free. all in one. ui/ux only. published design skills, emil kowalski motion rules, and copy-ready prompts you can paste into claude. bookmark it if it helps.
use this when you want a ui/ux pass with fable 5.2, and you do not want a general coding handbook.
you get a short path: pick a skill or prompt, paste it, then finish the visual work in one go.
this page is only ui/ux. copy a prompt when you want something paste-ready. open a skill or repo when you want a published tool. every third-party card names its author and links the source. numbers are from live fetches on 12 nov 2026, late evening ist. if a figure was not checked, the card is not here.
use this when you are about to ask fable 5.2 to design, restyle, or review an interface.
you get the official reminder that effort is a choice, applied here to ui work.
the official fable 5.2 prompting guide says to calibrate effort on purpose, not to reuse a setting from an older model. for a full ui pass, say the effort in the prompt. api default is high. claude apps and claude code default to medium. if the visual work is wide, say so up front.
use this when the model stops after comments, or leaves states, motion, and mobile as leftovers.
you get the official 'finish the task' rule, pointed at interface work.
the official guide says to tell fable 5.2 to finish the whole task, especially at low or medium effort. for ui, that means typography, color, layout, empty and error states, motion, and a narrow screen, not a list of leftover notes. say that in the prompt if you want a complete pass.
use this when you keep saying 'make it look better' and the page stays generic.
you get a short order of operations: skill first, then a prompt, then a review.
a vague 'make it nicer' request is how a lot of templated ui gets made. start with a published skill (frontend-design, impeccable, or emil's motion skills), name the product and the audience, then run a copy-ready prompt. review against the brief, not against a vibe. this page's prompts are written in that order.
use this when you are choosing a model for a design, landing, or visual pass, not for a general coding handbook.
you get a short, official-docs-only card. no invented ui benchmark. community visual posts are labeled as demonstrations.
official docs do not name a 'ui model'. anthropic's public guidance still tells you to start most work on opus 5.5, and it lists fable 5.2 for long-horizon work such as planning, migrations, and leading subagents. that is task guidance, not a design benchmark.
design job
what official docs actually say
what this page will not claim
a full ui or landing pass, or a long visual refactor
fable 5.2 is documented for long-horizon and planning-heavy work. set effort on purpose (api default high; claude code default medium).
that fable 'wins' design. no official ui leaderboard is cited here.
a small visual fix, a component, or a tight restyle
anthropic says start most work on opus 5.5, and sonnet 5.5 for well-scoped work. that guidance is general, not ui-specific.
a measured winner for buttons, type, or color.
you already work in the openai stack
openai's public start-here for complex work is gpt-6 astra. that is vendor guidance, not a ui test.
third-party scores, or any unlisted model name.
some public x posts show visual work made with fable 5.2 at high effort. those are demonstrations, not official results. one example is @other__reality (92 bookmarks, 13,449 views in the 12 nov 2026 fetch).
published skills and repos. star counts are live github figures from 13 nov 2026, about 11:30 pm ist (asia/calcutta). x metrics are from the same evening. a star count is for the whole repo unless the card says otherwise.
design skill · beginner · 10 min to install · then keep
use this when you want published motion and interface guidance, not a vibe prompt.
you get a mit library of skills for designers and engineers, plus the aiforui docs.
the repo's skill folders, as listed on github at this check, include animate, animate-expo, animation-performance, animation-vocabulary, apple-design, ask-sonner, break-ui, choreograph, emil-design-eng, find-animation-opportunities, gestures, improve-animations, learn-taste, mobile-native, pick-ui-library, prototype, remove-animations, review-animations, and write-swift. use the published easing and press rules from this set when you paste the motion prompts on this page.
by emil kowalski. announced and updated on his account, including this 26 oct 2026 post.
github, 13 nov 2026, about 11:30 pm ist: 56,597 stars, 3,311 forks, mit. x on that post, same evening: 1,244 bookmarks, 2,012 likes, 72,813 views.
design skill · beginner · 8 min to read · then keep
use this when the interface looks like a template, and you need a point of view before you code.
you get official guidance on distinctive type, palette, layout, and restraint, from anthropic's public skills repo.
the skill asks you to plan color, type, and layout against the brief before you build, and to avoid a set of generated defaults (a cream-and-clay kit, a black-and-acid kit, fade-and-slide on every section, and similar tells). it is not fable-specific. the typography and landing prompts on this page lean on it, in new wording.
github, 13 nov 2026, about 11:30 pm ist: the parent repo anthropics/skills has 194,050 stars and 22,818 forks. that count is for the whole repo, not this one skill. x, same evening: @pato__project 4,376 bookmarks, 174,314 views; @charliejhills 901 bookmarks, 47,388 views.
use this when you want a published skill that can critique polish and catch slop before you ship.
you get the public impeccable repo and site, with a live star count, not a figure copied from a thread.
the repo describes itself as a design vocabulary for a coding agent: steer with polish, review with critique, and catch ai slop before it ships. use the live github count if you quote size. do not reuse the 85k figure from the october post.
by pbakaus. listed in @AmirMushich's 23 oct 2026 repo roundup. that post used older star figures; the count below is from github at this check.
github, 13 nov 2026, about 11:30 pm ist: 109,429 stars, 6,521 forks, apache-2.0. x on the roundup, same evening: 3,247 bookmarks, 1,508 likes, 114,149 views.
use this when you want a skill that proposes patterns, palettes, and type for a brief.
you get the public nextlevelbuilder skill, with a live star count.
the repo describes itself as design intelligence for ui/ux across platforms. it is a community skill, not an anthropic official skill. check the license and the docs on the site before you drop it into a client repo.
by nextlevelbuilder. named in @X1Kafadzhy, @AmirMushich, and other october and november 2026 lists. post star figures lagged github; use the live count below.
github, 13 nov 2026, about 11:30 pm ist: 149,014 stars, 15,764 forks, mit. x, same evening: @X1Kafadzhy 259 bookmarks, 13,235 views.
use this when you want a brand's design.md as a file an agent can actually follow.
you get a public collection of design.md files, plus the project site.
the repo is a collection of design.md files drawn from brand design systems. the description says you can drop one into a project and let an agent generate a matching ui. that is a workflow claim from the repo, not a test run on this page.
by VoltAgent. named next to ui ux pro max in @X1Kafadzhy's 9 nov 2026 skills post.
github, 13 nov 2026, about 11:30 pm ist: 127,381 stars, 14,107 forks, mit.
use this when you want vercel's public skill collection, including web design guidelines mentioned in community lists.
you get the official vercel-labs/agent-skills repo, with a live star count.
this is vercel's official public collection of agent skills. the repo is broader than ui. if you only want design guidance, open the web design skill inside it rather than treating the star count as a ui-only figure.
by vercel labs. web-design-guidelines from this collection is named in @Shahporan_Razu's 26 oct 2026 five-skills post.
github, 13 nov 2026, about 11:30 pm ist: 34,009 stars, 2,967 forks.
use this when a figma file keeps moving, and the agent is coding from a stale frame.
you get an open figma-to-code change tracker, credited with its real (small) star count.
the repo describes a figma plugin, api, and agent tooling that tell a coding agent which design changes are not built yet. the post has a lot of views and very few bookmarks and stars, so treat reach and adoption as different facts. the view count is from the x fetch, not a separate analytics product.
original, copy-ready ui/ux prompts written for fable 5.2. they lean on official fable prompting, anthropic frontend-design, and emil kowalski's published motion rules, in new wording.
use this when the interface already moves, and you cannot tell if the motion is helping.
you are reviewing motion in this repo. follow emil kowalski's published skills (animate, review-animations, emil-design-eng, gestures). do not invent taste.
for each animated thing, first decide if it should move at all:
- 100+ times a day, including keyboard shortcuts and a command palette: no animation.
- tens of times a day (hover, list move): almost none, or none.
- occasional (modal, drawer, toast): standard motion.
- rare or first-time: delight is allowed.
then check:
- enter and exit use ease-out. movement on screen uses ease-in-out. hover and color use ease. loops use linear. never ease-in on ui.
- if you need a custom curve, use the published ones: --ease-out cubic-bezier(0.23, 1, 0.32, 1); --ease-in-out cubic-bezier(0.77, 0, 0.175, 1); --ease-drawer cubic-bezier(0.32, 0.72, 0, 1). do not invent cubic-bezier(0.4, 0, 0.2, 1).
- durations: press 100-160ms, tooltip 125-200ms, dropdown 150-250ms, modal and drawer 200-500ms, ordinary ui under 300ms.
- never enter from scale(0). start at 0.9 or higher, plus opacity.
- pressable elements use transform: scale(0.97) on :active, 160ms ease-out.
- gate hover with @media (hover: hover) and (pointer: fine).
- prefers-reduced-motion means fewer and gentler, not zero. keep opacity and color. drop movement and position.
- animate transform and opacity unless clip-path is the point, or height on an accordion.
write a table: file, what moves, frequency, purpose, verdict, one-line fix. do not implement yet.
you get a table of what moves, how often, and what to change, using emil kowalski's published rules.
use this when buttons and rows highlight, but they never feel pressed, or the press state shifts neighbours.
add honest press states to interactive controls in this ui.
use transform: scale(0.97) on :active, transition transform 160ms ease-out, following emil kowalski's published press guidance. do not change layout, margin, or width on press. do not use a hard brightness filter as the only feedback.
keep a visible :focus-visible ring. if a control is disabled, it does not scale and it does not look pressable.
show the css and the list of selectors you changed.
you get press feedback with scale(0.97), 160ms ease-out, and no layout jump.
use this when hover styles linger on a phone, or a sticky hover breaks a tap target.
find hover styles and hover motion in this ui that can stick on a touch screen.
wrap hover motion in @media (hover: hover) and (pointer: fine). keep :active and :focus-visible working without hover.
make hit targets at least 44px where they are meant to be tapped. do not shrink text to make room.
list every selector you changed, and call out any hover-only content a phone user would miss.
you get hover motion that only runs on a fine pointer, with a clean tap path.
use this when the reduced-motion branch either does nothing, or it still slides things around.
do a prefers-reduced-motion pass on this ui, following emil kowalski's published rule: fewer and gentler, not zero.
keep opacity and color transitions. drop movement, position, and scale on that branch.
do not disable the interface. do not hide content that only appeared mid-animation.
show the media queries and a short before/after for the three most visible motions.
you get a reduced-motion pass that keeps opacity and color, and drops movement.
use this when a dialog pops from nowhere, or focus falls behind it.
rebuild this modal (or the main dialog in the repo) as a complete pattern.
open from the control that triggered it, not from the center of the viewport unless there is no source.
use ease-out to enter, and a 200-500ms duration. never scale from 0. trap focus while it is open, and return focus to the trigger on close. escape and a scrim click both close it, unless the action is destructive and needs a confirm.
respect prefers-reduced-motion. write the markup, styles, and the focus behavior.
you get a modal that opens from its source, uses published easing, and returns focus.
use this when you need a sheet or drawer, and the current one feels like a modal glued to the bottom.
implement a bottom sheet for this ui.
when the user drags, the sheet should follow the finger. when they let go, continue with the last speed, then settle. emil kowalski's vaul drawer and the published --ease-drawer curve (cubic-bezier(0.32, 0.72, 0, 1)) are the references, not a bounce you invent.
offer a drag handle, a scrim, and a way to close from the keyboard. do not lock the page scroll in a broken state. include a reduced-motion close that does not fling.
you get a sheet with drag, a published drawer curve, and a calm close.
use this when cards or popovers grow out of a point and feel cheap.
search this ui for entrances that use scale(0), width/height from 0, or a pop that starts invisible and tiny.
replace them with a start of scale(0.9) or higher plus opacity, ease-out, under 300ms for ordinary ui.
do not add bounce. show a table of file, selector, old behavior, and the new values.
you get a list of scale(0) uses, each rewritten to start at 0.9 or higher plus opacity.
use this when a gallery or feed animates in so slowly that people click the wrong thing.
this list or gallery staggers in. keep a light stagger if it helps orientation, but:
- no item should wait so long that it feels broken.
- items must be clickable as soon as they are visible.
- the stagger is for first paint or a filter change, not for every tiny update.
- prefers-reduced-motion drops the stagger and uses a short opacity fade, or nothing.
show the timing function and the cap you used on delay.
you get a short stagger that does not delay interaction, and a reduced-motion branch.
use this when icon tooltips each wait, so a toolbar feels laggy.
fix tooltip timing on this toolbar or icon group.
the first tooltip in a group may delay 125-200ms. once one is open, moving to a neighbour should show the next tip with no delay, until the pointer leaves the group.
do not use tooltips for essential labels on a phone. if the control needs a text label, give it one.
include a reduced-motion path and keyboard focus.
you get a group of tooltips where only the first has a delay.
use this when the tab highlight jumps, or both panels sit on screen during a fade.
rebuild the tab indicator and panel swap on this component.
the active indicator should travel to the selected tab, not crossfade between two highlights. keep the motion short (about 150-250ms, ease-in-out).
panels should not sit stacked and half-visible. if you animate the panel, prefer a short opacity change. do not use scale(0).
include keyboard arrows, aria roles, and a reduced-motion branch.
you get an indicator that travels, and a panel change that does not blur together.
use this when the signed-in app uses the same flourish as the marketing hero, and it gets in the way.
audit motion in this repo and split it into two policies.
marketing or first-run screens may have one orchestrated moment. the product ui should follow emil kowalski's frequency rules: things people do all day do not animate for delight.
write a short policy, then list the motions you would remove from the product and the one moment you would keep on the marketing surface. do not implement until the list is clear.
you get two motion policies: one for marketing, one for daily product use.
use this when every block slides up on scroll, and the page feels generated.
this page fades or slides every section as it enters the viewport. that is a common generated default.
keep at most one orchestrated entrance (hero or first screen). remove scroll-triggered fade-and-slide-up on the rest. hover on every card should not travel.
motion that answers a click (open, expand, confirm) can stay if it shows what changed.
show what you removed and what you kept. respect prefers-reduced-motion.
you get one planned entrance, with the rest of the page staying still.
use this when the happy path looks finished, and the other three states look like leftovers.
design empty, loading, and error states for this view as one family, not three one-offs.
each state should say what is going on, and what the person can do next. errors name the problem and a fix, in the product voice. they do not apologize, and they are not vague.
an empty state is an invitation to act, not a cute dead end. loading should not jump the layout when data arrives.
use the existing type and color tokens. show all three at desktop and at a 390px width.
you get three states that share type, color, and a next action.
use this when validation pops in, the layout shifts, and the first error is off screen.
fix form validation in this ui.
show errors next to the field, not only in a banner at the top. do not shove the page when an error appears. move focus to the first invalid field, and announce the error to a screen reader.
keep a visible label. do not use placeholder as the only label. disable the submit state honestly while a request is in flight, then return it.
show the happy path, one field error, and a server error.
you get inline errors, stable layout, and focus on the first problem.
use this when the desktop mock looks fine, and nobody has checked 390px.
do a visual qa pass on this ui at 1280px and at 390px (a real phone width, not a scaled desktop).
look for overflow, overlapping type, hit targets under 44px, text that wraps into orphans, and motions that misbehave on touch.
write a punch list with file or component, width, issue, and a fix. then fix them. do not stop at the list.
you get a visual pass at two widths, with a punch list of real breaks.
use this when every card uses a two-word name and a perfect thumbnail.
stress this ui with awkward, realistic content.
use long names, one-character names, missing images, empty lists, huge numbers, zero, and a 12-item overflow where the mock had three.
fix truncation, wrapping, and image fallbacks so the layout holds. do not invent placeholder latin if the product has a real voice.
show screenshots or marked-up examples for the worst three cases, before and after.
you get a pass that uses long names, missing images, zero, and too many items.
use this when you have references that almost match, and the live ui does not.
here are screenshots or references for this product. extract a small token set (color, type, radius, spacing) that can cover all of them.
then list where the live ui drifts: alignment, density, type size, color, and motion. fix the drift. if a reference disagrees with another, say which one you followed and why.
do not average them into a generic look.
you get shared tokens and a list of the differences you will not ignore.
use this when it 'is responsive' but the header, inputs, and bottom bar fight the browser chrome.
do a mobile web pass on this app.
account for safe-area insets, 44px targets, and input zoom (keep form text at a readable size). do not hide actions in hover-only menus.
if there is a bottom bar, it should not cover content or the home indicator. test at 390px width.
list the issues, then fix them.
you get safe-area, 44px targets, and a layout that uses the narrow screen on purpose.
use this when padding is a mix of 11, 14, 18, and 22, and the page feels restless.
extract or invent a small spacing scale (for example 4, 8, 12, 16, 24, 32, 48) and apply it to this view.
replace one-off paddings and gaps with the scale. then look at the page and remove one increment of space everywhere it is not earning its keep.
do not flatten a layout that needs a clear section break. show the scale and the files you touched.
you get a named spacing scale and a tighter, more even page.
use this when icons, type, and avatars sit on a grid and still look drunk.
do an alignment pass on this view.
line up icon and label pairs, avatar and name rows, and numbers in tables. watch for optical misalignment (a play icon that looks left of center, a typographic quote that should hang).
do not add new decoration. show the before and after for the three worst rows.
you get a pass on optical alignment, hanging punctuation, and icon-text pairing.
use this when a dashboard is packed for a power user, and a first-time user cannot parse it.
review density on this dashboard or list.
pick a default that a new user can scan. if you keep a compact mode, do not let hit targets fall below 44px, and do not shrink type below a readable body size.
show both densities at 1280px and the default at 390px.
you get one default density and, if needed, a compact option that does not shrink hit targets below 44px.
use this when a sticky bar covers filters, or two sticky rows stack into a wall.
fix sticky positioning in this ui.
keep a single sticky header unless a second one is truly needed. it should not cover the first row of content or the primary action. add a light edge or shadow so content can scroll underneath without looking broken.
on a phone, account for safe-area. show 1280px and 390px.
you get one sticky layer, a clear shadow or edge, and room to scroll.
use this when the page uses six font sizes that were picked in the moment.
set a type scale for this ui.
use one typeface, or two if they are clearly distinct. give each size a role (display, title, body, caption, numeric). do not pick sizes at random.
keep body line length under 80 characters. give body a line-height that fits the face (a bit more if it is a serif).
do not accent a single word in a headline just to look designed. do not use tracked-out all-caps labels unless the brief asks for them.
show a specimen, then apply the scale to this view.
you get a named type scale, at most two families, and a specimen.
use this when the ui is inter plus a serif because that is what the model likes, not because of the brief.
the brief is this product and this audience. propose typography that belongs to that brief, not a default stack you would use on any saas page.
you may use one family. if you use two, say why. avoid the usual generated tells: a cream page with a terracotta accent and a fashion-serif headline, or a near-black page with one acid accent, unless the brief actually asks for that.
show three type options, pick one, and apply it to the main view. name the faces and the fallbacks.
you get a deliberate pair (or one family) with a reason tied to the product.
use this when body text runs wide, or every heading has a leftover label above it.
do a typography hygiene pass on this page.
keep body lines under 80 characters. fix awkward rags and orphans where you can do it without weird breaks. remove typographic labels, middle-dot meta strings, and all-caps eyebrows that do not add information.
numbers in tables should be tabular if the face supports it. do not add a new typeface in this pass unless the current one cannot do the job.
you get readable measure, a cleaner rag, and fewer typographic gadgets.
use this when a light theme passed, and the dark theme text is smoke on charcoal.
audit color contrast on this ui in light and dark, if both exist. if there is only one theme, audit that, and say so.
body text, muted text, links, and buttons need to be readable against their surfaces. call out any muted gray that fails for small type. do not 'fix' contrast by making everything high-contrast black and white if the product has a real palette.
list the failing pairs with the approximate use, then fix them. do not invent exact contrast ratios you did not measure.
you get a contrast pass for body, muted text, and links, in light and dark.
use this when the product has a rainbow of accents and no hierarchy.
propose a palette of 4-6 named hex colors for this product, with roles (background, surface, text, muted, accent, danger).
ground it in the subject matter of the brief. do not default to a cream-and-clay kit or a black-and-acid kit unless the brief points there.
use the accent sparingly. status colors (success, warning, danger) should not be the brand accent unless they really are.
show the palette, then apply it to one main view in light, and dark if the product has it.
you get 4-6 named colors, with roles, tied to the brief.
use this when dark mode is just the light theme with the whites flipped, and images look neon.
if this product has a dark mode, design it as a theme, not as an invert.
set surfaces, borders, and text on purpose. wash or re-export images and illustrations that blow out. shadows are usually quieter in dark than in light.
keep contrast readable. show the same view in both themes, and list the tokens you changed.
you get a dark theme with its own surfaces, borders, and elevation.
use this when success, warning, and links all use the same accent, so nothing reads as a status.
audit how color is used for meaning in this ui.
links, the brand accent, success, warning, and danger should not all share one hue. a form error should not look like a marketing highlight.
propose a small semantic set and apply it to buttons, alerts, and form errors. keep it readable in both themes if both exist.
you get semantic colors that do not fight the brand accent.
use this when the first screen looks considered, and every new screen invents values again.
extract design tokens from this ui (or from the brief, if the ui is new).
include color, type, spacing, radius, and motion (easing and duration), using emil kowalski's published motion values where you need motion tokens.
name them in language a human can read. then point the next screens at the tokens, and list any one-off values you refused to add.
do not invent a huge token system for a small page.
you get a short token file for color, type, space, radius, and motion.
use this when notifications stack, cover actions, or stay forever.
implement or fix toasts in this ui. emil kowalski's sonner is the reference for tone, not a visual clone you have to copy pixel for pixel.
a toast should not shove the page. it should be dismissible, should not cover the primary action, and should not stack without a cap. use a short, honest duration. errors can stay longer than success.
match the action name ('published' after a 'publish' click). include reduced motion. show success, error, and a stacked case.
you get a toast pattern with a short life, a way to dismiss, and no layout shove.
use this when every secondary flow is a centered modal, including long forms.
look at this ui's overlays. if a flow is long, on a phone, or feels like a panel (filters, details, a compose box), use a drawer or sheet, not a centered modal.
keep a modal for short, focused decisions. use the published drawer curve and press/focus rules already on this page.
write the rule in one short paragraph, then implement the one flow that most needs the change.
you get a rule, then a drawer for the long or mobile-heavy flow.
use this when the table wraps badly, numbers are proportional, and there is no empty state.
redesign this data table so it can be scanned.
align numbers to the right and use tabular figures if the face allows it. keep the header quiet. wrap or truncate text on purpose, not by accident. offer a way to handle overflow at 390px (horizontal scroll with a clue, or a stacked card), and an empty state.
do not add a rainbow of row colors. show 1280px and 390px.
you get alignment, tabular numbers, a quiet header, and overflow that does not break.
use this when the custom dropdown looks nice and fails arrow keys, typeahead, or escape.
rebuild this select, combobox, or custom dropdown so a keyboard user can finish the task.
arrows, typeahead if it is a combobox, escape to close, and focus back to the trigger. do not replace a native select with a div unless you are ready to do this work.
label it visibly. include an empty and a disabled state. gate hover, and keep 44px targets on a phone.
you get a select pattern with keyboard, focus, and a clear closed state.
use this when the sidebar looks even, and the current section is almost invisible.
redesign the primary navigation for this product.
make the current section obvious. group items by task, not by the team's org chart. on a phone, do not hide the current location in a mystery icon.
keep labels in plain language. include a collapsed or mobile pattern. show both widths.
you get clear current state, a sensible order, and a mobile path.
use this when the palette is used all day, and it still plays a long entrance.
build or fix the command palette in this app.
people will open it many times a day, so do not animate it. emil kowalski's frequency rule applies: no animation, ever, for this kind of control.
focus the input immediately. keyboard navigation must work. show recent or suggested commands without getting in the way, and a calm empty search state.
keep it usable at 390px.
you get a fast palette with no motion, good search, and a quiet empty state.
use this when filter chips wrap badly, or there is no way to clear one without clearing all.
implement or fix filter chips for this list.
selected state should be obvious. each chip should be clearable, and there should be a clear-all that is hard to hit by accident.
plan overflow at 390px. do not use color alone to mean selected. 44px targets on a phone.
you get a chip group with clear, overflow, and a visible selected state.
use this when the list jumps, or it never admits it is finished.
fix how this list continues.
if you paginate, do not reset scroll in a rude way, and show where the person is. if you infinite-scroll, include a load-more fallback and a clear end state. do not leave a spinner up when there is nothing left.
keep item focus if the person was keyboard-navigating. show an empty first page too.
you get a pagination or load-more pattern that keeps place and has an end state.
use this when the landing hero is a gradient, a big number, and a generic headline.
redesign the landing hero for this product.
lead with the most characteristic thing in its world: a real screenshot, a short demo, a sentence only this product can say. do not default to a big number, a gradient wash, and a stacked saas card row unless the brief truly wants that.
one primary action, one secondary if needed. real copy, no lorem. show 1280px and 390px.
you get a hero that uses the product's real subject, and one clear action.
use this when the logo wall and the '10x faster' line cannot be backed up.
review the social proof on this landing page.
if a number, quote, or logo cannot be sourced, remove it or replace it with a claim you can stand behind. do not invent counts, ratings, or customer names.
a small set of real logos or a short, attributed quote is enough. say what you cut and why.
you get proof you can actually stand behind, or the section removed.
use this when features, quotes, and a second hero are all saying the same thing in different type.
map this landing page section by section: job, audience moment, and the one action (if any).
merge or cut sections that repeat. keep a single memorable element, and let the rest be quiet. this is a restraint pass, not a new illustration pass.
show the new outline, then implement the cut.
use this when the pricing grid highlights the 'popular' plan so hard that the others are unreadable.
redesign this pricing table so plans can be compared.
do not hide the important limit in a tooltip. if a plan is recommended, say why in one line, without drowning the others.
stack cleanly at 390px. if prices vary by region or billing period, make the toggle obvious. do not invent discounts.
you get a comparable table, honest footnotes, and a usable mobile stack.
use this when the page has four 'get started' buttons, or the button says 'submit'.
audit calls to action on this landing or onboarding flow.
a primary button should say what happens ('start free workspace', 'book a demo'), not 'submit' or 'click here'. use the same name later in the flow (a toast should not rename the action).
keep one primary cta in each viewport. secondary actions stay quieter. show the before/after labels.
you get one clear primary action, in the product's words.
use this when the footer is a sitemap of dead links and a newsletter box with no promise.
redesign the footer for this site.
keep the links people actually need (product, docs, legal, contact). drop placeholder columns. if you include a signup, say what they are signing up for.
check it at 390px. do not add a massive logo row you cannot verify.
you get a short footer with real links and no fake legal theater.
use this when you are about to build a landing page and you do not want the usual template.
before writing layout code for this landing page, write a compact design plan:
- color: 4-6 named hex values.
- type: the faces and their roles.
- layout: one-sentence structure, plus a simple ascii wire of the first screen.
- principles: what will make this page specific to the brief.
then review the plan against the brief. if a choice is the generic default you would produce for any similar product, change it and say why. only then build the page from the revised plan.
do not start with a card kit.
you get a short design plan (color, type, layout, principles), reviewed against the brief, then the build.
use this when a product hero has a 3d scene, and it is also on every button.
review 3d or heavy visual effects in this ui.
keep at most one 3d moment (usually the hero or a product viewer). do not put a 3d flourish on every control.
provide a still image fallback and a prefers-reduced-motion path that does not auto-orbit. make sure the scene does not block the primary cta.
show what you kept and what you removed.
you get one 3d moment, a static fallback, and a reduced-motion path.
use this when the model is lit like a studio, and the page is a flat dark theme (or the reverse).
this page mixes a 3d object with 2d ui, and they do not share a world.
match lighting, background, and shadow to the surrounding theme (light or dark). if you cannot relight the asset, change the frame around it so the cutout does not look pasted.
do not add extra glow to hide the mismatch. show before and after.
you get a lighting and background pass so the object sits in the same world as the ui.
use this when the scene looks fine on a desktop gpu, and a phone heats up or stutters.
profile and, if needed, simplify the 3d or canvas hero on this page.
set a budget you can explain (resolution, polygon or effect count, whether it runs when off-screen). on a small screen, prefer a still or a simpler scene.
do not autoplay sound. pause the scene when it leaves the viewport if that is cheap to do. say what you changed and what you could not measure.
you get a documented budget, a simpler mobile asset or a still, and no surprise autoplay audio.
use this when a viewer for a 3d object also makes the menus and sliders 3d, and they are hard to use.
separate the 3d subject from the chrome in this viewer.
menus, sliders, and labels should stay 2d, readable, and at least 44px where they are tappable. do not put essential labels only inside the scene.
include a keyboard path for the main actions. show the viewer at 1280px and 390px.
you get flat, readable controls around the object.
use this when the model writes a careful review and leaves the pixels unchanged.
you are doing a ui pass on this repo. finish the work in the files, not as a punch list.
cover typography, color, layout, empty/error/loading, motion, and a 390px width. follow emil kowalski's published motion rules where motion is involved.
if something cannot be finished, say why in a short leftover list. do not use the leftover list as the deliverable.
you get an implemented pass, plus a short leftover list only for what truly could not ship.
use this when you need easing and duration you can paste, not a curve you invent.
you get the published ease-out, ease-in-out, and drawer curves, plus the duration bands used in this page's prompts.
--ease-out is cubic-bezier(0.23, 1, 0.32, 1). --ease-in-out is cubic-bezier(0.77, 0, 0.175, 1). --ease-drawer is cubic-bezier(0.32, 0.72, 0, 1). press is about 160ms at scale(0.97). never ease-in on ui. never enter from scale(0). reduced motion is gentler, not zero. this page's own motion follows those rules.
use this when you want a three-step motion loop: find, add, then review.
you get the official folder names in emil's repo, so you can install the right skills.
find-animation-opportunities is for deciding where motion belongs. animate is for adding it. review-animations is the pass afterwards. improve-animations, remove-animations, choreograph, and animation-performance sit in the same skills folder if you need those jobs.
from emil kowalski. a 17 oct 2026 note from @jimmyyangrz described this three-step loop. the folders were confirmed in the repo at this check.
use this when the interaction has to follow a finger, or you are on iOS-style mobile ui.
you get the two skill folders that cover tracking, release, and mobile-native motion.
gestures is the skill to open when a sheet or control should track the pointer and keep moving after a release. mobile-native is the skill when the work is in that language. apple-design also lives in this same repo if you want the human-interface side, rather than a third-party copy.
use this when you want apple-style interface rules as a skill, and you want the file that actually lives in this library.
you get a confirmed path: emilkowalski/skills/skills/apple-design, not an unofficial fork.
posts about 'apple-design' are easy to remix into unofficial copies. use this official folder if you want the skill from the same library as animate and gestures. the skill is about feel (how a surface follows a finger, what happens on release), not about frosting every panel.
from emil kowalski. community posts such as @Whispertgx (4 nov 2026) talked about an apple-design skill; the folder is in this repo at this check.
parent repo, 13 nov 2026, about 11:30 pm ist: 56,597 stars.
use this when you want the agent to pick up a design-engineering point of view, not only a curve.
you get two skill folders for taste and design engineering, in the same repo.
learn-taste and emil-design-eng are the skills to load when the problem is judgment, not a single animation. read the essay if you want the argument in long form. do not replace these with a prompt that only says 'have taste'.
use this when you want the community roundup that people actually bookmarked this month, not a new invented list.
you get two high-save october 2026 posts, plus a reminder to verify every repo they name.
these posts are lists, not official docs. they name overlapping skills (frontend-design, emil's motion work, and others). if you follow a link from either thread, check the github repo yourself. this page already verifies the main ones in the skills section.
use this when the motion is in, and you need words for what it is doing, or it is janky.
you get two more official folders: animation-vocabulary and animation-performance.
vocabulary is for naming the motion so a review can be specific. performance is for when the animation is doing too much work. there is also a performance-cheatsheet.md at the repo root.
use this when you want to see someone actually try the skills, not only a star count.
you get a 30 oct 2026 post, and a 4 nov note that points at it.
these are demonstration posts, not a controlled study. they are here so you can see the skills used on a real page (a 'motion specimen' in the follow-up note). judge the work yourself.
x, 13 nov 2026, about 11:30 pm ist: the 30 oct post had 1,028 bookmarks, 937 likes, 115,599 views. the 4 nov note had 39 bookmarks, 20 likes, 4,196 views.
use this when the page is full of labels, numbers, and cards that do not mean anything.
you get a short reading of anthropic's frontend-design advice on structure, in new words.
borders, eyebrows, numbers, and dividers should encode something about the content. numbered markers make sense on a real sequence, and they look like kit when the content is not a sequence. spend boldness in one place, and cut a decoration that is not doing a job. the layout prompts on this page start from that idea.
paraphrased from anthropic, frontend-design. not a verbatim extract.
use this when you want the 23 oct 2026 roundup, with star counts updated at this check.
you get the post, plus live github figures for the repos this page verified.
use this as a map, then open the verified cards on this page for current stars. do not quote the post's '550k total' or its per-repo figures; they did not match github on 12 nov 2026.
from @AmirMushich, 23 oct 2026. the star totals in that post were already stale by this check, so they are not repeated here.
x, 13 nov 2026, about 11:30 pm ist: 3,247 bookmarks, 1,508 likes, 114,149 views. live github for repos named on this page: impeccable 109,429, ui ux pro max 149,014, emil /skills 56,597.
use this when you need to stress a layout, or you want a prototype skill from the same library as the motion ones.
you get two official folders in emilkowalski/skills.
break-ui is the skill to reach for when you want the layout tested against awkward cases. prototype is for early interface work in the same library. they sit next to pick-ui-library if you are choosing a kit rather than inventing one.
use this when you are on react native and want one public shelf instead of a scavenger hunt.
you get volodymyr's 30 oct 2026 post and the nativebloom skills tab.
the post groups react native agent skills: design, animations (emil kowalski), and ui kits, and says 18 were on the shelf at the time. this page did not re-count the shelf. open the tab and verify a skill before you install it.
use this when you want a recent example of a dense product ui, without the post's career hype.
you get the 5 nov 2026 post, as a demonstration, not as a promise.
the post shows a medical x-ray dashboard and says it was built with claude code and free github tools, without traditional design software. that is the author's claim. this page does not repeat the 'designers are in danger' framing, and it does not treat the post as a controlled study.
use this when you want a short ui checklist before you invent a new one.
you get the ui-cheatsheet.md file confirmed at the repo root.
this is a file in the same library as the skills, not a third-party gist. read it alongside the motion tokens. if it conflicts with a client brief, the brief wins.
use this when you are about to pick faces, and you can feel the default stack arriving.
you get a short, original wording of the official typography guidance.
one family is enough; two is fine if they are clearly different. pick faces for this brief, not the pair you use on every project. keep body lines under 80 characters. do not accent a single word in a headline just to look designed, and do not lean on tracked all-caps labels as a style. typography carries personality, and it is part of the design, not a delivery vehicle.
use this when the palette looks like every other model-made page this month.
you get the default looks frontend-design warns about, in new words, so you can choose something else.
the skill flags recurring defaults: a warm cream field with a terracotta accent, a near-black field with one acid or vermilion accent, a hairline broadsheet, and a saas-card kit with one radius and a soft grey shadow. those looks are fine when the brief asks for them. they are a problem when they appear no matter what the product is.
use this when you are about to generate a page and you want a uniqueness check first.
you get the official two-pass process, written short.
write a compact plan: 4-6 named colors, type roles, a layout idea, and a few principles. review that plan against the brief. if a choice is the generic default you would produce for any similar product, change it and say why. only then write the code. the 'plan the page' prompt on this page follows that order.
use this when you want a starting palette and type pair from a published skill, then you will edit it.
you get ui ux pro max, as a starting point, with its live github size.
treat the output as a proposal. check contrast yourself, and swap anything that is just a generated default. this page does not claim the skill 'matches' a brand until you review it.
by nextlevelbuilder. also named for palettes and font pairs in @X1Kafadzhy.
github, 13 nov 2026, about 11:30 pm ist: 149,014 stars.
use this when you have permission to match a known system, and you do not want to invent tokens from a screenshot alone.
you get awesome-design-md as a file source, not as a license to copy a brand you do not have the rights to use.
a design.md is useful when you are allowed to follow that system. it is not a shortcut to ship a lookalike of a brand you do not represent. read the file, then check contrast and voice against your brief.
use this when the type is careful, and the labels still say 'submit' and 'error occurred'.
you get a short, original wording of frontend-design's writing advice.
name things the way a person understands them, not the way the system is built. a button should say the action, and the same word should show up after the action works. errors should say what happened and how to fix it. empty states should invite a next step. this is typography because the words sit in the type, and they are the interface.
use this when every section is trying to be the moment.
you get the restraint rule from frontend-design, in new words.
spend boldness in one place. keep a quality floor (it works on a phone, focus is visible, reduced motion is respected, contrast holds) without announcing that floor. if you can take one accessory off and the page is stronger, take it off.
use this when you need notifications that do not shove the page around.
you get emil kowalski's opinionated react toast, with a live star count and the official site.
the repo describes it as an opinionated toast component for react. use it, or use it as a reference for tone and stacking, as the toast prompt on this page does. you do not have to clone the look.
by emil kowalski, emilkowalski/sonner. the skills repo also has an ask-sonner folder.
github, 13 nov 2026, about 11:30 pm ist: 13,213 stars, 472 forks, mit.
use this when you need a drawer or bottom sheet that can follow a drag.
you get emil kowalski's react drawer, with a live star count and the official site.
the repo describes it as a drawer component for react. the sheet prompt on this page points at vaul and at the published drawer easing curve, without asking you to copy the pixels.
use this when you want a kit with a point of view, and you are tired of the same default set.
you get the 8 nov 2026 list, with the author's own paid-content notes kept in place.
the post names looseleaf (paper and tape), plopui (marker and sticker), inkra (pencil and wash), goopui (vue/nuxt slime-like motion), solid glass (glass plus agent states), affectui (agent tree and traces, some paid), prestige ui (react/tailwind/motion, some paid), and boardui (dashboards and agent workspaces, some paid). this page did not verify github stars for those eight, so none are invented here. open the post for the links, and check a license before you ship one.
from @Justin1024go, 8 nov 2026. the post argues that a good component library can matter more than a design skill. that is an opinion, not a measured result.
x, 13 nov 2026, about 11:30 pm ist: 1,199 bookmarks, 707 likes, 42,617 views.
use this when you want a paper, tape, and sketched-chart look, and you have checked the license.
you get the looseleaf site as linked in the 8 nov 2026 post.
the post describes paper texture, tape, hand-drawn charts, and full page blocks. this page did not re-audit the library. confirm the package, license, and maintenance before you depend on it.
use this when you want a sticker-and-marker kit, and you have checked the license.
you get the plopui site from the same post.
the post describes blue marker energy, sticker avatars, and a press state that squashes. treat that as the author's description. check the site and the license yourself.
use this when you want a pencil and wash look, including calendars and maps.
you get the inkra site from the same post.
the post describes pencil stroke and wash, including calendar and map treatments. confirm what is actually in the library before you plan a whole product around it.
use this when you want slime-like motion components, and you are in a vue or nuxt stack.
you get the goop docs link from the same post.
the post says goopui is vue and nuxt first, with components that drip or stretch. motion like that still needs a reduced-motion path if you put it in a product. this page did not verify framework support beyond the post and the docs url.
use this when you want a quieter glass treatment, including agent status and diffs.
you get the solid glass site from the same post.
the post describes restrained glass, with agent state, tool steps, and a diff view. glass can wreck contrast, so check text on those surfaces yourself.
design skill · beginner · 5 min some parts paid, per the post
use this when you are looking at the rest of that list, and you need the paid-content warning in one place.
you get the three libraries the post marked as partly paid, with their sites.
the post says affectui has some paid parts (the agent tree in the screenshot is called pro), prestige ui includes paid content, and boardui includes paid components. this page does not price them or endorse a plan. if you only want free tools, start with the libraries above, or with sonner and vaul.
use this when you are about to generate a landing page and you can already see the gradient arriving.
you get frontend-design's hero advice, in new words.
open with the most characteristic thing in the product's world, in a form that fits: a sentence, an image, a live demo, or something else you chose. a big number, a small label, supporting stats, and a gradient accent is the default treatment. only use it if it really is the best option for this brief.
use this when you are tempted to start in a card kit and 'make it unique later'.
you get the two-pass process, pointed at marketing pages.
write color, type, layout, and principles first. review them against the brief. then build. the 'plan the page' prompt on this page is the paste-ready version of that. a unique landing page is usually a unique plan, not a unique hover.
use this when you want a skill that claims landing-page coverage, with a live star count.
you get the repo's public description and topics, not a quality score.
a topic tag is not a guarantee. if you use the skill for a landing page, still run the restraint and proof prompts on this page, and do not invent metrics in the hero.
by nextlevelbuilder. github topics on the repo at this check include landing-page, ui-design, and uikit.
github, 13 nov 2026, about 11:30 pm ist: 149,014 stars.
use this when you want the page to follow a written system rather than a moodboard only.
you get awesome-design-md, whose topics at this check include landing-page and design-tokens.
if you drop a design.md into a landing build, make sure you are allowed to use that system, and that the copy on the page is still true. a matching skin does not fix a fake number.
by VoltAgent.
github, 13 nov 2026, about 11:30 pm ist: 127,381 stars.
use this when you want a tiny starting prompt, then you will replace the generic bits.
you get an original rewrite of a 27 oct 2026 prompt, not a paste of the original text.
a safer starter than 'make a modern premium landing page' is: name the product, the audience, and the one action. ask for a type scale, a 4-6 color palette, one memorable hero, and real copy. then say what to avoid (fake stats, a gradient kit, fade-up on every section). the longer landing prompts on this page go further.
the idea of a compact landing-page starter prompt is from @Numan_Ai12, 27 oct 2026. the wording below is new. the original post had 5 bookmarks and 299 views in the 12 nov 2026 fetch.
use this when the draft hero has a user count, a rating, or a logo wall you cannot source.
you get a hard rule for this page, and for any landing prompt you copy from it.
if you cannot source a number, a quote, or a logo, it does not go on the page. that is the same standard used to build this resource. the 'social-proof' prompt on this page is there so an agent will cut unverified claims instead of decorating them.
use this when you want a recent example of visual work attributed to fable 5.2, without turning it into a benchmark.
you get the 9 nov 2026 post, with its live x metrics.
the author says fable 5.2 at high reasoning produced an animated clip from a short brief, using an existing animation base. that is a demonstration, not an official visual score. it is here because the page is about ui/ux and visual work with this model, not because a winner was measured.
3d and visual · intermediate · 10 min if you open it
use this when you are doing stylized 2d/3d motion and want a public course-style post, not a product-ui kit.
you get the 27 oct 2026 post and article title, without repeating unverified view counts from the text.
the post describes a short cartoon format and a course of prompts, skills, and a review loop. it mentions opus 5.5 in the article title, not fable 5.2. a view count quoted inside the post was not independently checked here, so it is not repeated. use this as a motion-studio reference, and keep product chrome 2d if you mix this with an app ui.
from @Y0urDumble, 27 oct 2026. the attached article title, as returned by x, is "how to build a 2.5d motion studio with opus 5.5 (full course)".
x, 13 nov 2026, about 11:30 pm ist: 3,194 bookmarks, 1,011 likes, 1,005,701 views.
3d and visual · beginner · 2 min author comparison, not a benchmark
use this when you want a same-prompt visual anecdote, clearly labeled as one person's take.
you get the 12 nov 2026 post, with no invented winner.
the author ran a 'visually impressive 2.5d snake game in one html file' brief on fable 5.2 and gemini 4 carbon, and said carbon looked and felt better while fable was faster. that is one person's take on one brief. it is not a leaderboard.
use this when a scene is doing hero, decoration, and button work at the same time.
you get a short rule, aligned with the 3d prompts on this page.
if you use 3d, use it in one place. give it a still fallback and a reduced-motion path. keep the menus readable. a product ui should still work if the scene never loads. the 3d prompts on this page are written to that standard.
3d and visual · beginner · 2 min author comparison, not a benchmark
use this when you want the 11 nov 2026 iart note, without treating it as a scoreboard.
you get the post, clearly labeled as the author's run.
the author said they ran seven motion-graphics briefs through sonnet 5.5, opus 5.5, and fable 5.2 before shipping fable in iart. the public text was truncated in the fetch, and this page does not invent the outcome. open the post if you want the author's own conclusion.
use this when sliders and labels are inside the scene, and nobody can use them.
you get the same rule as the 'product chrome' prompt, as a reminder card.
the object can be 3d. the controls should stay 2d, labeled, and tappable. if a label only exists inside the scene, it will fail on contrast, on a phone, and with reduced motion.
use this when you are starting a ui repo and you want the first message to load skills, not adjectives.
paste this at the start of a ui thread. if a skill is not installed, install it or skip that line, do not hallucinate its rules.
you are the design lead on this repo. before you restyle or invent ui, load published skills, not adjectives.
preferred skills, if they are available in this environment:
- anthropic frontend-design, for art direction, type, and restraint
- emil kowalski's skills (especially emil-design-eng, animate, review-animations, gestures, learn-taste)
- impeccable, if we need a critique pass
then do this, in order:
1. name the product, the audience, and the job of this screen.
2. write a short plan: 4-6 colors, type roles, layout idea, motion policy (emil's frequency and easing rules).
3. review the plan against the brief. if a choice is the generic default you would produce for any similar product, change it and say why.
4. only then implement. finish typography, color, layout, empty/error/loading, motion, and a 390px width. do not stop at a punch list.
do not invent metrics, quotes, or logos. do not use ease-in on ui. do not enter from scale(0). reduced motion is gentler, not zero.
you get a paste-ready setup that names published skills and asks for a plan first.
written for this page. it leans on anthropic frontend-design and emil kowalski's published skills, in new wording.
use this when the layout is fine, and you only want a motion pass.
you are doing a motion pass only. do not restyle type or color unless a motion change requires it.
follow emil kowalski's published skills: find-animation-opportunities, animate, review-animations, gestures, animation-performance.
rules you must not break:
- 100+ times a day: no animation.
- ease-out to enter and exit, ease-in-out for movement, ease for hover and color, linear for loops. never ease-in on ui.
- published curves only. press is scale(0.97) at 160ms ease-out.
- never scale(0). start from 0.9 or higher plus opacity.
- gate hover. reduced motion is gentler, not zero.
- only one marketing entrance, if any. the product ui stays quiet.
first write a table of what should move. then implement. then review.
you get a short brief that loads emil's motion skills and forbids delight on daily actions.
written for this page, from emil kowalski's published motion rules.
use this when you are starting a marketing page and you want the brief to block the usual kit.
you are designing a landing page for this product. plan first, then build.
load anthropic frontend-design if it is available. write a short plan (color, type, layout, principles) and review it against the brief before you write layout code.
requirements:
- lead with the characteristic thing in this product's world, not a generic gradient hero.
- at most two typefaces, a 4-6 color palette, body lines under 80 characters.
- one memorable element. cut decoration that is not doing a job.
- no invented metrics, quotes, ratings, or logos. if proof cannot be sourced, omit it.
- one primary call to action, named as the action the person will get.
- no fade-and-slide-up on every section. respect prefers-reduced-motion.
- finish at 1280px and 390px, with real copy, not lorem.
if a choice is the default you would produce for any similar product, change it and say why.
you get a paste-ready setup that asks for a plan, real proof, and one hero moment.
written for this page, from frontend-design principles, in new wording.
use this when you want official effort guidance in a sentence you can paste.
use fable 5.2 at high effort for this ui pass (say medium instead if the work is small). finish typography, color, layout, empty/error/loading, motion, and a 390px width in the files. do not stop at a list of notes. do not invent metrics. follow emil kowalski's published motion rules if anything moves.
you get a one-liner that names effort, then asks for a finished ui pass.
effort defaults and the "finish the task" reminder are from anthropic, prompting claude fable 5.2. the ui scope is from this page.
use this when a first ui already exists, and you want a review against published skills, then fixes.
review this ui against published guidance, then fix it.
use anthropic frontend-design for art direction, type, structure, and restraint. use impeccable if it is available, for a slop and polish critique. use emil kowalski's review-animations (and gestures, if needed) for motion.
write a short review: what is specific to the brief, what looks like a generated default, what fails on a 390px screen, what fails reduced motion or contrast. then implement the fixes in the files. do not leave the review as the deliverable.
do not invent new brand metrics or quotes.
you get a prompt that reviews, then implements, instead of only commenting.
written for this page. it names frontend-design, impeccable, and emil's review-animations.
use this when the first view is good, and you do not want the next view to invent a new palette.
extract a small token set from the current ui: color, type, spacing, radius, and motion (use emil kowalski's published easing and duration values for motion). put it in one file. every new screen in this pass must use those tokens. if you need a new value, say why before you add it. do not grow a huge token system for a small product.
you get a short instruction to extract tokens and reuse them.
use this when you want the official prompting guide, even though this page is only about ui.
you get the primary source for effort, finishing the task, and the other official rules.
this page only keeps the rules that help a visual pass (calibrate effort, finish the work). the rest of the official guide is still worth reading if you use fable for more than ui. it is the source to trust over any thread.
use this when you want anthropic's own model-choice page, not a thread.
you get the official choosing-a-model doc.
read this alongside the short model card on this page. official docs still do not name a ui-specialist model. do not infer a design winner from a general task table.
use this when you want the official skill text, not only this page's paraphrase.
you get the skill file in anthropics/skills.
read the source if you are going to argue with a default look, or if you are writing your own design skill. the prompts on this page remix the ideas; they do not replace the file.
use this when you want emil's docs and the essay, not only the github tree.
you get the skill documentation site and the essay.
use the docs when you are installing or calling a skill. use the essay when you want the argument for giving an agent taste, instead of only a list of easing curves.