Changelog

Shared links kept showing the old artwork. The picture has a new address now

Dragging the changelog rail could catch a dot from the hidden one

Every site in the library now previews live, at both breakpoints

Nine sites were stuck showing a screenshot for no reason. Seven are live again

Removed seven sites that render blank or will not load at all

Opening a site loaded it twice. Now it loads once

An off-screen copy of the panel could take down the preview you were watching

Colours are drawn as one palette, each band sized by how much the site uses it

Type specimens lead with the face, a pangram and its character set

A specimen now says when it is showing a fallback instead of the real font

A font that never loaded left its specimen almost invisible

A specimen no longer claims a fallback when the real face is available

A slow site gets longer to arrive before the preview gives up on it

Specimens no longer sit near-invisible while their font is downloading

Live previews crashed on almost every site the moment they loaded

Previews now run on their own domain, walled off from the library

A site that uses local storage no longer breaks its own preview

A hidden copy of the panel could tear down the preview you were watching

Live previews gave up after three seconds on almost every site

A script error inside a page no longer counts as a page that failed

The mobile preview showed a screenshot instead of the live site

A capture now stands in only when the live page will not load

Removed 31 sites that would not render in the preview panel

Every remaining card previews live or opens straight to its capture

Ten sites were only failing because the audit tested them all at once

Most sites showed “This page couldn’t load” in the preview. That was our bug, not theirs

Others loaded but stayed blank, waiting on a request the preview was blocking

Previews weigh 86% less. A capture is sized for the panel, not shipped whole

Opening the same site twice no longer refetches the entire page

Drag the changelog rail to scrub through releases. The date keeps up

Scrolling back up left the rail sitting on a release you had already passed

Arrow keys step through releases one at a time

Dropped the Figma capture pipeline. It produced a capture for none of 278 sites

A nightly job recaptured the whole library to fill a column nothing ever wrote

Dropped the Excel importer, the Mobbin seed list, and the share-image fallback

The library is filed by what kind of site it is, not by its customer’s industry

hex.inc, a brand and product studio, had been filed under E-commerce

Product, Studio, Editorial, Company, Portfolio, Store, Venue, Event — in fixed order, not by size

Three quarters of the shelf used to sit in SaaS or uncategorised

Every site now has a real mobile capture, shot at phone width from the start

Mobile captures were desktop pages squeezed to 390px, not the mobile layout

Six sites showed their own share image in place of a screenshot

Twelve mobile captures were blank, or empty files that loaded nothing

Sites the live preview can never reach now open straight to their capture

Two sites whose domains stopped resolving have been removed

Dark mode sits on a new surface ladder. Background, cards and popovers step apart

Muted text on dark was too faint to read. Every level is lighter now

Card hover is a fill, not a hard border. Softer in both themes

Sidebar categories sit tighter together

New artwork when the site is shared, replacing a card from March

The site answered on two addresses at once. www now redirects to hitmanslibrary.xyz

The site read as not secure. One card stored its image over http, and a single insecure image drops the padlock for the whole page. Scraped URLs are upgraded to https on the way out now, so it cannot come back

Muted text was too faint to read. The lightest level measured 1.7:1 on white; every level now clears 4.5:1 in both themes

Type scale up one step across the board — the smallest labels were 10px

The rule above the corner links stopped in mid-air on About and Changelog. It follows the sidebar on the gallery and spans the page where there is none

A mark for the library — three books on a shelf, one leaning

The mark is the loading screen. Books arrive onto the shelf, hold, clear, repeat

Right-click the logo — or long-press it on a phone — to copy or download it as SVG

Favicon, touch icon and app icon are generated from the same geometry the site draws

Request a site, About and Changelog were all hidden below 640px — a phone reached the gallery and nothing else

About and Changelog sit in the bottom-left corner of the window, at every width. They list the pages you are not on

A changelog minimap — one dot per release, grouped by month. In the gutter on desktop, under the nav on mobile

Month markers on the changelog timeline, so a jump lands somewhere marked

sort=top could be set from a bookmark but never cleared

An About page — how a spreadsheet of links became this

Both signatures draw themselves when they scroll into view

Hover the word spreadsheet to see the original sheet

Changelog bylines show faces. Each line keeps an initial

Detail panel reads Preview / Mobile / Colors / Type. Mobile is a tab now, not a toolbar sitting over the preview

Live link moved into the panel header, beside the domain

Sort moved beside the result count. Categories select one at a time

Presentation mode and the filter chip row removed

An empty screenshot retries at viewport size; the shared browser reconnects after a crash instead of poisoning every capture after it

Duplicate detection ignores trailing slashes and query strings

The selected sort pill was near-white on near-white in dark mode — 1.02:1, unreadable

Live previews gave up on any failed image. Only real script errors count now

Capture now waits for a page to finish painting itself rather than to finish loading. joyco.studio was a flat blue panel with a logo on it, therawmaterials.com a cream rectangle and one hairline, datum.xyz an empty white page with a nav bar — all three had complete data behind them. The test is whether the frame is still getting busier: content arriving pushes it up, a preloader holds it flat and low, and two samples without a rise means the page has arrived. therawmaterials climbs through three stages over four seconds before it settles

How busy a frame is gets measured by asking the JPEG encoder, not by decoding the image. A flat colour costs almost nothing to store and a painted page costs a great deal, so the length of the buffer is the answer already — which is what makes this affordable to do four or five times a second inside a serverless function

A capture that comes back nearly empty is now retaken once. Walking a page can restart an intro, and a scroll-triggered reveal may not have run at all, so the second attempt is compared against the first and the busier frame wins. A site that really is that sparse keeps the picture it had

A tool that re-photographs the sources whose stored capture is a picture of nothing much, judging a replacement on two measures: how much is in the frame overall, and how much of the first screenful is given over to a single flat colour. The second is what catches a hero caught mid-transition — busy further down the page, blank across the top, which the first measure scores as perfectly healthy

Fourteen cards replaced. joyco.studio, datum.xyz, shuttle.zip, therawmaterials.com, eandco, naotofukasawa and lyannetonk were blank or splash frames; variant.ai, payjustnow and waabi.ai were caught mid-transition. One site, occupied.unadsgn.tw, has stopped answering altogether — its host resolves and then never responds, which is now recorded rather than left looking like a capture that failed

Extraction now refuses to write when a page renders nothing readable. newterritory.studio fell over behind a Kirby error page served with a 200, and a backfill run captured that error page and filed it as the studio’s design — because from the pipeline’s point of view it was a page, and it loaded. No typography, no assets and almost no colour is not a site with an unusual design; it is a page that did not render. The previous capture is kept and the reason recorded, since a stale picture of the real site beats a fresh picture of somebody’s stack trace. New Territory’s card has been put back

Spotify Design was filed under design.spotify.com, which does not resolve and never has. The entry had no screenshot, which meant it was not in the gallery at all — it existed only as a row. Repointed at spotify.design, where it extracts 12 colours, both typefaces and 43 assets

Assets no longer decide whether a site counts as broken. A canvas game and a WebGL studio site have nothing in the document to extract, and no number of re-attempts changes that — three of them were being queued nightly forever on the strength of a panel they can never fill. A site is complete when it can be seen, read for colour and read for type; an empty assets tab now explains itself rather than being treated as a failure

Four sites added this morning, before the extraction fixes landed, were still holding what the old pipeline gave them. Re-running the backfill recovered Raycast (16 colours, 82 assets), Aside (14 colours, 72 assets), illoca (6 colours, 26 assets) and Sleep Well Creatives, whose screenshot had never been captured. 283 of the 284 sites in the library now render every panel they can

Background images were only ever looked for on a guessed list of tags — sections, headers, anything with class="hero". That list matches inline style attributes and little else, so a site that art-directs properly, with its imagery in a stylesheet, read as having no imagery at all. The document is now walked. cameronsworld.net is built from 335 CSS backgrounds and had been reporting zero

Some sites keep their entire document inside a frame, and everything the extractor does was reading the top document only. Same-origin child frames are now read as well — including the about:blank kind, which inherits its parent origin and gets filled by script after load, and which had been skipped on the assumption it was empty. Going looking only happens when the main document comes up thin, so ordinary pages are never trawled and nobody else’s advertising ends up filed as a site’s design language

Screenshots clamped their height against Chrome’s pixel ceiling but never their width, and the ceiling applies to both. A page measuring 11520 wide blew the budget no matter how short the clip was, and came back empty every time. Width is now clamped too — and a document much wider than its own viewport is treated as overflow rather than layout, because a rogue absolutely-positioned element or a carousel measured at full length is not what the page looks like to a visitor

Three sites with complete colour, typography and asset data were missing from the gallery entirely. The gallery only shows sources that have a screenshot, but the check for damaged sites only counted the three panels — so a site with everything except a picture passed every test while being invisible. The screenshot is now part of what counts as complete, which is what it always was to anyone looking

Request a site from the header, or from the moment it actually occurs to you — search for something that is not here, and the empty result offers to request the thing you just searched for, rather than making you type it a second time somewhere else

The address resolves as you type it. If the site is already in the library the form stops being a form and becomes a link to it. If somebody has already asked, it says how many. If it is genuinely new, it shows you the site — name, preview image, favicon — so you can confirm it is the one you meant before asking

stripe.com, https://www.Stripe.com/, and stripe.com/?utm_source=twitter are one site, not three. Addresses are reduced to a common form before anything is compared, so the duplicate check works on what people actually type and two people asking for the same site increments a count instead of filing a second request

Your browser remembers what you asked for and checks back on it. There are no accounts, so without this the flow would end at "thanks" and you would never learn whether it worked. Reopen the form later and your requests are listed as pending, passed, or added — and if it was added, there is a link straight to it

A review queue in admin, ordered by how many people asked. Approving runs the same extraction the manual add form uses — one path into the library rather than two that drift apart. A failed approval leaves the request pending, because that is something to retry rather than something to lose. Passing on a site remembers the decision, but a fresh request revives it

Rate limiting, which the site had none of anywhere. It lives in the database rather than in memory, because each serverless instance has its own memory and loses it on every cold start — a counter held there would reset constantly and be trivially outrun. It also runs before an address is resolved rather than after, so a stream of invalid addresses cannot spend DNS lookups without spending quota. If the check itself fails it lets the request through: a limiter that takes the whole feature down with a database hiccup is the worse failure

Typography was being written to a database column that does not exist. The insert raised an error every single time, and the error was caught and thrown away, so the failure was completely silent. Seventy-eight sites hold no typography at all because of one wrong column name

Extraction gave a page fifteen seconds to go completely quiet on the network before it would read anything, and if that never happened it threw the whole attempt away. A site with analytics, a chat widget, or an open websocket never goes quiet — so the timeout fired, and everything fell back to scraping colours out of raw HTML. That produced one unusable row and no type or assets, which is exactly what 55 sites were left holding. Getting the page is now what has to succeed; a quiet network is waited for briefly and then stopped caring about

Assets were being read from a page that was moving. The asset scan ran at the same moment as the screenshot capture, and the screenshot capture scrolls — so images that load lazily were still empty when they were measured and got discarded as tracking pixels, and the logo detector, which looks for something near the top of the page, was looking at a viewport that had already scrolled past the header. The page is now brought to a complete stop first, once, and everything reads from there

Sites behind Cloudflare were written off before the browser was ever asked. A plain request runs first to collect the title and preview image, and a 403 from it ended the entire attempt — even though the headless browser renders those sites perfectly well. Stripe, Wise, Mercury, Revolut and Airwallex had all been sitting in the library with nothing extracted for this reason

Any page whose source contained the word "blocked" anywhere in it — a CSS class, a variable name, a sentence of copy — was rejected as a bot wall. Real challenge pages are now identified by the machinery they actually carry

Tall pages came back with no screenshot. The height was capped at 12,000 pixels, but captures run at double density, so Chrome was being asked for 24,000 — past the limit where it silently returns an empty image instead of an error. The cap is now derived from the density actually in use

One unreadable colour value discarded an entire palette. The colour parser does not always fail politely — it can throw from inside itself — and that happened outside the code guarding it. A site with 136 extracted colours was recorded as having none

Extraction failures have never actually been recorded. The statement that writes the reason gave the database no way to work out what type the message was, so it raised every time, and the error was silently discarded. The explanation shown when a site has no data was never being written in the first place

The most common "brand colour" in the entire library was #0000EE — the blue Chrome paints unstyled links. Sites that never styled their links were having a browser default filed as their brand

A failure in one extractor no longer takes the others down. Colours, typography and assets are each caught separately, so a site with awkward assets still yields its palette and its type

Heading fonts are now found by looking for the largest visible text on the page rather than the first h1 — which missed every site that builds headings out of divs, and every site keeping a hidden h1 for search engines in a font nobody sees. Body font is whichever family covers the most of the page, not whatever the first paragraph happened to use. And a monospace font now has to prove it is monospaced by measuring itself, which stopped Stripe being recorded as setting its code in Söhne

Assets now include CSS background images, the largest image in a srcset rather than whichever one the browser picked, and anything inside a shadow root — which is most component-driven sites of the last few years. Logo detection gained two more ways to find a mark

The sidebar was the least designed surface on the page — a plain list of text with a 2px mark beside whichever row was selected, which read as a list with a pointer next to it rather than a row that was on. Selected rows now fill: a solid surface, full ink, one weight up. Losing the mark also frees the leading edge, so labels start where the padding says they do instead of after a rule holding its place whether or not it was visible. Rows sit taller, the gaps between them close up so the column reads as one thing, and the counts share a fixed right-aligned column so the digits line up down the list

The type scale runs from 10px to 32px instead of from 10px to 14px. Five of its seven tiers used to sit inside a four-pixel range, which is close enough that a card title and a category label read as the same voice. There is now a display tier, a heading tier, and a size meant for actually reading prose

The admin screens had been re-declaring the type scale by hand — forty-two separate hand-written sizes that duplicated tiers already in the system, and three that had no tier at all

Waiting for webfonts before reading type. Fonts routinely arrive after a page first paints, and reading before they land reports the fallback — which is how sites with real typography were recorded as using Arial

A backfill tool that re-runs extraction across every site the pipeline previously failed on, and marks anything still unrecoverable with an honest reason rather than leaving it looking merely empty

Clicking a card did nothing unless you hit the title text exactly. The whole card is meant to be one target, carried by an invisible layer stretched across it — but pressing any control scaled it slightly, and that scale made the title button the anchor for its own overlay, shrinking the target to the size of the title mid-press. The mouse then lifted somewhere the target no longer was, so the click dissolved. The opt-out meant to prevent this had been written as a weaker rule than the one it was overriding, so it never applied

A tablet was the worst place to browse the library. The three-pane split opened at 768px, which left the grid 384px to work with and cut cards to 158px wide — titles ran six lines and no two cards were the same height. The split now waits for 1280px, so everything below it gets the mobile layout with more room rather than the desktop one crushed into it. A small laptop at 1024px went from two 222px cards — narrower than a phone gets — to two at 484px

Card titles carried two instructions that both set how the element displays, and the wrong one won, so the one-line clamp never applied at all. Every title now truncates as intended, which is what finally made card heights uniform across a row

The industry label was told to truncate and also told it could not shrink, which are contradictory — so instead of an ellipsis it overflowed and got sliced mid-word at the card edge. It reads "FINANCE", not "FINANC"

Every control in the header was smaller than a fingertip: sort buttons 23px tall, the changelog link 17px. All of them now carry a 44px touch target while looking exactly as they did — the target grew, the design did not

A skip link. With up to 128 cards loaded and two stops on each, reaching the grid by keyboard meant tabbing past everything else first

Near-black colour swatches were invisible in dark mode, sitting on an almost-black background behind a hairline too faint to separate them. They have a real ring now

The preloader was tuned for a two-column grid that no longer exists — it was fetching images the browser then reported as unused, while missing the one that actually anchored the page load. It now follows the layout it really has

Nothing in the app confirmed a press. All 27 controls now scale to 0.97 on press over 140ms — the single biggest reason the interface read as unresponsive, no matter how fast it actually was

Cards are too large to scale without looking cheap, so pressing one lights its border instead. The press registers through the whole card, wherever you hit it

No hover state was gated to fine pointers, so tapping on a phone left cards zoomed and buttons lit long after you had moved on. Movement-bearing hover is now pointer-gated

One easing curve was doing every job. Split into three: the signature curve for entering and leaving, symmetric acceleration for movement across the screen, and a plainer curve for hover and press that never overshoots

Reduced motion was flattening every transition to 0.01ms, including the colour and opacity changes that tell you something responded. It now neutralises movement only, and keeps the feedback

Keyboard-driven actions no longer animate: arrow keys through presentation mode and the detail tabs are repeated constantly, and animating them made every step feel like it lagged the key. Pointer and swipe navigation keep the motion

The detail tab underline moved from 300ms to 200ms on a movement curve — it is clicked often enough that the old timing read as drag

Two places used transition-all, animating layout properties off the GPU

Two quick category clicks both branched from the same URL snapshot, so the second silently dropped the first — picking SaaS then Finance left you with only Finance. Consecutive edits now compose

Filtering left you where you were scrolled, so a new result set opened halfway down and read as though nothing had changed. Changing a filter returns you to the top; opening a site still does not move the page

The sidebar All row was counting the filtered results, so selecting Finance made it read "All 13"

Selected categories now carry a rule on the leading edge and go to full ink, instead of a background wash you had to look for. The rule holds its slot when inactive, so nothing shifts

A filter bar above the grid shows everything currently applied — categories, tags and the search term as removable chips, with a result count and Clear all. Applied state used to be split between the sidebar, the header and a separate tag list

Result sets crossfade with the grid height held, rather than the whole grid dropping to 40% opacity and the page collapsing between sets

The favicon was v0's logo — the SVG shipped in the template was theirs, not ours. Rebuilt as the library's own hash mark, legible at 16px. Also dropped leftover v0 authorship metadata

Admin was a client-side illusion: the gate was a sessionStorage flag, so every write route — delete, extract, re-extract, bulk import, deduplicate, Figma backfill, mobile capture, element capture — was callable by anyone with curl. All of them now verify an HMAC-signed httpOnly session cookie server-side, with a bearer-token path for scripts

The link proxy validated only that a URL started with http — it would happily fetch cloud metadata endpoints and internal hosts and hand back the response. It now resolves DNS and rejects loopback, link-local, private, CGNAT and multicast ranges, re-checking on every redirect hop

The gallery is server-rendered: the grid, categories and counts arrive as HTML. Crawlers previously saw 141 characters of text on a library of 199 sites

Every view has a URL — category, tag, search, sort and the open site all live in the query string, so views are shareable and the back button works. A ?site= link opens that site even when it sits outside the first page of results

robots.txt, sitemap.xml and a web manifest, none of which existed; SVG favicon instead of a 28KB JPEG; filtered views are marked noindex so they stop competing with the index

Tag filtering never worked — clicking a tag refetched with the tag silently dropped, so nothing changed. Tags now filter properly, stack with categories, and show as removable chips

Titles were raw <title> tags: HTML entities rendered literally ("International Tax &amp; Legal") and brands repeated the domain right below them. Entities are decoded and echoed brand suffixes trimmed, but only when what remains still describes the page

Broken thumbnails: some stored URLs kept their HTML entities and 404d; those are decoded on read. A new scripts/repair-screenshots.mjs finds sites whose screenshot uploaded as zero bytes (5 of 199) and re-extracts them

Root cause of those empty screenshots: Chrome returns an empty buffer rather than an error when a page is too tall to encode, and the result was uploaded unchecked. Captures are now clamped to 12000px and an empty buffer is never uploaded. Three of the five recovered by falling back to the site OG image; the two that resist capture entirely (sarvam.ai, vap.studio) show their domain, as designed

Search no longer replaces the grid with 32 skeletons on every keystroke — results stay put and dim while refetching, and out-of-order responses can no longer overwrite newer ones

The preloader was a flat 2.1s block unrelated to loading, and it server-rendered on every route, so /changelog shipped a blank overlay reading 000. It now races the page with a 900ms ceiling, is client-only, and sits out entirely for reduced-motion visitors

Cards: the visit link was invisible until hover, so touch users could never reach it — now always visible. Removed a link nested inside a role="button", and the five colour swatches announce as one palette instead of five hex codes

List and category responses are cached at the edge instead of no-store on every request; sort is its own labelled row on mobile rather than sharing a scroll strip with categories

Escape now closes the detail panel at every breakpoint, and the mobile focus trap no longer runs against a hidden sheet on desktop

Foundation tokens: ink levels (100/62/40/24) replace eleven ad-hoc text opacities; edge weights (70/50/30) replace four border treatments; 4px radius everywhere; one signature easing with four tokenized durations

Type scale: micro 10px / meta 11px / ui 12px / body 13px / title 14px — nothing in the UI renders below 10px anymore (specimen footers were 7.5px)

Roomier inspector: desktop split is now sidebar 2 / gallery 6 / panel 4 — specimens, palettes, and asset grids finally breathe; cards are wider at typical desktop widths (2-up, 3-up on very large screens)

Press / to focus search (keycap hint in the field); panel load failures now offer Retry

Panel tabs: sliding active underline (shared layout animation), legible 10px labels

One spinner, one copy-feedback pattern (in-place check morph, no layout shift — export buttons no longer jump), one section-label treatment, one focus ring

Calmer motion: scale-pop removed from panel content, card hover eased to 1.015 on the signature curve, all durations tokenized

Skeleton cards now match real card geometry exactly; mobile filter pills returned to library materials (mono, tabular counts); sheet respects safe-area

Presentation mode: white-alpha steps mirroring the ink system, keycap hints for pointer users, tokenized progress motion

Type specimens: fade-in-up keyframe was never defined in globals — specimens could render invisible; now defined

viewport themeColor now matches the real background tokens (#f7f7f5 / #0e0e0e)

Removed dead components theme-toggle.tsx and typography-display.tsx; duplicate Google Fonts link on specimens

Design cards: subtler hover scale (1.02×), visit link replaced with a blurred pill that reveals on hover, metadata section has cleaner hierarchy — title in tight tracking, domain in 40% opacity mono, industry in uppercase caps

Header title: "Hitman's Library" split so the apostrophe-s renders in font-light — intentional weight contrast, not a typo

Sidebar: active category now uses bg-muted/70 instead of full bg-foreground inversion — less aggressive, more editorial

Panel tabs: labels now uppercase with wider letter-spacing, 1.5px border indicator, slightly smaller icons — tighter and more refined

Panel header: hostname uses tighter tracking, industry/tag shows as uppercase badge below

Empty panel state: stale "figma" removed from hint text, type treatment more restrained

Globals: 2.2% SVG fractal noise grain overlay on body — gives all surfaces a material, warm quality instead of flat digital feel

Skeleton cards match the updated 3px border-radius of real cards

Removed Figma tab entirely — panel-tabs, site-detail-panel updated; figma-tab.tsx deleted

Preview tab loading state: clean domain name + three staggered pulsing dots — no screenshot background while the iframe loads. What you see while loading matches what you'll see when it's ready

Preview tab proxy-failed state: domain name + "Live preview unavailable" label + "Open site ↗" CTA. Honest about what's happening instead of silently showing a static screenshot as if it were live

Presentation mode proxy-failed state: subtle dot-grid background, large domain name, industry label, color palette swatches, and a "Visit site ↗" link. Blocked sites now feel like reference cards, not broken frames

Design card selected state: top accent line (2px) appears when a card is active — stronger visual signal that pairs with the brighter border

Design card color swatches are slightly larger (14px) for better readability at a glance

Empty panel state: "Select a site" now shows the list of available tabs (preview · colors · type · assets · figma) as a quiet visual hint of what's inside

Layers button in the Figma tab — opens the site in a capture window using Figma's own publicly-hosted capture.js script (the same tool that powers their AI integrations). Click "Copy to clipboard" in the toolbar that appears, then ⌘V in Figma to get fully editable text, shapes, and layout — not a rasterized PNG

Element selection in capture mode — the capture toolbar lets you select any component (nav, card, pricing section) instead of capturing the full page, so you can pull only what you need into Figma

Presentation mode shows the site thumbnail immediately while the proxy iframe loads — navigating between sites now feels instant instead of flashing to black

Presentation mode: added always-visible prev/next buttons to the bottom HUD — arrows are now immediately discoverable instead of requiring hover over the preview area

Sidebar active category now has a left border accent — clearer visual signal for which filter is selected, especially when multiple industries are active

Panel tabs: active tab label is now font-medium — stronger differentiation from inactive tabs at a glance

Added sandbox attribute to all proxy iframes — the root cause of "Application error" crashes. The proxy serves external sites at the hitmanslibrary.xyz origin, so their JavaScript ran same-origin, letting any unhandled JS error in those sites fire window.onerror on the parent and trigger the Next.js error page. Sandboxing the iframes without allow-same-origin isolates errors to the iframe while keeping postMessage (used for proxy-failed signals and the Figma picker) fully functional.

Presentation mode: minimal overlay HUD — controls float at top-right and fade in after a beat so the first thing you see is the site, not chrome

Presentation mode: nav arrows appear only on hover over the preview area, keeping the experience immersive when you're just watching

Presentation mode: thin animated progress bar at the very top shows your position across the full collection at a glance

Presentation mode: bottom HUD shows domain, industry, counter, and color palette in one compact row instead of two separate top/bottom bars

Presentation mode: info row slides in the direction of navigation — left when going back, right when going forward — giving a physical sense of movement

Presentation mode: swipe left/right to navigate on touch devices — works on any phone or tablet without tapping arrows

Header icon buttons increased from 32px to 36px — easier to tap accurately on touch screens

Mobile category filter pills increased from 28px to 40px height — were consistently too small to tap without mis-tapping adjacent pill

Sidebar filter rows increased to 36px — matches minimum recommended touch target for dense lists

Detail panel action buttons (re-extract, close) increased from 32px to 36px

Panel tabs increased from 40px to 44px height — meets Apple HIG minimum touch target recommendation

Color copy button expanded to 28px minimum — was previously a 12px hit target from a 4px padding box

Type specimen action icons (copy, Google Fonts link) increased from 24px to 32px

Figma tab now appears in the detail panel — it was fully implemented but never wired into the panel tabs

Replaced transition-all on all buttons and links with transition-colors — eliminates browser compositing cost across all CSS properties on every click target

Sidebar and mobile filter "All" count now shows the real total (e.g. 177) instead of the current page load (32)

Removed invalid workflow key from next.config.mjs that was generating startup warnings on every dev server boot

Created missing UI components: Label, Select, Badge, Tabs, ScrollArea, Textarea, Slider — many node and panel components were importing these non-existent files

Created missing lib/node-utils module — all node components importing getStatusColor() were silently broken at runtime

Node components now use correct @xyflow/react v12 NodeProps generic — NodeProps<FlowNode<DataType>> instead of NodeProps<DataType>

fetch() timeout option removed — not part of RequestInit; replaced with AbortController signal for proper 8s timeout on URL extraction

motion/react Variants type — spring transition type is now a string literal (as const) so it satisfies AnimationGeneratorType

unknown → ReactNode errors in 6 node files — output conditions now use !! to guarantee boolean before &&

Missing CheckCircle import in excel-parser node

prompt-node onFocus handler accepts SyntheticEvent instead of MouseEvent

Implicit any on Slider onValueChange callbacks — destructured value now typed as number[]

Implicit any on forEach callbacks in browser-extraction.ts — all font string arrays now typed explicitly

Removed unused DesignBrowser component — eliminates dead code and reduces bundle

Type specimen sample text and glyphs now render at correct opacity — was using hsl() around an oklch token which browsers silently ignored

Design cards are now keyboard-navigable — Tab to reach, Enter or Space to open the detail panel

Status colors (success, error, loading) now use design tokens — consistent across nodes, type specimens, figma tab, and changelog

Consolidated to single icon library (@phosphor-icons) — removed lucide-react dependency

Search inputs now have accessible labels for screen readers via aria-label

Filter and sort buttons now announce their active/inactive state to screen readers via aria-pressed

Color swatch copy button now announces its action and value to screen readers

Re-extract button uses aria-label instead of title — now accessible to screen readers and keyboard users

Mobile detail sheet upgraded to native <dialog> with showModal() — gains proper focus trapping, Escape to close, and system backdrop

Presentation mode screenshot alt text now describes the site by name, not just domain

Extracted shared NodeStatus component — eliminates duplicated status icon logic across all node components

Google Fonts stylesheet now only injected once per URL — prevents duplicate network requests when multiple type specimens share a font

Theme toggle button now announces the destination mode ("Switch to dark mode") rather than just "Toggle theme"

Preloader is now hidden from screen readers via aria-hidden — decorative counter is not meaningful content

Preloader setTimeout calls now properly cleaned up on unmount — eliminates potential state updates after component is removed

Color swatches on gallery cards now have role="img" and aria-label with the hex value for screen readers

Replaced all spring/bounce easing (0.34,1.56,0.64,1) with ease-out-expo (0.22,1,0.36,1) — animations feel faster and more intentional

Extracted DesignCard component and getDomain helper to their own files — reduces page.tsx by ~130 lines

Assets tab checkerboard background now uses design tokens — adapts correctly to light and dark mode

Preview tab screenshot image now uses loading="lazy" — deferred until the panel is actually viewed

Re-extract and close icon buttons in the detail panel increased from 28px to 32px for easier touch activation

Package name corrected from "ai-agent-builder" to "hitmans-library"

Figma tab — copy Desktop or Mobile screenshot as a PNG image, paste directly into Figma with ⌘V — no html.to.design plugin required

Sort pills in header — New / Old / A–Z / Top (quality)

Multi-select industry filters — hold multiple categories at once

Card hover reveals a visit ↗ button without opening the detail panel

Tags shown on gallery cards and in the detail panel header

Detail panel hostname is now a direct link to the site

All 177 thumbnails now use captured blob screenshots — eliminates blank cards from rate-limited external screenshot services

Sidebar category counts now match exactly what the gallery shows

Font 404 errors — removed dead @font-face declarations for missing SuisseIntl files

Extraction error UI in Preview tab was never visible due to an impossible condition

Re-extract could crash Colors/Type/Assets tabs by not normalizing null arrays

Quality sort silently fell back to "newest" — now correctly sorts by extracted quality score

SVG assets tab had an XSS vector from inline rendering of raw DB content

Breakpoint selector in Figma tab — switch between Auto / 390 / 768 / 1440px and see the site respond live

Full-page copy — one click captures the entire page as Figma layers, not just a single element

Client-side exception on some sites — API error responses now handled gracefully, no more crash on load

Site cards not visible — gallery query was referencing non-existent DB columns, now uses safe correlated subqueries

Retina screenshots — desktop now captured at 2× DPR, mobile at 3×, with lazy-image scroll-through before capture

Card image hover — CSS transition replaces motion.img for better performance and no layout shift

Element picker — hover any element to highlight it, click to copy directly to Figma

Full-page Figma layers pre-loaded on tab open — copy is instant, no wait

HTML proxy — loads any site inside the panel regardless of X-Frame-Options restrictions

Live preview — embedded iframe replaces static screenshots entirely

Figma tab — dedicated panel tab with auto-capture on first open

Search input in header — search across all sites in real time

Panel header redesigned — compact icon row for visit, re-extract, and close

Preview simplified — responsive iframe with no desktop/mobile toggle needed

Filter effect dep used JSON.stringify on every render — replaced with stable primitives

Copy to Figma — paste any captured site as editable layers directly into Figma

Mobile screenshots alongside desktop — captured at 390×844 during extraction

Desktop / Mobile toggle in the preview tab

Palette export — copy colors as CSS custom properties or Tailwind config in one click

Bulk add — paste multiple URLs at once and watch them process in a live queue

Preloader — 000→100 counter on first visit

HEX and OKLCH values are now separately copyable from the colors tab

Full gallery layout redesign — cleaner grid, tighter type, better dark mode

Subtle sound effects on copy and successful extraction

Admin CMS — passcode protection and one-click duplicate removal

Admin page was unscrollable due to a global overflow:hidden — resolved

Skeleton loading on first page load and category filtering

HEX / OKLCH format toggle on the colors tab

Dark / light theme toggle — instant, no flash of wrong theme

Category labels rewritten to better reflect actual design context

Mobile sheet — safe-area insets and 44px touch targets throughout

Prevented scroll chaining between preview panel and page body on mobile

Edge case: achromatic and scientific-notation OKLCH values now parse correctly

Live website screenshots displayed on every design card

Switched to a reliable headless screenshot pipeline after testing multiple providers

Copy feedback on color swatches — animated toast on click

OG image and social meta for sharing the library

Microinteractions across the UI following Emil Kowalski's principles

Color palette extraction — brand colors with OKLCH values

Typography extraction — font families, roles, and Google Fonts links

Asset catalog — logos, icons, images discovered per site

Industry categorization for every captured site

Browse and search across your entire design library

Excel import for bulk seeding the library