Video structured data uses accurate segment semantics
Video cases no longer label the Prompt link as a video segment in structured data. Original prompts, copy actions, and source links remain available.
Experience
Directory landing pages reuse generated listings
Default case, creator, and Skills listings avoid repeated rendering. Search, filters, and pagination keep their existing results, and publishing or editing content refreshes the directories. Full Agent snapshots now identify content versions instead of changing solely because of a generation timestamp.
Experience
Copy web reproduction prompts with execution requirements
Web cases now offer a reproduction prompt that combines the unchanged original with runtime, asset checks, and acceptance requirements. The original and translations can still be copied separately. Copying does not generate a new work.
Experience
Content edits refresh the relevant case and creator pages
Title, translation and score edits refresh the relevant case and creator pages first. Cards embedded in other details update with their cache window or the next release; removals, merges and creator identity changes still refresh all related details.
Experience
Today's retest in the daily digest only shows the last three days
The retest slot used to show the most recent retest even when it was ten days old, which made it look like there was a new one every day. It now only features cases retested within the last three days and falls back to the review pick otherwise. Retests scored by judge models but not yet reviewed by a person now show their stability score.
Experience
Code-rendered videos are one click away from AI Coding (UI)
Code-rendered videos live under AI Video, so visitors browsing AI Coding (UI) rarely found them. The output-form row there now ends with a Code-rendered → link that jumps straight to them, keeping the current search, model and sort.
ExperienceOpen
Model pages are paginated, and the daily digest structured data lists only that day's cases
Model pages used to show only the newest 24 cases and sent everything else to a filtered library view that search engines cannot follow. Each model page now has a pagination bar at the bottom, and page 2 onward has its own address such as /models/veo/page/2, 24 cases per page, newest first, in both languages. Page 1 keeps its address, and the button that opens the library filtered by model is still there, below the pagination bar.
The daily digest structured data used to list all of the site's roughly 1,700 cases. It now lists only the few dozen that actually appear on that issue, picks first and new arrivals after.
Experience
Stronger internal links: jump-ahead pagination and more ways onward from case and creator pages
The library pagination now also lists jump pages such as 10, 20 and 30 next to the nearby pages, so any page of the 69-page list is at most 4 clicks away. On phones it still shows only the first, last and adjacent pages, so the bar does not wrap into several rows.
Case pages gain two blocks below related cases: more from the same creator and the latest cases on the same model, up to 6 each. Creator pages list all of that creator's cases after the 3 representative ones. The footer links to the model and creator pages, and every tab of the two home rankings is now written into the page itself.
Measured across the site's links, the median number of clicks from the home page to a case page drops from 4 to 3, and the deepest case goes from 38 clicks to 7.
ExperienceDataAPI
Titles drop the repeated model label, a Code-rendered video filter, and a dedicated Opus 5.5 model page
101 cases split from Douyin image sets all had titles shaped like Series Midjourney Prompt: scene, so five rows on the same leaderboard shared one prefix. They now read Series: scene, and the model name stays in the card tags and recommended models. Another 8 titles were just the opening of the prompt itself, up to 92 characters; they have been rewritten as one-line titles, and the publishing pipeline now rewrites both kinds automatically.
With AI Video selected, the generation-form row on the case library gains a Code-rendered option for videos written as code (single-page HTML, GSAP, Three.js) without any video generation model. The first 10 were all written by Claude Opus 5.5. The public API form parameter accepts code-generated-video.
Claude Opus 5.5 now has its own page at /models/claude-opus-5-5 instead of sharing the generic Claude page, and model chips on case pages link straight to it.
Skill detail pages now show Case evidence before creator versions, so the work comes first.
ExperienceOpen
GIF cards become looping video, out-of-range pages return a real 404
The 33 cases whose main media is a GIF now show a static first frame on the card, and only mount a muted looping MP4 once the page has loaded, the browser is idle, and the card is in the viewport; it pauses when the card leaves the viewport, and reduced-motion users only see the first frame. The original GIFs totaled 200.5MB; the MP4s are 6.1MB plus 0.9MB of first frames. On a list page with GIFs above the fold, such as page 6 of the web category, time to usable on Fast 4G dropped from 50.9s to 1.1s. Cases without a loop video still show the original GIF.
With 1649 published cases the last page of the library is page 69. Plain pagination URLs past the last page, such as /cases?page=70, now return a real 404 in both languages instead of 200 with noindex. The rule is generated at build time from the published count; out-of-range pages that also carry filter, search, or model parameters stay 200 with noindex.
The Code-Rendered Explainer Animation skill gained an admission check: when packages are generated, the full prompt is checked for actual code-rendering intent. The SpaceX Mars landing page case, whose prompt only describes visuals, was removed, taking the skill from 7 cases to 6.
The sister repo awesome-seedance now exports every video skill from the site each day, 21 official skills and 40 creator methods, each with author name and install command, plus a new list page. It used to count by scraping the /skills list page, which after the section split only showed 13 creator methods.
FeatureExperience
/skills splits official and creator skills, 7 new skills
The /skills list is now three sections: GoodCase official skills, creator methods, and Carl's toolbox. Official cards carry an orange GoodCase Official badge; creator methods carry an outlined badge naming their parent skill. Creator methods without a distinct style tag are no longer listed and live in a creator versions block on the parent skill's page, with links and packages unchanged. A new naming check blocks duplicates, so pairs like Pets and Animals as the Lead next to Some Creator · Pets and Animals as the Lead no longer appear.
Seven new skills take in cases that previously fit nowhere: cinematic drama shorts, stage performance and dance MVs, code-rendered explainer animation, cinematic stills, photo re-creation comparisons, interactive tools and simulations, and open-source hardware projects. General skills went from 27 to 34, with 94 installable packages.
Categories were tightened by task: product ads only match ad intent and no longer pull in lifestyle vlogs; animation consistency excludes code-rendered and character-free animation; landing pages now match whole words, shrinking from 199 cases to 61, with a separate section-level recreation spec skill split out. The product ad, POV / vlog, landing page, and animation consistency pages gained a how-to section (what it is for, what to prepare, steps, failure modes and limits, retest coverage) and 5 representative cases; the landing page notes that Claude has not been verified by retests.
Related cases at the bottom of a detail page now prefer the same task and would rather show fewer, instead of padding with same-model or alphabetical picks; when same-task matches run short, original-title similarity fills in. Separately, 64 cases that belonged to no skill were labeled by a model reading the prompt, reviewed one by one, and placed into existing skills. Together, cases with zero related picks fell from 630 to 166.
Experience
Detail-page prompts re-typeset, header stays put between pages, library covers lazy-load
Prompts on case detail pages are now 16px with 1.75 line height, cutting line length from about 158 characters to about 79. Long prompts show 12 lines by default and expand on demand, the copy bar sticks within view, desktop shows the prompt and copy button above the fold, and on mobile the media moved above the title.
Header and footer moved into a shared site shell, so they no longer vanish and reappear when you change pages; the top nav marks the current page with an orange line on paper, and the mobile sticky bar gained a menu. Text that may contain Chinese is now at least 12px, and standalone buttons have 44px touch targets.
The card entrance animation is gone. It made cards appear, hide, then fade in, at one point hid all 24 cards on a mobile page, and blocked keyboard access; without it, the cards reachable by Tab went from 2 to 24. Double lines between cards are now 1px, and paging jumps straight to the top.
Video covers on case cards are now lazy-loaded images that only download near the viewport and pick a size by screen density. Time to usable on /cases (p75) went from 1.76s to 1.10s on Fast 4G and from 4.33s to 2.21s on Slow 4G, and above-the-fold cover downloads on 1x screens dropped from 631KB to 258KB.
The homepage AI hot list grew from 8 items to 10. The Feishu group digest now counts cases published since the previous digest was sent, so a night like 10-03, when 161 cases went live at once and the next digest reported 0, will not repeat; the missed digest was re-sent to the group.
TransparencyAPIData
Retest notes spell out what scores mean, public API stops exposing pipeline tags, pagination signals fixed
The retest section gained three notes: when recommended and retest models appear side by side, it says this is a cross-model retest; it states that the retest score is the AI judges' average for a single generation and should not be read as a success rate; and the footer now says two or three AI judges according to the record's actual panel. Several representative cases also carry an editor-written retest reading.
The 30-second promo case made directly by Opus 5.5 was recreated by Claude Sonnet 5 from the same prompt and entered the retest line: Claude, GPT, and DeepSeek judges gave 70.8, 58.9, and 74.6, for a retest score of 68.1, moving the case's stability score from 0 to 68. For open briefs like a site promo, a different model will always produce different footage, so the judges no longer compare the result against the original cover; they judge consistency within the video and across aspect ratios, and the 9:16 version is shown to them as well.
The public /api/public/cases endpoint, detail-page tags, JSON-LD, and on-site search no longer include ingestion and review pipeline tags such as auto-approved or trial-x-socialdata; trust-related tags human-reviewed, not-rerun, verified-rerun, and reconstructed-prompt are still returned. Agents that filter by tag should take note.
English pages that showed Chinese titles and summaries received 609 English display texts, and English-source cases received 179 Chinese translations, bringing both gaps to 0. From page 2 on, library canonical and hreflang point to the page itself and invalid page numbers return a real 404; model tags on detail pages that map exactly to a model page now link there; and when the case list fails to load it no longer falls back to sample data, and the page carries noindex.
Two intake gates were tightened: Xiaohongshu leads must have over 1000 likes, include a prompt, and be posted within 7 days; and when a creator already has 3 similar published cases in the same category within 30 days, further candidates from that series are held back, so one creator's look-alike series cannot flood the list. Both apply only to new candidates; existing cases are untouched.
DataExperience
161 human-reviewed cases go live in one night, site visuals pulled back into the design system
On 10-03, 162 candidates were approved by hand in the operator console, making 163 waiting to publish. Before publishing, 17 unusable titles were rewritten, such as a model name plus colon plus the prompt's first sentence, a numbered title like Film Narrative Case 241802, English sentence fragments, or a whole prompt used as the title. That night 161 were published: 1 was blocked by dedup for having exactly the same prompt as a published case, and 1 was routed as a reproduction. Published cases reached 1649.
The publish line used to run only when that round of auto-review approved something; the 10-01 round approved 0, so human-approved candidates sat waiting too. Publishing now runs every round, and human-approved candidates go live at the next 07:00 run.
A cleanup pass removed design-system violations from the front end: three orange gradients, a radial background glow, soft shadows, frosted glass, and pastel fills are gone, the Magic Study entry changed from a gold rounded pill to a standard button, and creator avatars are square. Copy buttons, rankings, thumbnails, and small labels were merged into shared components, so the same element now looks the same on every page.
DataTransparency
Model detection gaps closed, video skills reach 18, Grok image retests cover 287 cases
Model detection on the X ingestion line moved from keyword substrings to ordered rules: it now catches Wan (Tongyi Wanxiang), which was missed entirely, GPT Image written with spaces or hyphens like GPT- image 2, and MiniMax H3 written as just H3; Seedance, GPT Image, and Midjourney prefer version numbers, so Seedance 2.5 and Seedance 2.0 now match their model pages. On September 15, 30 existing cases were relabeled under the new rules.
Recommended models for web cases now mean the models named by the original source plus the models a retest judged as reproducing it. Web sources rarely name a model, so under this rule 60 web cases that passed retests gained GPT-5.6 Sol; 235 web cases still have no model and will get one only after passing a retest.
Two new general video skills, food close-ups and mukbang ASMR, and gameplay capture with livestream overlays, took the count from 16 to 18.
A Grok Imagine image gap run covered 42 cases. After judging, 18 turned out to depend on a user-uploaded input image that the rules at the time missed, so their scores reflect the missing input and no stability score was written; the other 24 were written, with a mean of 78.2 and a median of 81.9. Of 318 published image cases, those with a valid Grok retest went from 263 to 287.
DataTransparency
Stability score becomes a per-model mean; the dumpling case gains an Opus 5.5 web retest
The homepage has always described the stability score as the mean of independent retests across models, but each retest round overwrote the previous score in the database, so other models' runs did not count. Scores are now grouped by retest model, each model contributes its latest conclusive result, and the mean is taken; failed generations and zero scores are excluded, a case scores 0 and shows a failed retest only when every run failed, and three-tier model-vote verdicts are left out. All three retest lines (image, web, video) write stability scores this way.
The interactive 3D dumpling feast web case gained a retest generated by Claude Opus 5.5 from the same prompt, bundled into a single HTML file and scored by two judges at 72.9 and 90.2, for a retest score of 81.6. The interaction probe clicked an empty link in the header instead of a dumpling, and the Claude judge docked interaction points for it; this is published as is. The case's stability score stays at 86.
Once published cases passed 1400, the full-table cache for case card data exceeded the 2MB per-entry limit and was silently skipped, so every request refetched about 2.9MB from the origin. It is now cached in chunks of 400 rows, and the reassembled result matches the old full table row for row.
Transparency
Both Dumpling Club builds move back under the original case's reproductions
The two Dumpling Club builds use the same prompt as the published interactive dumpling feast case once whitespace is normalized, so they now sit in that case's community reproductions section as Carl's reproductions and open as runnable 3D pages. The GLB the original prompt mentions was not provided with the case, so a substitute model was used, and that note stays; the generating model was not verified and is left blank.
The Lab collection entry added the day before was removed from the homepage, and the old collection page and old dumpling showcase now redirect to the original case's reproductions. The 3D portfolio and pelican SVG from the same batch have no confirmed parent case yet; their pages stay, with the claim of originality withdrawn and search indexing disabled.
Feature
General video skills grow from 5 to 16, creator method cards no longer share titles
The first batch added 4: early DV home video, epic fantasy and sci-fi spectacle, twist comedy skits, and 3D cartoon character shorts. Categories come from the sister repo awesome-seedance's hand classification of 463 Seedance cases, with at least 15 cases and 12 creators per category, and cases outside the matches were also checked one by one for false inclusions.
The second batch added 7: travel walks, pets and animals, vehicles, fashion portraits, horror and suspense, extreme sports, and storyboard grid to video. While checking false matches, broad words such as fitness, skincare, chicken, and sequence were dropped because they pulled in unrelated cases.
Creator method cards used to reuse the parent skill's title, so the same POV / Vlog Presence card could appear several times. Titles now use a style keyword extracted from the creator's prompts, for example AIwithkhan · Korean alley summer family DV, and all 33 cards are distinct. Across both batches, 68 packages passed end-to-end verification.
DataExperience
117 more Grok image retests, library Latest sort now uses the date a case was added
Grok Imagine retests were run for 117 more image cases, all generated successfully with evidence published. Of 262 published image cases, 252 now have Grok generation evidence; the remaining 10 lack the original input image and cannot be run. This batch only has a first-pass score from a single GPT judge (mean 80.8, median 82); a retest score is only produced once both judges have scored, so stability scores did not change.
The new evidence images briefly failed to load: the publisher probed the public URL before uploading, and Cloudflare cached that 404. It now checks storage directly and adds a content version to new URLs, and all 117 images read back correctly.
The library's Latest sort now uses the date a case was added to the site, so a newly added case with an older source post is no longer buried; card dates now read Added followed by the date, while detail pages still show the original post date. Retest evidence that has a generated result but no judge receipt yet shows AI score pending.
A small key below the homepage's main actions leads to Carl's Magic Study; it only navigates on click, and the homepage does not preload any 3D assets.
Data
New cases get Chinese titles and translations automatically, and no longer 404 after publishing
Between September 5 and 13, auto-published cases piled up 93 fallback titles in the form model name: first prompt sentence, and 191 English cases had no Chinese translation, so the morning digest opened to a screen of English. All of them were filled in on the night of the 13th; from the 14th the publish line generates titles and English-to-Chinese translations right after publishing, machine translation never overwrites a Chinese title already in place, and the cache is refreshed as soon as the writes finish.
Fixed new cases returning 404 after publishing: the rebuild triggered after publishing was skipped whenever the latest commit on main was a routine chore commit, so new cases only appeared after some later build. Production builds now always run.
Translation validation normalizes timecode notation before comparing, so correct translations such as 0-3秒 rendered as 0-3s are no longer rejected, while changed numbers or segment counts are still caught.
ExperienceFeature
Daily digest lists every new case, skills can be favorited, filters follow real counts
The daily digest gained a New today list after the two picks, showing up to 20 cases with a link to the library for the rest and a one-line empty state on days with nothing new. The homepage digest entry now reports the real count for the day.
Skills can be favorited: every card on /skills has a star in its top-right corner, and skill detail pages have a favorite button too. The favorites page shows cases and skills together with a type label in each card's top-left corner; favorites still live in your browser, and existing ones load as before.
In the library's form filter, Motion / 3D and 3D Games merged into Motion / 3D / Games; old links still open, and the public API's form values are unchanged. Category pills in the library and on /skills follow real counts, so categories with zero cases (AI copywriting at the time) no longer appear.
Homepage stats now link to the matching lists; the This week's new models row ranks by release freshness, case heat, and case count; and the installable skills section ranks by evidence-case heat plus on-site views over the last 7 days.
The model evaluation section on the connect page is now a single instruction you can paste whole into an agent, with one copy button; the changelog page gained a Feishu group QR code in the top right for getting updates first.
APIExperience
Public API q= is now relevance-ranked; search highlight and multi-word phrase bonus fixed
The q= parameter on /api/public/cases changed from filter-in-default-order to relevance-ranked, and now also matches source and prompt text (previously only title, summary, creator, recommended models, and tags). This is a public contract change: with q, items[0] is the best match rather than the hottest case, so agents that used items[0] as the top case should drop q for that; llms.txt, the Agent API docs, and the goodcase skill now say so.
Three on-site search fixes: hits on full-width letters or ligatures (characters whose length changes under normalization) no longer lose their highlight or shift the snippet; multi-word queries such as 可灵 运镜 apply the phrase bonus per token synonym group, so a Chinese-titled and an English-titled version of the same case score symmetrically; creator cards no longer show a leaderboard rank while a search query is active.
ExperienceTransparency
Skill cards get collage covers, media rehosting becomes a publish gate, share image shows real site stats
Every card on /skills now uses a collage of thumbnails from its top three evidence cases as the cover, card content dropped from eight layers to five, and the whole card is clickable. Opening a skill unfolds the cover, title, and summary in place into the detail page, with the same paired transition on the way back, all within the frozen design system (0 radius, 1px black border, single accent).
255 published cases had media hosted on X's CDN (twimg), which all four mainland-China probes failed to reach, so those cases never loaded there. The publish script now rehosts media to media.goodcase.ai as a mandatory step: a case is published only after rehosting succeeds, failures stay in the candidate pool with the reason recorded, and no more mainland-invisible cases are produced. Existing external links are being backfilled, after which a self-hosting constraint goes onto the database.
The homepage share image (the preview shown when a link is posted to social platforms) gained the booth-poster line "An AI case site where authors burn their own money to verify prompts actually work" and four real site stats (cases, creators, models, skills) sourced from the same data as the homepage footer; if the lookup fails the stats row is omitted entirely rather than printing sample numbers.
/connect and /agent-api gained a Model Evaluation section: an agent with the goodcase skill installed can sample cases by category, pull the on-site retest baseline, run the target model, and produce a comparison table, with curl examples and trigger phrases; llms.txt links to it.
Cross-model retest verdicts in the public awesome-seedance repo now follow human verdict > score threshold > status fallback. Batches without human review were previously counted as fully reproduced, so the README showed MiniMax H3 Max at 100%; threshold mapping gives 169 reproduced / 74 degraded, i.e. 69.5%, matching the on-site stability scores.
DataTransparency
Stability ranking expands: full H3 Max video retests and Grok image retest evidence published
All 324 text-to-video cases were retested with MiniMax H3 Max via fal.ai: 321 generated and scored by two judges (mean 77, median 80), stability scores written to production, every evidence video uploaded to our own storage, and the public retest manifest grew from 199 to 520 entries; 3 cases were rejected by the platform's content moderation.
Grok Imagine retest evidence for 136 of the 138 image cases was added to the public manifest (520 to 656); this batch previously had only scores and no published evidence. The remaining 2 are template fragments whose prompts are a few words long and cannot be retested.
Two retest-rule changes: a generation rejected by the model provider (content moderation or parameter validation) is now recorded as a definite failed retest with a score of 0, evidence level L2, and the failure reason published, instead of being left blank; video scoring gained an audio dimension (loudness, silence ratio, spectrogram, transcript when there is speech), with weights rebalanced to prompt adherence 35 / reference fidelity 25 / visual quality 15 / text detail 10 / audio 15, and older records can be re-judged under the new rubric while keeping their previous score.
The public awesome-seedance repo now syncs automatically, keyed on a content hash, instead of manual refreshes: the export drops 11 cases relayed from third parties, each case carries its last 5 retest records and a verdict summary, and the GitHub index collector runs on a schedule.
The daily auto-publish gate rose from 20 to 40 cases to work through the candidate backlog alongside the doubled auto-review quota.
DataTransparency
Near-duplicate prompt merging goes live with permanent redirects; full-table reads no longer silently cap at 1000
Cases with highly similar prompts are merged into one primary case, with the other members attached as external reproductions: clustering uses 4-gram similarity and containment, the primary prefers the original author and already-published cases, and merge pairs can be pinned by hand. The first merge batch has run in production; each merged slug gets a build-time 308 redirect to its primary for both the Chinese and English paths, and the detail page and public API check redirects before returning 404.
Ingestion dedup now catches the two link forms of the same tweet (x.com/i/status/ID and x.com/author/status/ID), the main source of earlier duplicate cases; 31 duplicate candidates were rejected outright.
Full-table case reads switched to pagination with a verified total: the database API returns at most 1000 rows per request by default, and with more than 900 published cases the sitemap, full listings, and retest scripts could silently drop entries. A mismatch between rows fetched and the authoritative count now raises an error instead of returning fewer cases.
Six component-library source adapters now extract a reusable method, so summaries for newly ingested cases no longer degrade into a generic template sentence; a cleanup script for the 98 existing fallback summaries is ready, replacing each with real wording drawn from its source.
ExperienceCommunity
Homepage rankings now match height, stability copy updated, new Join Us page
The Platform Heat Ranking and Stability Ranking cards on the homepage now match height: once the grid stretches both cards to the same height, the footer no longer trails the list with a stray gap underneath; the stability ranking's footer gained an explanation of its own scoring (what the stability score and the reproduced n/m count mean), mirroring the heat ranking's note.
The unmeasured-stability cell now reads "Vote for a paid retest" instead of "Vote to retest", making clear the vote is for someone to actually pay for and run the retest.
Top nav gained a Join Us page that gathers the three ways to take part: submit, feedback, and agent access. The submission page and the feedback form reserve a Feishu group invite block below them, live as soon as the group link is published.
FeatureAPI
Agent API gains category sampling and per-model retest lookup; coding cases get a Form filter
The public /api/public/cases endpoint accepts four new parameters, form, exclude, sample, and seed: filter or exclude by sub-form, draw up to 50 random cases, and reproduce the same draw from the same seed. A new /api/public/cases/{slug}/retests endpoint returns a case's stability score, retest records grouped by model, and the human review verdict. The goodcase skill adds a model-evaluation flow on top: when a new model ships, anyone with the skill installed can sample the same question set by category and compare results against on-site retests with the same yardstick.
The AI Coding category gets a second-level Form filter with three stackable tags, Web UI, Motion 3D, and Agent Tool, so landing-page templates, Three.js effects, and CLI tools no longer share one bucket; 323 of the 328 published coding cases were back-tagged, and collectors now tag at the source.
All 744 English-only cases now have Chinese translations, and the backfill script supports the en→zh direction; browsing an English-sourced case on the Chinese site no longer shows only the untranslated original.
New Douyin and Xiaohongshu keyword lead collection runs every three days using the platforms' own low-follower viral label, and the daily auto-review quota doubled to 40 judged / 20 approved, so the candidate backlog should clear within a week.
ExperienceData
Model tag on case cards, longer prompt preview, tighter Connect and About copy
Case cards now carry a model chip (the first recommended model), and the source-heat cell names its platform (e.g. "Source heat · X"), so you can tell at a glance which model made it and where the heat comes from.
The prompt preview on cards grew from two lines to four, truncation now measures visual width instead of character count so mixed Chinese/English prompts fill the box, and tweet lead-ins like "made with X, prompt 👇" are stripped on cards so the preview starts at the prompt itself.
The X collection line no longer writes the boilerplate "AI video case, full prompt shared…" summary; cards show the author's own post text instead, and a replacement script rewrites the 151 published cases and 106 candidates that still carry the template.
Connect page headline now reads "Bring GoodCase into your workflow" and all four command blocks have a copy button; the Open Source and About section titles fit on one line with their sub-copy removed, the feedback form is centered, and the changelog gained the 15 missing entries from Aug 14 to Sep 1.
DataTransparency
Source heat algorithm adjusted: velocity weight cut to 15%, completeness becomes a gate
Weights changed from interaction 60% + velocity 30% + completeness 10% to interaction 85% + velocity 15%: production data shows a 0.68 rank correlation between interaction volume and interaction velocity — since ingestion only ever captures one snapshot, velocity is largely a monotonic re-ranking of the same interaction signal, so the old weights double-counted it.
Data completeness no longer contributes to the score; it's now a hard gate at >=50 to be ranked. Completeness is computed against each platform's actually available fields — X does not expose a public save/bookmark count, so its completeness only counts likes, comments, and shares, and is no longer docked for a field it can never report.
When interaction velocity is missing, it no longer defaults to a neutral 50 — it now falls back to the item's own interaction percentile, and the source-heat note says so explicitly.
When a same-platform cohort has fewer than 20 scored cases, percentiles shrink toward the median (50) so platforms with very few samples (GitHub, Douyin) can't dominate or bottom out the ranking off a handful of coarse percentile jumps.
ExperienceTransparency
Case detail pages simplified, original Skills get real screenshots
Case detail pages dropped the AI-generated summary paragraph, the large author card, and the reusable-method panels, keeping only the original post's result, the prompt, and retest evidence — less AI-written text that could pass for verified fact.
Retest evidence switched from a stacked list to per-model tabs, making it easier to see which model reproduced a prompt and which didn't when comparing several.
Creator identity on the homepage and detail pages was scaled down to a single byline that still links to the creator page; the homepage's dedicated creator section and stats were removed to keep the focus on cases.
All six of Carl's original Skills now show real screenshots of the tool in action, replacing the earlier placeholder-cover notice.
FeatureData
New AI Hot Topics section on the homepage
Below Carl's banner, a new Today's AI Hot Topics section pulls from AI HOT's public rankings and hides itself automatically if the data hasn't refreshed in 72 hours, so the homepage never shows stale content.
Each headline links straight to its source, and hovering shows how many outlets are covering that topic.
FeatureData
Auto-publish pipeline and time-decayed current heat ranking
New candidates can now clear an automatic review channel: three models vote and a unanimous majority approve publishes directly, with anything ambiguous or short of votes still falling back to human review.
Case ranking now factors in time decay: new, actively spreading cases rank higher while older ones settle down naturally, kept at a 15% floor so a good older case never bottoms out to zero.
The homepage's two ranking boards relabeled their score from heat to current heat, to distinguish it from the source heat snapshot still shown on case cards and detail pages.
ExperienceData
Homepage and page-level feedback fixes
Homepage cover images are now deduplicated across sections, so the same picture no longer shows up in the hero, the rankings, the Skills section, and the creator section at once.
Case detail pages deduped the tag row and changed the evidence-level label from Evidence L1 to a plain Source verified; derived Skill cards gained a thumbnail from their source case.
The /skills page moved from category-grouped sections to a single flat grid with a cover image on every card, plus a new Carl's Originals filter for the six original Skills; the mobile product bar's three external links now match the desktop tab strip's color coding.
Fixed old favorites disappearing: the favorites page only fetched the newest 50 cases and filtered locally, so once the library passed 700 cases older favorites fell out of that window; it now looks up favorited cases directly instead.
ExperienceData
Homepage redesign feedback fixes
The homepage's Our Output case slot now shows the single highest-heat case in the whole library, instead of the first item after deduping the top-10-by-heat and top-10-by-stability sets; the matching creator slot follows that same case's actual author.
The Skills showcase now only shows Skills that are actually packaged and installable, hiding the section entirely when there's no real data instead of filling space with GitHub repo art.
Model boards with only one or two cases used to look sparse; they now pad out to five rows with high-heat cases that haven't been stability-tested yet, visually distinguished from tested rows.
The mobile homepage got a section-by-section pass: the hero headline shrinks on small screens, the output cards use a shorter aspect ratio, and the model tags and ranking tabs now scroll horizontally instead of wrapping into several lines.
FeatureExperience
Homepage redesign, phases one through four
The hero section gained an Our Output trio showing one real case, one Skill, and one creator side by side; the search box moved to the center of the top nav.
The ranking section became two switchable-tab boards: heat ranked by platform, and stability ranked by model, the latter now showing real retest reproduction rates.
Added a Trending this week section pulling the top six cases by real on-site view counts over the last 7 days.
Added a banner featuring Carl himself with a few highlight moments; the desktop top bar became a tab-strip layout, showing GoodCase.ai alongside its three sibling sites (AI News Radar, AI Map, LearnPrompt).
Open
Connect page gains an Open Source & Related section
The /connect page gained a section linking to GoodCase Lite (the open-source version of the site) and Awesome Seedance (a verified prompt library), for anyone who wants to run their own copy or borrow template prompts.
FeatureExperience
Daily digest gains AI intel, top bar added site-wide
The /daily page gained a Today's AI Intel section pulling the day's top picks from AI News Radar, hiding itself automatically if it hasn't refreshed in 36 hours.
A thin Projects bar now runs across the top of every page, linking to sibling sites AI News Radar, AI Map, and LearnPrompt.
The /skills page gained a dedicated Carl's Original Skills group with real install commands for six Skills (Laoyeye, Carl Skills, Partner Skill, and more), plus a progress bar tracking total GitHub stars across all of LearnPrompt's repos.
DataTransparency
GitHub index ingestion goes live, new /standards page
The case library added a GitHub-indexed source, only accepting entries with a known author, original prompt, publish date, and verifiable engagement; the new heat gate is strict enough that roughly one in six candidates clears it.
Added a public /standards page that writes out the inclusion criteria that previously only lived in internal docs.
ExperienceData
Homepage dedup, retest verdicts now decided by three-model vote
The homepage's top-3 hero cases, Hot this week, and the deep case pool used to overlap heavily and show the same handful of cases; the three sections are now mutually exclusive, and the deep case pool is topped back up to 24.
193 retest records that had sat with images but no verdict for over two weeks now get judged by three models voting reproduced, degraded, or failed, closing that gap.
Transparency
Case detail pages gain a provenance statement
Right below the full prompt on case detail pages, a new statement discloses whether that prompt has been checked against its original source post, with a link to verify it yourself; unverified cases are labeled as such instead of getting a misleading checkmark.
ExperienceInfrastructure
DSH announcement becomes a welcome modal, faster loading in mainland China
The DSH plugin ranking announcement, previously an easy-to-miss banner, is now a centered modal shown on first visit, and won't repeat once dismissed.
Media now routes through a mainland-China CDN, cutting measured time-to-first-byte from roughly 200ms down to the teens across several mainland cities.
DataTransparency
Case summaries and site search no longer leak source code
Some AI Coding (UI) cases had their summary field filled directly with component source code, which leaked into share cards and on-site search results; these now use a cleaned or fact-based fallback description instead.
The machine-readable full snapshot llms-full.txt used to silently drop to 3.4% of its content under load while still claiming to be complete; it now reads the full set reliably and states the real coverage in its header if anything is missing.
ExperienceInfrastructure
Homepage speed trio and a rolling new-in-24-hours badge
The three hero images no longer lazy-load, and the media host domains get a preconnect hint, so the first screen renders faster.
The homepage stats row gained a +N in the last 24 hours badge, computed on a rolling 24-hour window and hidden entirely when there's nothing new.
Experience
Fixed intermittent blank 404s on creator pages
Some creator pages would get treated as nonexistent and show a 404 whenever their data fetch failed, instead of a retryable error; failed fetches now surface a clear error state instead of being mistaken for a missing page.
BrandExperience
Site positioning updated to AI Prompt Case Collection
The site positioning changed from “AI Case Evidence Library” to “AI Prompt Case Collection”, applied across the header, footer, homepage sample label, share cards, and feeds in both languages.
Fixed blurry Chinese nav text on some devices: the monospace font stack now falls back to proper CJK fonts instead of the device default.
The header tagline and nav links switched from gray to black with larger sizes for better legibility on small and low-DPI screens.
Experience
Cards default back to full color, new scroll-reveal animation
Based on user feedback: card images are no longer grayscale by default with color only on hover — they now show full color right away, no more hovering each card one by one.
Case lists and the homepage deck gained a scroll-reveal entrance animation: cards fade and slide in as you scroll, replacing the old grayscale-to-color hover interaction; devices with reduced motion enabled skip the animation.
TransparencyData
Image, UI, and video retest outputs are now viewable
Case detail pages gain an independent retest-evidence section: generated images and videos are directly viewable, while UI retests include desktop and mobile captures plus a download of the generated HTML source.
Every retest file is explicitly labeled as a retest output, not the creator's original. Only the latest same-model result is expanded by default; earlier independent runs move into a numbered history to avoid looking like different models.
Legacy Codex image-generation records are grouped under GPT Image 2 in the UI, while original run IDs and model-receipt disclosures remain intact.
Model, test date, final score, and both AI judges' raw scores are public. Automated scores stay labeled as automated, and unverified generator identities are disclosed instead of implied.
The first public batch contains 181 retest records: 87 image, 89 UI, and 5 MiniMax H3 768p video records. Failed generations remain visible as no-artifact evidence.
FeatureExperience
Weekly hot, digest archive, and stability sample scores
New Hot this week on the homepage: the six most-engaged cases from recent source posts, with the subtitle stating the actual data window; max two slots per creator.
Daily digest back-issues: month-grouped archive plus prev/next paging across 28 issues; historical issues never show that day's retest block — retest evidence is only presented as of now.
Navigation reordered: Rankings comes first and jumps straight to the two homepage boards, Daily moves to the end; each board now links to the full sorted case list.
Search understands Chinese/English synonyms: 千问, 可灵, 智谱, 即梦 and more now match English-named cases; the public API benefits too.
Stability sample scores (BETA) rolling out: the first 81 image cases now carry sample-based stability scores, labeled separately from human retest verdicts; other categories follow.
ExperienceInfrastructure
Mainland mirror entry, faster edges, and share cards
New mainland-friendly entry goodcase.carlwow.com, same content as the main domain; measured good reachability across mainland China.
Edge caching overhauled: half the site used to recompute on every request; everything now serves from edge caches, cutting cold TTFB from ~2s to the 0.4s range.
Social share cards completed site-wide: case cards composite the actual artwork, creator and Skill pages no longer share without an image; fixed WeChat picking another case's image when sharing video cases.
Publishing a retest verdict now triggers a site rebuild automatically — new verdicts appear within minutes.
ExperienceTransparency
Author-side dates for creators and Skills
Creator cards and pages show a Latest work date — the newest source publish time among their published cases.
The Skill catalog and detail pages show a Latest example date from their supporting cases.
Only author-side dates are shown: they reflect ecosystem activity; editorial actions on our side never produce public timestamps.
FeatureAPI
Daily digest and Agent API launch; retest lab starts running
New daily digest (/daily): two cases a day — one fresh hit from the last 14 days and one worth revisiting. Selection is deterministic by date, so everyone sees the same pair, with a dedicated RSS feed.
Agent API formalized (/agent-api): the public API stays usable without a key, now with documented rate limits; integrators can request an API key for higher quotas. Every case response now carries a provenance field confirming the prompt was verified against its source post.
The retest lab is running: the first 10 image and web cases have been regenerated with current models for side-by-side comparison, prioritized by retest votes. Verdicts will appear after human review.
ExperienceMobile
Real-device fixes: full-color images on mobile, vote anywhere
Images on phones are no longer stuck in grayscale: the hover-to-color effect is kept only for devices with a pointer; touch screens show full color right away.
Retest votes can now be cast directly from any card, not just the detail page; a voted case shows its voted state everywhere.
Fixed the prompt-translation toggle: on Chinese-original cases, tapping the English translation used to do nothing; it now switches correctly and remembers your preference per interface language.
The install command on Skill pages no longer stretches mobile pages sideways; it scrolls within its own box and stays fully copyable.
The source-post like count on detail pages now matches the size and style of the neighboring action buttons.
ExperienceData
Self-hosted video playback and equal-height cards
All case videos and posters now live on our own storage — playback no longer depends on third-party CDNs; failed loads show a clear message with retry and a link to the source post.
32 cases had their prompts restored to the full original-post version (some had been truncated upstream; the longest grew from 253 to over 10,000 characters), with method breakdowns rewritten to match.
Cards in the same row are now strictly equal-height: title, summary, and tag zones have fixed heights, and prompt blocks and stat rows align across cards.
Cases awaiting a stability retest now show their current vote count.
FeaturesTransparency
Account-free likes and retest votes
Likes are now real site-wide counts instead of local-only numbers: no signup, one count per browser per case, visible across devices.
Cases without a measured stability score accept retest votes, with public running counts.
Detail pages show the source post's real like count alongside, clearly labeled as an ingestion-time snapshot, never mixed into site counts.
The public API gains real likedCount and retestVoteCount fields for agents.
DataTransparency
Provenance cleanup: 33 cases unpublished
Every case ingested via aggregator channels was checked against its original source; 33 whose prompts could not be matched to the source post (likely reverse-engineered, or the author never published a prompt) were unpublished.
Two cases were restored to the author's original Chinese prompt, keeping the previous English version as a translation.
The ingestion pipeline gained three provenance gates: aggregator reverse-engineering markers are preserved, flagged candidates no longer auto-enter review, and prompts are automatically matched against source-post text at ingestion.
One case with a publicly disputed attribution now carries a neutral note on its detail page.
ExperienceFeatures
Site-wide speedup and model browsing on production
Major page-weight cuts: the case list dropped from 4.5MB to about 180KB, creator pages from 1.1MB to about 97KB.
Model browsing launched: filter cases by model family (Seedance, GPT Image, and more), with a model strip on the home page.
Full bilingual support; switching languages keeps you on the current page.
Case and creator lists are paginated; card media adapts to aspect ratio, so vertical videos and UI screenshots are no longer cropped.
FeaturesTransparency
Creator methods and shared Skills derived from Cases
Creator credits on Case cards now link directly to creator pages, whose Case count covers all published work.
A shared Skill appears only after the same method is evidenced by at least three Cases across two creators.
A creator method requires at least three matching Cases by the same creator; source popularity is not part of the threshold.
Skills have shareable evidence pages but remain outside the primary navigation and are not a separate submission type.
OperationsOpen
Operator metrics and agent entry points
The operator page now summarizes anonymous visits, case activity, submissions, public evidence, and hidden historical data from the last 30 days.
Test feedback is explicitly marked and archived separately from real user suggestions.
Added /llms.txt so agents can discover the public API, RSS, Skill, and evidence-use rules from one address.
The public API and case library now support AI copy and AI hardware categories consistently.
ExperienceBrand
Agent access moved to the top and brand icons aligned
Agent Access is now the orange primary action in the header and home hero.
Changelog and Feedback moved into the top navigation; the footer now contains only the brand and page note.
Browser favicon, Apple icon, and the in-product thumbs-up mark now use one identity.
DataTransparency
The popularity ranking became a traceable Source Heat Ranking
Source Heat reads likes, comments, reposts, saves, and capture time from the original post instead of using a static popularity score.
Interactions use likes + 2×comments + 3×reposts + 4×saves; the final score combines within-platform interaction percentile, velocity, and data completeness.
Cross-platform ranking is normalized within each platform first; missing fields remain null rather than being penalized as zero.
Cases without a verifiable interaction snapshot can remain public but do not enter the ranking.
Features
Search, favorites, RSS, Connect, and share posters
The case library now supports keyword search across models, methods, and creators.
Favorites are stored locally in the browser and require no account.
Added an RSS feed at /feed.xml, with new cases delivered automatically.
Added the Connect page for Skill, RSS, and API access.
Every case can generate a QR-coded share poster.
Open
Open API and Skill launched
Released a free public API with anonymous access for case lists and case details.
Released the GoodCase Skill so AI assistants can search and compare cases in conversation.
Experience
Cases no longer require registration
Removed the login wall so every public case is readable without an account.
Full prompts and retest records are public; likes remain a local interest marker.
Launch
GoodCase launched
The first 12 real AI cases launched across video, coding, and image creation.
Each case surfaces source heat, stability, creators, prompts, and reproducibility signals.
Roadmap
Next
Stability scores will increasingly come from Lab retests instead of manual labels.
The submission workflow will continue to improve as more reproducible cases arrive.
