Roster mapping each role (researcher, curator, reviewer, copy editor,
picker, conductor) to its pipeline stage and canonical prompt file,
the shared house rules, the day-in-the-life commit flow, and the
copy editor's subagent brief verbatim — the one prompt not stored as
a skill; daily-pipeline stage 4 now references it as canonical.
Claude-Session: https://claude.ai/code/session_01WZaczDJjL3xZ3u5spsN5AL
One written standard for every audience-facing word: concise, to the
point, inspiring delight through specificity — with word budgets per
copy type (reviews had drifted from a ~370-word median to 500-1000+).
All writing skills now reference the charter, and daily-pipeline gains
a stage-4 copy-edit pass: a subagent that didn't write the day's copy
tightens it against the charter, prose only, diff reviewed before
commit. Digest gains a copy-edit line.
Claude-Session: https://claude.ai/code/session_01WZaczDJjL3xZ3u5spsN5AL
New rule at every selection stage: daily-finds drops such items at
capture (categorical, overrides design/gift appeal; borderline cases
skip too), build-bundle and cluster-bundles treat legacy alcohol
finds as ineligible for new bundles, and build-review refuses to
graduate them even on explicit slug. Existing finds stay published
per forward-only decision; CLAUDE.md records the policy centrally.
Claude-Session: https://claude.ai/code/session_01WZaczDJjL3xZ3u5spsN5AL
The real failure mode is 0-1 images (wall of text); two heroes read
fine on the 1-2 column layouts. Keep ~4 as the aim — prefer
hero-bearing candidates among equal fits, more images for bigger
bundles — since 2 among 8 items still looks thin on the wide
4-column grid. Restores the 3-item minimum bundle size.
Claude-Session: https://claude.ai/code/session_01WZaczDJjL3xZ3u5spsN5AL
Bundles now ship only when at least four of their items have a hero.*
image, purely so the grid never renders image-thin. Clusters that miss
the floor are held back and logged rather than padded or written thin;
the weekly catch widens its window to older unbundled hero-bearing
finds instead of holding. Capture gets a nudge to try a second image
source before giving up on a hero.
Claude-Session: https://claude.ai/code/session_01WZaczDJjL3xZ3u5spsN5AL
Rotated from herculaneum-scroll-first-read (held since 07-13).
The best hero image of the run is deliberately NOT the pick: Lai Tak Tsuen's
atrium is the most arresting thing we published today, but its residents have
posted notices asking photographers to stay out, and the homepage hero is this
site's loudest amplifier — pointing it there would generate exactly the traffic
they are trying to stop. Same shape as the rule against heroing anything that
reads as endorsing harm, just quieter. The review still runs; it just never
gets the tile.
The M0 Pura wins the rest on the axes the pick is judged by: a press shot that
shows the whole product in one frame (it is tiny, and there is nothing on it),
a virtue legible from the tile without domain knowledge, and nothing remotely
controversial about it.
Three graduations, and the fact-check earned its keep on every one.
Shanling M0 Pura (music): the 250mW balanced figure needs an adapter Shanling
sells separately — out of the box it is 100mW, on a single 3.5mm jack. The
'cannot stream' pitch holds (no Wi-Fi, no apps) but the review declines to sell
the myth: Bluetooth-receiver mode will happily play your phone's Spotify. Read
alongside the existing Zenteek review — the same argument with a battery in it.
Lunacy (play): runs the original unmodified After Dark module, not a
reimplementation — confirmed. Corrects the find's attribution of More After
Dark to Apple (it was Berkeley Systems) and disambiguates Jeff Halter, who made
Lunacy, from Ben Haller, who wrote Lunatic Fringe in 1991. Ships no game data;
you supply your own copy. Same developer as the Stacks HyperCard player, which
is the actual story.
Lai Tak Tsuen Estate (places — new category): the find was substantially wrong
and both it and this morning's bundle have been corrected in place. Not two
1975 towers but four cylinders across two twin-cylinder blocks, completed 1976.
And access is not 'a guard checks resident cards' — there are coded gates and
posted notices prohibiting entry and photography, put up because the Ghost in
the Shell shoot turned a residential estate into a backdrop. The bundle no
longer invites anyone through the gate, and this is emphatically not the pick.
Pool was at exactly 25 (the nominal cluster threshold) but yesterday's cluster
run already judged its remainder un-clusterable, so re-clustering would only
re-litigate that. Two tight 3-item themes instead, neither padded: three places
built for living/working rather than visiting (the run's one real patch to a
travel axis that has been thin for a week), and three things that argue for the
handmade page by being one — which finally surfaces london-notebook-writing,
shortlisted-but-invisible since 07-04.
Dedup was the story: 56 candidates surfaced, 31 already captured (the daily
feeds largely re-served yesterday's items), 2 previously rejected on merit,
5 failed the bar today.
Shortlisted 10. Two items were demoted off the shortlist by the fact-check
step rather than before it — the Coleman stove's Snow Peak IGT compatibility
is not a claim Coleman makes (and it is backorder-only), and Shortverse is a
filmmaker hosting platform, not the viewer-facing short-film streamer it
looked like. Both stay as finds, honestly written.
5 Amazon links. Source maintenance: Kevin Kelly demoted weekly -> as-needed
(publishes, but the Technium is essays, never things); The Browser's free feed
now carries only 2 of 5 daily picks; two Reddit 429s recorded but not acted on
(same-domain subs succeeded, so it reads as throttling, not a block).
- astro 7.0.7, @astrojs/mdx 7, sitemap 3.7, rss 4.0.19, check 0.9.9
- trailingSlash 'always' → 'ignore': Astro 7 generates prerender paths for
dynamic extension endpoints (/categories/<c>/rss.xml) with a trailing
slash under 'always' but builds their route patterns without one, so the
build fails with "Missing parameter"; directory-format output is
identical either way. Also removes the now-unneeded dev feed-slash shim.
- compressHTML: true pinned to keep pre-v7 whitespace behavior
- zod 4: import z from 'astro/zod' (astro:content re-export deprecated),
z.string().url() → z.url()
- npm audit fix (fast-uri, fast-xml-builder, yaml)
Claude-Session: https://claude.ai/code/session_01WZaczDJjL3xZ3u5spsN5AL
Fact-check against folo.is, its pricing page, and the GitHub README
corrected the license (AGPL-3.0, not GPL-3.0) and added mobile apps,
real pricing tiers, and the self-hosting caveat the find omitted.
Topics — the curated subscription layer: 18-value controlled vocabulary
in src/lib/topics.ts, enum-enforced by the content schema on finds and
reviews so unattended runs can't drift it. /topics/ index +
/topics/<t>/ pages with finds first-class alongside reviews, per-topic
RSS feeds that include finds (unlike category/tag feeds), "By topic"
section on /feeds/, Topics footer link and llms.txt section, dev-shim
coverage for the new feed URLs. daily-finds and build-review skills now
classify at capture/graduation time; "giftable" is a tag, never a
topic.
Per-source indexes: /sources/<slug>/ lists every review and find from a
source (catalog ∪ names in content, so retired sources resolve);
find/review colophons link "(more from this source)"; /sources/ cards
link "everything from this source →"; sub-3-item pages noindexed.
Claude-Session: https://claude.ai/code/session_01WZaczDJjL3xZ3u5spsN5AL
Fact-check against tintd.app and the App Store surfaced a version split:
the direct download is the real Tintd 2 (macOS 14+, one-time $9.99+),
while the Mac App Store listing is still the older v1.3.4 (macOS 13+,
subscription option) — Apple's review queue hasn't caught up yet.
864 finds + 64 reviews assigned 1-3 topics each from the new 18-topic
controlled vocabulary (src/lib/topics.ts, lands in the next site
commit). The retired gifts topic becomes a giftable tag on the 140
finds that carried it; 5 finds legitimately match no topic.
Claude-Session: https://claude.ai/code/session_01WZaczDJjL3xZ3u5spsN5AL
Fact-check against indieseek.co corrected the find's "no panel" claim
(there's a secondary status view) and added shortcut mechanics,
format limits, and pricing.
Tag pages are now generated from the union of entry and find tags —
1,179 tags that existed only on finds (most tag links on find pages)
previously 404'd. Tagged finds render in a compact secondary "Finds"
list below the entry grid, keeping the editorial tier on top. Find-only
tag pages with fewer than 3 finds are noindexed until they fill out.
Tag feeds stay entries-only; the RSS pill and autodiscovery link only
render where a feed exists. Category pages unchanged (finds have no
category).
Claude-Session: https://claude.ai/code/session_01WZaczDJjL3xZ3u5spsN5AL
With trailingSlash 'always', astro dev only matches dynamic-route
endpoints when the request has a trailing slash, so the RSS pill links
(/tags/<t>/rss.xml, /categories/<c>/rss.xml) 404'd locally while the
built site served them fine. Dev-only vite middleware rewrites those
requests to the slashed form.
Claude-Session: https://claude.ai/code/session_01WZaczDJjL3xZ3u5spsN5AL
Autodiscovery alone is invisible since browsers dropped feed-detection
UI — add a small RSS pill in the subheader (new feedHref prop on
BaseLayout) linking to the page's scoped feed.
Claude-Session: https://claude.ai/code/session_01WZaczDJjL3xZ3u5spsN5AL
Feeds: JSON Feed 1.1 at /feed.json, per-category and per-tag RSS at
/categories/<name>/rss.xml and /tags/<name>/rss.xml alongside the main
/rss.xml — all from a shared src/lib/feed.ts item source. Category/tag
pages advertise their scoped feed via autodiscovery; new /feeds/
directory page (footer-linked) lists every feed with counts; RSS feeds
gain an atom:link self reference and a feed.xsl stylesheet so they
render readably in browsers.
Programmatic readers: build-generated /llms.txt (site map with every
review/bundle/post), SearchAction on the WebSite JSON-LD wired to
/search/?q=, and a "For robots" section on /feeds/ documenting
search.json, llms.txt, the sitemap, and JSON-LD coverage.
Claude-Session: https://claude.ai/code/session_01WZaczDJjL3xZ3u5spsN5AL
The width-tier system (full-bleed wide tier, body max-width tweens,
absolute-positioned FLIP masonry) is gone. Grid pages are 4 columns in a
55.5rem container (>=69rem viewport), everything else 3, with the existing
1-2 column steps on small screens. Masonry returns to real flex columns.
This also fixes bundle pages rendering as a diagonal card cascade: the old
.masonry > .masonry-tile { position: absolute } tied with FindCard's scoped
position: relative on specificity, and on bundle pages the FindCard rule won
the cascade order, leaving tiles in normal flow with masonry translate()
offsets stacked on top. Flex columns don't position tiles, so the tie is gone.
Kept from the motion pass: staggered tile reveal on initial load and
"Load more" (now via animation-delay), card hover lifts, button tactility.
- Images/videos fade in on load instead of popping (cached media renders
instantly; nothing is hidden without JS)
- Cross-document view transitions: soft cross-fade between pages where
supported
- Masonry reveals ripple: initial grid and "Load more" batches stagger in
per-tile instead of blinking on at once
- Card hover: gentle -2px lift + slow image zoom on hero, home/grid tiles,
and find cards; find-card lift switched to `translate` so it composes
with the masonry's inline transform (it was silently dead before), and
its transition list unified with .masonry-tile's to end their cascade tie
- Tactile press/lift on "Load more" and the external CTA button
- Accent-tinted ::selection, visible :focus-visible rings, smooth anchor
scrolling; global prefers-reduced-motion catch-all disables it all
Claude-Session: https://claude.ai/code/session_01ADR9r5rw5NuvvnCxfYbMuc
Named width tiers (mobile/narrow/regular/wide) with body max-width easing
between them; grid pages opt in via BaseLayout `wide`. Masonry is now
absolute-positioned with FLIP glides on column-count changes: targets are
computed from the *final* container width (reading --max, which flips
instantly while max-width tweens), mid-glide column changes retarget
instead of being dropped, and container height eases so content below
rides along.
Claude-Session: https://claude.ai/code/session_01ADR9r5rw5NuvvnCxfYbMuc
rzen