on, thinking about, or doing.
/now/llms.txt: Plain-text render of /now with the freshness badge baked in, designed for AI grounding when the agent only cares about the "what's current" beat.
/now/rss.xml: Change-feed: a single-item RSS whose pubDate moves only when /now actually changes — gives agents a cheap signal to re-poll.

Every article, glossary term, press entry, and stream item also get an accompanying llms.txt sibling — same content as the rendered page, with an attribution header at the top.

Build-pipeline highlights

Auto-derived /now sections. The "Recently" block on the /now page pulls from Stream and Writing at build time, so the page stays fresh even when the handwritten content ages.
Citation reciprocity in JSON-LD. Press entries emitted twice (standalone NewsArticle nodes plus compact citation arrays inside Person/ProfilePage). Glossary terms get stable @id URLs so AI tools can resolve entities.
Conditional GET on llms surfaces. S3 sets ETag + Last-Modified; CloudFront passes them through. Agents re-polling llms.txt get 304 Not Modified when nothing has changed. Zero bytes, zero cost.
Site-wide Link response header. CloudFront advertises discovery entry points on every response. Agents can find llms.txt, agent-skills, and ai-policy without crawling.
IndexNow on every deploy. Pings Bing, Yandex, Seznam, and Naver — including the llms surface URLs explicitly, not just the HTML pages.
Per-article enrichment. wordCount, timeRequired (ISO 8601 at 225 WPM), and SpeakableSpecification for voice-readable slices. Glossary mentions inside articles auto-emit about edges to the canonical term — no per-post frontmatter needed.

Design principles

Type: Three faces, three jobs. -webkit-font-smoothing: antialiased throughout.

Aa

Literata Display · Serif
High-contrast serif for headlines.

Aa

Inter Body · Sans
A clean sans for sustained reading.

Aa

JetBrains Mono Mono · Code
For code and the footer terminal.
Colour: Two themes, swapped via the footer toggle. Single accent. SVG logos use fill="currentColor" so they theme-adapt without a second asset. Dark-mode ink is deliberately dimmed to #d8d4ce rather than near-white, to counteract the optical bolding (halation) that light text picks up on emissive screens.

Light

Surface#f9f8f6

Surface alt#eeebe4

Ink#2b3139

Muted#6b6359

Primary#3348a5

Dark

Surface#151515

Surface alt#1b1b1b

Ink#d8d4ce

Muted#b8b4b1

Primary#6970bf
Voice: Direct.

Operator-first.

British English.

Honest about what failed.

Specific over generic.

Same voice across the homepage, the 404 page, and the footer terminal.
Architecture: Single source → many surfaces. Edit once; everything downstream rebuilds together. No manual sync needed.

src/content/MDX · MD

src/data/TypeScript

HTML

JSON-LD

llms.txt

RSS

JSON Feed

OG cards

corpus.json

What's deliberately not on this site

No popups, exit-intents, or modals. The site never interrupts you to ask for an email.
No third-party tracking beyond Google Analytics. No Facebook pixel, no LinkedIn Insight tag, no session replay.
No comments. If you want to respond, write to me. I usually reply within a few business days.
No paywall, no gated content, no signup wall. Every essay, glossary term, and press entry is public on first visit.
No AI-generated content presented as written by me. When AI helps, it's editorial. Voice and judgment stay mine. Always.