v0.2.2

Research ledger

Working document behind the form-and-force skill. Every row records what was read or observed, how it was verified, and which skill file it shaped. Sections A to G were researched on 2026-10-22. Section H was added on 2026-11-10 with the 0.2.0 method.

Verification labels

  • Verified: read directly from the primary source, or measured with audit_page.mjs on the live site.
  • Inferred: a reasonable reading of evidence that was not stated outright, or that rests on screenshots rather than interaction.
  • Disputed: sources disagree, or the claim depends on the author's taste.

Source types: official spec · technical docs · award-gallery entry · studio case study · critical essay · independent analysis · opinion · measured (our own capture).

Contents

  • A. Awwwards and award-caliber work
  • B. Anti-slop and generic AI design
  • C. Agent Skills and packaging
  • D. Interaction and motion engineering
  • E. Accessibility and usability
  • F. Performance and quality
  • G. Measured corpus observations (own captures)
  • H. The creative method (0.2.0)

A. Awwwards and award-caliber work

#QuestionSourceTypeDateFindingStatusInformsCaveats
A1How are sites scored?Awwwards evaluationofficiallive 2026-10Design 40%, Usability 30%, Creativity 20%, Content 10%. At least 18 jurors; the 3 scores furthest from the mean are dropped. HM at jury score ≥ 6.5. SOTD goes to the highest scores, with no fixed cutoff. A Developer Award needs a dev-jury score > 7. SOTM is the top 8 of the month, re-judged.Verifiedjury.md (background only: the jury marks ten dimensions in words and does not use these weights as a score)The page names the criteria but does not define them. Definitions in third-party blogs (utsubo, hontran.dev) are unofficial.
A2What does the developer jury score?SOTD pages, e.g. Lando Norris and Igloo Incaward-gallery2024–2026Six dev sub-scores: Semantics/SEO, Animations/Transitions, Accessibility, WPO, Responsive Design, Markup/Meta-data.Verifiedjury.md §6 (dimensions 9 and 10), floor.md §2—
A3Where do award winners score low?15 SOTD pages with dev scores (table below)award-gallery2023–2026Accessibility is the lowest or tied-lowest dev sub-score on 14 of 15 (range 6.4–7.4). The exception is Igloo Inc, where Markup 6.4 is below Accessibility 6.6. Animations/Transitions is the highest or tied-highest on 13 of 15 (8.2–9.6). The exceptions are Dropbox Brand and Son Daven, where Responsive is highest.Verified (counted from the table)SKILL.md stance: match winners on motion, beat them on accessibilitySmall sample; the jury applies the rubric unevenly.
A4Which sites are recent standouts?Sites of the Month listaward-galleryto Sep 2026SOTY 2025 is Lando Norris (OFF+BRAND), SOTY 2024 is Igloo Inc (abeto). SOTMs from 2024–2026 span B2B logistics, hospitality, brand guidelines, dev tools, furniture and film.VerifiedCorpus selectionThe listing was missing some months.
A5Are award categories broader than "immersive"?SOTD pages for Terminal Industries, Floema, Dropbox Brand, Anime.jsaward-gallery2025–2026Business, corporate and e-commerce sites win when the concept fits the domain. Terminal Industries (yard logistics) and Floema (urban furniture) are both SOTM/SOTD.Verifiedreferences/project-modes.md—
A6Do award sites persist?DNS check on takingshape.adobe.com and vogue-talents-sept2021.monogrid.io; Wayback Machine snapshotsmeasured2026-10-22, re-checked 2026-10-23Both campaign sites from the seed list no longer resolve. Their archived copies hold only the JavaScript app shell, so the experiences cannot be replayed. Studio and Awwwards pages with stills remain (see SOURCE_STATUS.md).Verifiedresearch/RESEARCH_GAPS.md; project-modes.md (campaign lifespan and archiving)Archives do not preserve WebGL experiences.
A7What did the Deep Beauty experience do?Awwwards element page; MUDEC; Ogilvy work page; AQuest case page (HTTP 403)gallery + institution + agency2025; re-checked 2026-10-23Element tags "scroll" and "tunnel". Ogilvy confirms that AQuest built an immersive online version of the exhibition. Per secondary summaries only: free 3D navigation through rooms, 60+ artworks, a brush-like cursor, and sound.Inferred from secondary evidence (no experience URL found)Corpus §22, marked "not analysed"The AQuest page blocks automated access (403 on 2026-10-22 and 2026-10-23); MUDEC returned 403 to an automated client on 2026-10-23. Needs a manual browser check.

A3 source data: Awwwards developer sub-scores (from each SOTD page, retrieved 2026-10-22)

SiteSemantics/SEOAnimationsAccessibilityWPOResponsiveMarkup
Adobe Taking Shape7.409.006.807.207.807.80
House of Dreamers7.008.206.408.007.207.00
Half of Eight8.008.207.007.607.607.60
Lando Norris7.408.607.007.607.407.40
Igloo Inc6.609.606.608.008.406.40
Shopify Editions Winter '267.809.407.008.407.807.60
Dropbox Brand7.008.407.007.608.608.00
Anime.js7.409.007.407.808.007.40
Terminal Industries7.808.807.008.207.807.40
Oryzo AI7.608.607.008.607.807.20
Why Zero6.608.406.608.007.006.80
Floema7.808.207.007.607.607.60
Siena Film Foundation7.208.606.607.208.007.20
Bruno Simon7.008.606.608.208.206.80
Son Daven8.208.207.408.008.408.00

Bold marks the highest value per row and italics the lowest (ties included).

B. Anti-slop and generic AI design

#QuestionSourceTypeDateFindingStatusInformsCaveats
B1What makes a site look AI-generated?InterfaceKit: What makes a website look AI-generatedindependent analysispub 2026-02-26, upd 2026-09-07Five tests: substitution, decoration-off, second-screen drift, uncomfortable states, and product language. The verdict should rest on three criteria together (generic, unresolved, unreviewed), never on single style choices.Verifiedfloor.md (failure classes, substitution and removal tests)The publisher sells UI kits.
B2Why do AI sites converge?InterfaceKit: Why AI sites look the sameindependent analysispub 2026-02-12, upd 2026-08-17Main cause: underspecified briefs ("clean", "premium") that the model fills with familiar defaults. Contributing: template and framework defaults, iterative "make it pop" prompting, and no review against real content.Verified (the source calls it an editorial model, not measured)direction.md §2 (facts before ideas)—
B3Which surface tells recur?Snoddy: Why AI websites look the sameopinion2026-07-13Violet-to-indigo gradients, Inter, glass cards, centered hero plus three cards, stock shadcn blocks, rocket/bolt/check icons. Cause: models sample the high-probability centre.Verified (opinion)slop_scan.mjs rulesAuthor sells a scoring tool.
B4Does the "tasteful" alternative converge too?Morphic: Why AI websites look the sameindependent analysis (promotional)202617 generated homepages fell into about 5 creative families: cream/sand/stone, serif display with sans body, hairline rules. Scores were 4/5 cosmetic variation, 2/5 strategic, 1.5/5 typographic.Verified (editorial judgement, not statistics)Second-order defaults in floor.mdPromotional source with a small n.
B5What does Anthropic's own frontend guidance list?anthropics/skills frontend-designofficial (vendor skill)current mainA calibration list of defaults: warm cream with serif and terracotta; near-black with one acid accent; broadsheet hairlines; the SaaS card kit; template chrome (tracked caps eyebrows, middle-dot meta, spaced em-dash labels, → on links, mono labels). Also: "Spend your boldness in one place," and one orchestrated motion moment beats scattered fade-ups.Verifiedfloor.md (second-order list); motion-direction.mdCompatible with this skill; nothing here contradicts it.
B6How did earlier vendor guidance shape outputs?Claude cookbook: prompting for frontend aestheticsofficial2025-10-21Names "on distribution" convergence: Inter/Roboto/Arial, purple gradients on white. It recommended layered gradients and "atmosphere", and noted the model still converges on new defaults such as Space Grotesk.VerifiedCore insight: every anti-default list breeds the next default, so the skill derives decisions from the subject instead of from listsThe cookbook's "layer gradients" advice is what B5 now flags.
B7Award winners have their own convergenceOwn captures (G-section)measured2026-10-22Near-black with an acid-lime accent appears on Lando Norris (#D2FF00), Terminal Industries and Why Zero (#01C654). Floating pill nav, Lenis (6/21), preloaders and gates (House of Dreamers, Siena, Why Zero, Bruno Simon) are common.Verified (palette from computed styles and screenshots)"Award-site defaults" in floor.mdLando's lime is his real racing livery, so it is not a default there.

C. Agent Skills and packaging

#QuestionSourceTypeDateFindingStatusInformsCaveats
C1Metadata rulesAgent Skills overviewofficiallive 2026-10name: ≤ 64 characters, lowercase letters, digits and hyphens; no XML; must not contain "anthropic" or "claude". description: non-empty, ≤ 1024 characters, no XML, says what the skill does and when to use it. Three-level progressive disclosure (~100 tokens / < 5k / on demand).VerifiedSKILL.md frontmatter; tooling/validate_skill.py—
C2Authoring rulesBest practicesofficiallive 2026-10Body under 500 lines. References one level deep from SKILL.md. A table of contents in any reference over 100 lines. Third-person description. Checklists and feedback loops. Scripts handle errors themselves. No time-sensitive facts. Forward slashes. At least 3 evals.VerifiedWhole package structure; validator checks—
C3Open spec constraintsagentskills.io specificationofficial speclive 2026-10name must match the parent directory, with no leading, trailing or consecutive hyphens. Optional fields: license, compatibility (≤ 500 chars), metadata (string → string), allowed-tools (experimental). Conventional dirs: scripts/, references/, assets/. Validator: skills-ref validate.VerifiedDirectory name = form-and-force; metadata values quoted as strings—
C4Claude Code specificsClaude Code skills docsofficiallive 2026-10Locations: ~/.claude/skills, .claude/skills, plugin skills/. ${CLAUDE_SKILL_DIR} substitution. Description plus when_to_use truncated at 1,536 chars. Outside Claude Code, only name, description, license, compatibility, metadata and allowed-tools are allowed (others hard-fail on upload).VerifiedFrontmatter kept to portable fields—
C5Plugin distributionCreate a pluginofficiallive 2026-10.claude-plugin/plugin.json holds only the manifest; skills go in skills/<name>/SKILL.md. Test with --plugin-dir; validate with claude plugin validate.VerifiedRepo layout; README install steps—
C6Behavioural evalsPlugin evalsofficiallive 2026-10evals/<case>/prompt.md plus graders/*.md (regex, tool_used, tool_order, file_exists, llm, baseline). claude plugin eval . runs with- and without-plugin arms. Runs are sandboxed; scaffolds need --scaffold.Verifiedevals/ suiteEvery run is billed.
C7Skill security postureAnthropic: Agent Skillsofficial2025-10-16 (upd 2025-12-18: open standard)Audit bundled scripts; be wary of skills that fetch external content.VerifiedScripts contact only the URLs you pass them, never read secret files (inspect_repo lists their names), and slop_scan reports secret matches by file and line without the value. Regression tests cover both—

D. Interaction and motion engineering

#QuestionSourceTypeDateFindingStatusInformsCaveats
D1Which properties are cheap to animate?web.dev animations guidetechnical docsupd 2020-10-06Animate transform and opacity. Avoid layout- and paint-triggering properties. Do not apply will-change early: set it from JS just before a change and remove it afterwards. Layers cost memory.Verifiedmotion-craft.md §13; slop_scan layout-anim ruleOlder page, but the rendering model is unchanged.
D2GSAP in ReactGSAP React guidevendor docslive 2026-10useGSAP() from @gsap/react reverts animations, ScrollTriggers, Draggables and SplitText through gsap.context(). It takes scope, dependencies and revertOnUpdate; wrap event-time animations in contextSafe(); it handles Strict Mode double-invocation; add "use client" in the App Router.Verifiedmotion-craft.md—
D3GSAP licenceGSAP pricingvendorlive 2026-10"100% free for all users", including the plugins, supported by Webflow.Verified (page states it is free; licence text not reviewed)Library choice tableRead the licence terms for unusual uses.
D4ScrollTrigger pitfallsScrollTrigger docsvendor docslive 2026-10Create triggers in page order (or use refreshPriority). Never animate the pinned element itself. Do not put ScrollTriggers on child tweens of a timeline. content-visibility: auto breaks measurement. gsap.matchMedia() replaces ScrollTrigger.matchMedia(). Register the plugin or tree-shaking drops it. anticipatePin: 1 for big pins.Verifiedmotion-craft.md; slop_scan rule—
D5LenisLenis READMEvendor docslive 2026-10The package is lenis. Options include autoRaf, lerp (default 0.1), syncTouch off by default, and respectReducedMotion on by default. data-lenis-prevent stops smoothing on nested scrollers. GSAP sync: lenis.on('scroll', ScrollTrigger.update) plus gsap.ticker. Limits: no CSS scroll-snap, 60 fps cap in Safari, no smoothing over iframes.Verifiedmotion-craft.md (smooth-scroll policy)—
D6three.js disposalthree.js manual: How to dispose of objectsvendor docsrepo dev branchResources are never freed automatically. Call dispose() on geometry, material, texture, render target and skeleton; call ImageBitmap.close() yourself; scene.remove does not free anything; controls and renderer cannot be reused after dispose(); check renderer.info.Verifiedcreative-coding.md §7 (WebGL lifecycle); slop_scan rule—
D7View TransitionsMDN startViewTransition + BCD 8.1.7official docs + dataBCD 2026-10-22Same-document view transitions are Baseline 2025 (Chrome 111, Firefox 144, Safari 18). Cross-document @view-transition ships in Chrome 126 and Safari 18.2 but not Firefox.Verifiedmotion-craft.md (progressive enhancement)Re-check support at use time.
D8Scroll-driven animationsMDN + BCD 8.1.7official data2026-10-22animation-timeline/scroll()/view()/animation-range: Chrome 115, Safari 26, Firefox 159. animation-trigger is Chrome 146 only and experimental.Verifiedmotion-craft.md (native-first option, @supports fallback)Very recent cross-browser support; older Safari and Firefox remain in the wild.
D9Other modern primitivesBCD 8.1.7official data2026-10-22@starting-style (Chrome 117, Firefox 129, Safari 17.5); popover; <dialog>; inert; text-wrap: balance everywhere, pretty not in Firefox; interpolate-size Chromium only; anchor positioning now in Chrome, Firefox 147 and Safari 26.Verifiedmotion-craft.md, component-architecture.mdVolatile: the skill says "check support" instead of hard-coding versions.

E. Accessibility and usability

#QuestionSourceTypeDateFindingStatusInformsCaveats
E1Animation from interactionsWCAG 2.2 Understanding 2.3.3officialliveLevel AAA. Motion triggered by interaction can be disabled unless it is essential. Parallax is the canonical example. Scrolling the page itself is essential; added scroll animation is not. Techniques C39 and SCR40 use prefers-reduced-motion. Auto-playing motion falls under 2.2.2.Verifiedaccessibility.md; motion-direction.md §11It is AAA, but the skill treats it as a default floor for large motion.
E2What does reduce mean?MDN prefers-reduced-motionofficial docsliveIt means "removes, reduces, or replaces" motion, not "removes everything". The MDN example replaces a scale pulse with an opacity dissolve. Baseline widely available since 2020.Verified"Replace, don't delete" rule—
E3New in WCAG 2.2W3C: What's new in 2.2officiallive2.4.11 Focus Not Obscured (AA), 2.5.7 Dragging Movements (AA), 2.5.8 Target Size Minimum 24×24 (AA), 3.2.6 Consistent Help (A), 3.3.7 Redundant Entry (A), 3.3.8 Accessible Authentication (AA). 4.1.1 Parsing was removed.Verifiedaccessibility.md; audit_page checks (target size, focus obscured)—
E4Interaction patternsARIA APG patternsofficiallive30 patterns, including dialog, combobox, tabs, menu button, disclosure, listbox, carousel, slider, grid, treegrid and toolbar.Verifiedcomponent-architecture.md, atlas accessibility rowsAPG examples are not production-tested for every assistive technology.

F. Performance and quality

#QuestionSourceTypeDateFindingStatusInformsCaveats
F1Core Web Vitalsweb.dev vitalsofficialupd 2024-10-31Good thresholds at the 75th percentile, split by mobile and desktop: LCP ≤ 2.5 s, INP ≤ 200 ms, CLS ≤ 0.1. Lighthouse cannot measure INP and uses TBT as a proxy.Verifiedperformance-engineering.md budgetsLab ≠ field.
F2Font loadingweb.dev font best practicesofficialupd 2022-10-04Inline @font-face; preconnect with crossorigin; preload sparingly; WOFF2 only; subset with unicode-range (check the licence); font-display optional for body, swap for brand; size-adjust fallback to cut CLS.Verifiedtypography.md §11, performance-engineering.md—

G. Measured corpus observations (own captures)

Method: skills/form-and-force/scripts/audit_page.mjs in headless Chromium (Chrome for Testing build 1243, Playwright 1.56). Viewports were 390×844 (touch) and 1440×900, with wheel- or scroll-driven sequences, a 25-stop keyboard probe, and a prefers-reduced-motion: reduce pass. Each site was run once on 2026-10-22, unthrottled, so these are lab, single-run figures. Raw data is in research/captures/<site>/report.json. The contact sheets (sheet.jpg) show third-party sites and are kept locally, not redistributed.

#ObservationnStatusInforms
G1Pages without exactly one <h1>12/21Verifiedaccessibility.md: structure is still owed on immersive sites
G2Pages without a <main> landmark8/21Verifiedsame
G3Skip link present4/21 (Linear, Grafana, Pudding, Shopify Editions)Verifiedsame
G4Viewport blocks zoom3/21 (Oryzo, Why Zero, Teenage Engineering)Verifiedslop_scan C rule
G5Reduced-motion code detected / behaviour measurably changed16/21 detected; 5 changed (Linear 82→76 running animations, Stripe 18→1, Shopify 7→1, Floema 3→1, Lando Norris swaps the loader for a static portrait)Verified"Detection ≠ handling": verify the visual result, not the media query
G6Consent UI dominates the first keyboard and visual experienceFloema (modal over every viewport), Dropbox Brand and House of Dreamers (Tab cycles inside the consent iframe), GOV.UK and Pentagram (consent first)VerifiedAtlas §Footer & secondary: consent is part of the design system
G7Entry gates and preloadersHouse of Dreamers ("Enter"), Siena ("Enter"), Why Zero ("draw a zero"), Bruno Simon ("click to start"). Lando Norris's loader filled about 10 s of mobile capture.Verified (headless; GPU may slow WebGL)motion-direction.md §9: justify the gate, offer a skip, never fake progress
G8Heavy WebGL often stalls in headless desktopIgloo desktop stuck at loader; Bruno Simon mobile stuck at portal ringVerified (limitation of the tool)RESEARCH_GAPS.md; audit_page docs mention GPU limits
G9Lab layout instability on data-heavy UIGrafana Play CLS 0.85–1.06 while panels hydratedVerified (lab, scroll-influenced)Atlas §Dashboards: reserve panel geometry
G10Stacks in the wildLenis 6/21, three.js 9/21, WebGL contexts 11/21, Webflow 4/21, Vue/Nuxt 4/21Verified (detection heuristics)Library choice is contextual; no single stack wins
G11"AI-default" typefaces on top sitesLinear (Inter Variable), Half of Eight (Inter Display), Grafana (Inter)VerifiedA typeface is not slop by itself; defaults are judged in context

H. The creative method (0.2.0)

Sources behind the method that replaced the concept engine. How they were used, and what was deliberately not taken, is in CREATIVE_METHOD.md.

#QuestionSourceTypeDateFindingStatusInformsCaveats
H1Why does a model asked for variety return the typical answer?Zhang et al., Verbalized Sampling: How to Mitigate Mode Collapse and Unlock LLM Diversitypreprint, later PMLRread 2026-11-10The authors attribute mode collapse to typicality bias in the data, and report 1.6 to 2.1 times more diversity in creative writing when a model is asked for a distribution of answers instead of one.Verified (abstract and summaries)direction.md §1 and §5: the choices that decide distinctiveness are dealt by a script, not chosenText tasks, not design. The method here uses external randomness, which the paper does not test
H2Does AI assistance narrow what a group produces?Doshi and Hauser, Generative AI enhances individual creativity but reduces the collective diversity of novel content, Science Advances 10(28), 2024peer-reviewed experimentread 2026-11-10With AI story ideas, individual stories were rated more creative and were more similar to one another. 293 writers, 600 evaluators, pre-registered.Verified (abstract, data record)floor.md §7 and the deal.mjs history: sameness across projects is treated as a fault of the method, and recent commitments are steered away fromShort stories, one model generation
H3Do examples constrain the designs that follow them?Jansson and Smith, Design fixation, Design Studies 12(1), 1991, pp. 3 to 11peer-reviewed experimentread via reviews 2026-11-10Designers shown an example reproduced its features, including flawed ones. The effect was also reported for professionals.Inferred (reviews of the paper; the paper itself was not read)direction.md §12: references are looked at before the hand and closed before sketching; the skill's references carry no example layouts or palettesReplications vary in effect size
H4Can a model judge its own work fairly?Panickssery, Bowman and Feng, LLM Evaluators Recognize and Favor Their Own Generations, NeurIPS 2024peer-reviewedread 2026-11-10Evaluators scored their own outputs higher than human raters did, and self-preference rose with the ability to recognise their own text.Verified (abstract)jury.md §3 and §4: a fresh context where possible, pictures before rationale, a cold read that is not edited afterwardsSummarisation, not design. A fresh context of the same model is not an independent judge
H5Is there a compact vocabulary for the character of a movement?Laban effort factors (weight, time, space, flow) and the eight effort actions, from teaching material on Laban movement analysissecondary teaching sourcesread 2026-11-10Three factors combine into eight named actions, each playable bound or free. Sources agree on the eight and disagree on details of flow.Inferredmotion-direction.md §4 and motion_tokens.mjs: a working vocabulary for choosing a motion characterBorrowed as vocabulary only. The mapping from factors to durations and curves is this project's own and is not derived from Laban
H6How does another skill get different pages from one model?Leonxlnx/taste-skill 2.5.1, commit f6b7657, MITsource read2026-11-10A script deals the hero for a page at random, weighted by kind of business and away from recent deals; a checker gates the finish and counts rounds. Palettes, type and eleven hero layouts are libraries.Verified (source)deal.mjs (dealing away from recent work) and the jury's round limitsIts deck holds looks. This project's deck holds questions and structural principles only
H7What separates motion that works from motion that feels designed?emilkowalski/skills: emil-design-eng, animate, review-animations, choreograph, find-animation-opportunities, apple-design, animation-vocabularysource read2026-11-10A gate that asks frequency before purpose; interruption through transitions and springs; origin-aware surfaces; nothing from scale 0; asymmetric timing; casting and relating movers; a required list of what does not move.Verified (source)motion-direction.md §2, §5, §6, §10; motion-craft.md; the questions in motion_strip.mjsIts values are one designer's house values. Here they are bands, and tokens are derived per project
H8Do the design-facts measures separate decisions from defaults?Own measurement of 19 reference sites and 11 templatesmeasured2026-11-10 to 11Five questions separate well; two are inverted; two were removed; the instrument is blind to a template with a photograph.Verified (measured)jury.md §8, design_facts.mjsINSTRUMENT_CALIBRATION.md

Open disputes recorded

  • Gradient backgrounds. Recommended in B6 (2025) and flagged in B5 (current). Resolution: neither "use" nor "avoid"; a gradient needs a material or brand reason (Stripe's ribbon is a brand asset).
  • One-word accent in headlines. Flagged in B5 as a tell, but used to good effect on Lando Norris (serif/sans word swap) and as a two-tone continuation on Stripe and Linear. Resolution: allowed when the accent encodes meaning (a brand voice, a sentence hierarchy), not as default flair.
  • Award-score thresholds. Third-party claims of "8.0 for SOTD" conflict with the official page, which sets no fixed cutoff. Resolution: use the official text.