feat(writing): markdown-backed writing, and the first piece — a quote collection

[IngTian]

The owner: "i want the writing to be flexible. so essentially writing.ts displays some markdown notes i have. My favorite quotes is just a markdown file."

So /writing is markdown-backed now, and the first piece is a quote collection. Writing a post is no longer a code change — drop a .md into src/content/writing/, give it frontmatter, and it appears on /writing under its section with its own page at /writing/<slug>.

The split

data/writing.ts keeps only the taxonomy — what a section is, what belongs in it, what to say while it's empty — because that's site voice, not content. The pieces are files. buildKinds() is the single seam between them, and it returns the same WritingKind shape the page and lib/pageStops already consumed, so the side rail, the sorting, the section ids and all their existing tests are untouched. Nothing downstream had to learn that content moved.

An Astro content collection, not import.meta.glob

The repo already globs images that way, so the alternative was familiar. The schema is the reason: frontmatter is validated at build time, so a misspelled kind: or a missing date: fails npm run build and names the file. The glob alternative renders a broken row instead. CI is the merge gate here, so that's the difference between catching it and shipping it.

It earned its keep on the first build: YAML parses an unquoted 2026-08-15 as a Date object, not a string, so z.string() rejected it. Requiring quotes would have worked and been a permanent trap — the error only appears at build time and reports a type mismatch rather than the missing quotes. The field now accepts both and normalises to YYYY-MM-DD, so either spelling is correct and downstream always gets one shape.

draft: true keeps unfinished work out of the build, filtered once at the source rather than in the template, so the rail, the count and the sections can't disagree about what exists.

Three things the owner caught, and two had one cause

The prose styles never applied at all. No spacing between paragraphs, and the quote didn't read as a quote. Both were the same bug: the style block was scoped, and Astro rewrites scoped selectors with a [data-astro-cid-…] attribute — but the elements <Content /> emits are generated from markdown at build time and carry no such attribute. .prose blockquote compiled to .prose[cid] blockquote[cid] and matched nothing. Every rule was dead CSS; what was on screen was the browser's default <blockquote>.

The repo's guide warns that this attribute changes specificity. Worth recording that it also changes what a selector can reach — markdown output is outside the component's scope entirely, so no amount of specificity would have saved it. Now is:global, with every rule nested under .prose so it can't leak. Verified in the built bundle rather than by eye: ships as .prose blockquote{…} with zero prose[data-astro-cid occurrences.

A quote now looks like a quote. Even with the CSS live, the first treatment was a size bump and a left rule — which is what a callout looks like. It's a pull-quote now: a hanging opening mark in ochre doing the announcing, the words in the display serif as the loudest thing in the column, and the source in a separate register beneath — mono, small, em-dashed, set to the column's right edge where the eye already arrives after the quote's last line. The attribution is styled by position (blockquote + p), never by a class, so the file stays a plain .md with nothing to remember and still renders correctly anywhere else.

"Quotes" became "Miscellaneous." Owner: "having only one quote doc under quotes is weird." Right, and the fault is scale rather than taxonomy — a section named for a genre holding one file reads as an unfinished shelf. Named for the leftovers it holds one honestly, and will take the reading log and talk notes later without a rename. If quotes outgrow it, promoting them is a data edit. The schema enum, the taxonomy key, the frontmatter and the drift test all moved together, which is what that test exists to force.

A test I did not weaken

tests/writing.test.ts requires every section's empty-state line to exceed 40 characters — the rule that an empty section must name what's coming rather than say "coming soon". My first draft for the new section was literally "Nothing here yet.", which is exactly the filler that rule exists to catch. The copy changed, not the test.

New specs

tests/writingContent.test.ts (6 cases) covers the new seam, including a drift guard that parses the schema's kind enum out of content.config.ts and asserts it matches the taxonomy's keys. Both failure modes it covers are silent: a piece whose kind no section declares never renders, and a mis-grouped entry appears under the wrong heading. Neither throws.

The Carney citation has no source link, deliberately

The moment is corroborated three ways — his statement and Q&A on the collapsed Canada–U.S. trade negotiations, 22 August 2026 (NYT that day: "Carney Slams U.S.-Canada Trade Proposal and Vows Retaliation", described as a Saturday morning address; Canadian Press: "PM Carney explains why Canada left the U.S. negotiating table"; and an American Rhetoric transcript of the Q&A). But I could not retrieve the transcript text to confirm the wording: this session has no web search tool, and both sources carrying it refused the fetcher.

So the citation names the event and stops there. On a page whose entire value is exactness, a plausible-looking URL nobody checked is the one unrecoverable mistake. The words are the owner's transcription with "all of a sudden" and "house in order" set in their canonical forms, and they want checking against the audio before this is treated as settled.

Gates

npm run build clean; 819 tests across 46 files green (was 813/45). No browser pass — the owner asked for no more headless browsers after an earlier fleet took their machine down — so layout was checked by reading the built CSS and HTML, and by the owner's own eye on the served build.

🤖 Generated with Claude Code