Taste
Left alone, a model builds the average of the web. Every choice in it is plausible, and none of them
was actually decided. This skill makes the decisions explicitly and grounds them in two things the
average doesn't have: the subject's own material, and measured numbers from real work. Then it looks
at the result next to its reference.
Four beliefs behind every step:
The user's taste outranks every rule here. They are the only judge of whether it is
distinctive. In this skill's first blind test, the owner chose a richer no-skill page over the
skill's own sparser build, even though it broke several "rules". Ask what they love before
deciding anything.
Familiar structure, exact surface. People judge a page in about 50ms, and they like it when it
looks simple and typical of its kind. Novelty helps only while the page still reads as typical.
Keep the skeleton people expect, and spend novelty on the surface.
Taste is calibration, not avoidance. linear.app uses Inter, pills, gradients and borders, and
it still looks expensive because its values are precise:
- grounds 1–7 RGB steps apart
- four ink greys at 18.7 / 13.6 / 6.1 / 3.5:1
- weight 510
- −0.022em tracking on the headline
- a spacing jump from 12 to 32px
Copy relationships between values, not ingredients.
Say what to do, with numbers, and make it fast by construction. Naming what to avoid primes
it, and performance retrofitted at the end means deleting effects. Every decision below is written
positively, and its cost is chosen when it is planned. Then the page is seen next to its
reference, by someone who didn't build it.
<skill> is this folder. Working files go in design/ at the project root. The scripts need Node 22+
and Chrome, Chromium, Edge or Brave (set CHROME_PATH if it isn't found).
Small edits: the fast path
When design/DESIGN.md already exists and the request is a small edit to that design (copy, a
link, an icon, a colour, one component, one section's spacing), skip steps 1–4:
- Use the plan's tokens as they are. A new value is allowed only if the plan has no fitting one;
add it to the plan.
- Build the change (step 5's rules still apply).
- Lint, and run the probe at 1440 and 390 (step 6).
- Screenshot the changed area at both widths, with the previous version beside it, and compare them
yourself. Run the critic (step 7) only if the change touches the first screen or the hierarchy.
- Add one line to DESIGN.md: the date, the request, what changed.
Not a small edit, so run the full process: a new page or section, a change of direction, palette,
type or layout, or anything the user calls a redesign. If you are unsure, ask the user which one
they mean.
1. Ground the brief
Start design/DESIGN.md with:
The user's taste, first. If the user is there, ask in one message:
- Which 2–3 sites or apps do you love?
- Light or dark?
- Anything you already know you want or hate?
Record the answers verbatim; they outrank everything below. If design/critiques.md exists, read
it too.
Subject, audience, job. The job is one of persuade, operate, read or experience. Add the one
action the page exists for.
Mode. Mobile-led (consumer, retail, local, single-action pages, traffic from social or QR),
desktop-led (dashboards, pro tools, B2B and developer sites), or both (portfolios, mixed
marketing sites). Every mode is complete and comfortable on a phone; the mode only decides which
composition leads. Real analytics, if the user has them, beat this table.
Inventory. 6–10 things that belong to this subject and to no competitor: its name and what the
name means, its place, materials, numbers, constraints, rituals, the printed things of its world,
its real data.
Real content. Write the actual copy and data now, at full density. Use sample details where the
brief has gaps, and list them for the user to confirm. Proof is never invented: no customers,
quotes, logos or business metrics.
If the brief carries nothing specific to the subject, ask at most two more questions, in the same
message as the taste questions. Otherwise state your reading in one line and continue.
2. Measure a reference
The style lives in numbers that a screenshot doesn't carry.
- The user named sites they love: these are the references. For each, run
node <skill>/scripts/probe.mjs <url> --out design/refs/<slug>.json.
- No URL: open
<skill>/library/README.md. Pick the 1–2 profiles whose notes match this job and
tone, then read each tokens.json and notes.md.
- Only an image: write your estimates (grounds, inks, families, sizes, spacing) as numbers in
DESIGN.md and mark them as estimates for the user to correct.
For each reference, write one line saying what you borrow (its ink staircase, spacing rhythm, type
calibration) and what you leave (its subject, imagery and composition). Borrow calibration, never the
whole layout.
3. Diverge, then drop the most likely
First fix the skeleton: the structure people expect from this kind of page, such as a nav, a clear
first screen, the work, about and contact for a portfolio, or a list with filters for an app screen.
It is not up for divergence. Directions differ on the surface instead: type, the one visual, the one
motion, colour, and how the idea shows.
List five directions. For each, write one line plus your estimate of how likely it is that you would
have produced it for any similar brief. Discard the two most likely. From the rest, pick the one that
best carries an item from the inventory and fits the user's taste. Then write the idea sentence:
On this page, [something only this subject has] becomes [the device].
Then run the competitor test: put the nearest competitor's name in the sentence. If it still works,
it is a category idea. Go back to the inventory.
If the user asks for a minimal, simple or text-first page, or loves one (emilkowal.ski,
Apple-like), read references/restraint.md now. It says when restraint is enough, and which
few values must then be exact.
4. Write the token plan before any code
Add these to DESIGN.md, all as numbers, derived from the reference profile:
- Grounds: 1–3 of them, and how surfaces separate. Use a tone step (give its size), or hairlines
(give their alpha).
- Inks: a 3–4 step staircase, each with its contrast ratio on the main ground. Body text uses the
strongest or second ink and is at least 4.5:1.
- Accent: one hue (two at most), and the places it appears: the action, plus at most one more.
- Type:
- Families: at most two; one is normal.
- Weights: at most three.
- Sizes: 3–5 in use, with one decisive jump. Display is at least 3× body on pages and at least 1.6×
on app screens. On a phone, 2–3× body (about 32–44px at a 16px body), and the longest word
fits.
- Tracking and leading for each size. Large type gets tighter tracking (−0.01 to −0.04em) and
leading (0.95–1.15).
- Body: 15–20px on pages, 13–15px on app screens, line-height 1.4–1.6, lines of 45–75 characters.
On a phone, running text is at least 16px (lines of 30–60 characters), nothing is under 12px,
and weight 300 or lighter only appears at 20px and up.
- Spacing: a set of 5–7 values. The gap between groups is at least 2× the gap inside a group.
- Radii: one rule (square, small or full), decided by hierarchy. Never one radius on everything.
- Boldness: the one place it is spent: a scale jump, a field of colour, an image, a live element.
Everything else stays quiet.
Then review the plan line by line. Would you have written that line for any similar brief? If so,
revise it from the reference or the inventory, and note the change. Read design/history.md if it
exists, and don't repeat a past build's direction.
The budgets above are defaults. A measured reference overrides them (Linear runs four inks); write
down the override and its source. Where neither the reference nor the user decides something (dark
surfaces, button sizes, tracking at each size, hover and focus, motion timing), take the value from
references/defaults.md. Don't invent one.
Finally, write a cost line for every planned effect: which rendering stage it touches (composite,
paint or layout), and the cheap way to get it from references/performance.md. The signature
moment gets its own line, and the performance budget applies to it.
5. Build
- Tokens first, as CSS custom properties or a Tailwind theme. Components reference tokens only.
No value outside the plan.
- Fast from the first line:
- CSS compiled at build time, never a runtime compiler.
- Every
<img> has width and height.
- The hero image is never lazy and gets
fetchpriority="high".
- At most 2 font families, in WOFF2, with variable axes pinned to what is used.
- No script in
<head> without defer or type=module.
- The first screen first, at 1440 and 390 wide, until it is right. Then each section, each one
carrying the idea. Neighbouring sections never share a layout. A section exists only if the content
has a claim for it.
- Two compositions, one priority. 390 and 1440 share the content and its order; each is composed
for its width. The desktop uses the width (columns, a side element, a wider figure), never a
phone column floating in the middle. The phone gets one column of text, never two side by side.
- The phone layout: read
references/mobile.md before building it: the first screen, the one
primary action, visible navigation, forms, touch and comfort.
- Rich, not sparse. Every screen carries real content and enough parts to reward a look: meta,
figures, real fragments of the work. Group things to lower complexity. Don't delete them. Empty
space is spent once per page at most, on purpose.
- Every data view ships its states:
- loading: a skeleton shaped like the result
- first use: what this is, plus the next action
- filtered to nothing
- error: what happened, how to fix it, input kept
- sparse and dense
- Interaction:
- a real button for the action, labelled verb + noun
:focus-visible styles
- targets of at least 24px, and 44px or more on touch (bigger at the screen edges)
<a> for navigation
tabular-nums on figures
- Components: borrow the behaviour, own the look. Dialogs, menus, selects, comboboxes, tabs,
tooltips, toasts and date pickers come from native HTML or a tested headless primitive, never
hand-rolled. They are styled only with the plan's tokens, and a kit's default theme never ships.
Which primitive to use for each pattern and stack:
references/components.md.
- Product fragments: a product fragment shown as a picture (a hero mockup of the app) gets
data-ui and aria-hidden="true". That keeps it out of the contrast and target checks. Everything
a person can really use stays in.
- Motion: none yet. Get it right standing still.
For app screens (dashboards, settings, lists, editors), read references/craft.md before building.
6. Lint
node <skill>/scripts/lint.mjs <source files or folder>
node <skill>/scripts/probe.mjs <url or index.html> --check
node <skill>/scripts/probe.mjs <url or index.html> --check --width 390
node <skill>/scripts/probe.mjs <url or index.html> --check --width 360
The phone runs add their own checks: running text under 16px, form fields under 16px, touch targets
under 44px, a first screen that ends in empty space, a buried h1, centred paragraphs, thin small
text, overflowing headings and disabled zoom.
--check runs three passes:
- Design: contrast, targets, type and colour budgets.
- Cost: the page loaded at 4× CPU slowdown, checking LCP, CLS, long tasks, render-blocking
scripts, script and font bytes, blur and non-composited animations.
- Reduced motion: a reload with
prefers-reduced-motion on, which must leave nothing moving.
Fix every ERROR. A WARN stays only if DESIGN.md says why the reference, the subject or the user
earns it. The gate measures defects, not liking. A clean gate is the floor, not the verdict.
7. Look
- Shoot. Run
node <skill>/scripts/shoot.mjs <page> --ref <reference url> --out design/shots/r1. It captures the
page and the reference at 390, 1024 and 1440 wide, plus a blurred copy for the hierarchy test.
- Critique. Spawn a critic subagent with a fresh context, one that did not build the page. Give
it
<skill>/references/critic.md, the shots folder and DESIGN.md. If design/critiques.md
exists, give it that file too: it holds the user's own example critiques and outranks the rubric.
The critic returns ranked issues and one priority target. If your agent cannot run subagents, do
it as a separate pass: re-read references/critic.md, look only at the shots and DESIGN.md (not
the code), and write the critique before changing anything.
- Fix. Fix the priority target and at most two more issues, without breaking what already works.
- Compare. Shoot again into
r2. The critic compares the new version with the best so far,
pairwise. Keep the new version only if it is strictly better; otherwise revert it.
Run one round by default. Run a second only if round 1 found a P0. Never run more than three. Ship the
best version, not the last one.
8. Motion and finish
Only now add motion, and keep it within these limits:
- Required: motion that explains a change: things opening, closing, moving or confirming. Also
press feedback, and tracking an item from a list into its detail.
- Optional: one signature moment per page.
- Never: on keyboard-triggered or high-frequency actions. No scroll-jacking, and no parallax on
text.
- How:
- Platform first: CSS transitions,
@starting-style, same-document View Transitions.
Scroll-driven animation only inside @supports.
- JS only where the platform can't, and per-frame values never go through React state.
transform and opacity only. Timings from references/defaults.md: 70–150ms micro,
150–250ms small, 200–350ms modals, exits shorter.
- Ease-out, never from
scale(0).
- Always: a reduced-motion fallback (a cross-fade). Phone motion and comfort:
references/mobile.md.
Use the animate skill for anything beyond that. Then look in the mirror and remove one thing.
9. Hand over
Tell the user, briefly:
- the reference and what you borrowed from it
- the idea sentence
- the plan in five lines
- the sample content to confirm
- what the critic still flags
- the runner-up direction from step 3, in one line
Then append to design/history.md: the date, direction, grounds and accent, families, and what the
user said. The user is the judge of whether it is distinctive. Ask for their verdict in one question.
When it starts to look generic
references/tells.md lists the defaults that generated pages fall into, and pairs each one with the
move that replaces it. Read it during critique, not before planning. Tells are warnings, never
targets to minimise. One is allowed when DESIGN.md says why the reference, the subject or the user
earns it. If the user loves mono labels, they're in.
Files
| Path |
Use |
library/ |
Measured reference profiles (tokens.json + notes.md), indexed in library/README.md |
references/critic.md |
Brief for the critic (a subagent, or a separate pass) |
references/defaults.md |
Values for type, buttons, states, colour, dark mode and motion when nothing else decides |
references/performance.md |
Budget, the cheap way to get each costly effect, platform feature readiness |
references/craft.md |
App-screen rules: states, feedback, density, copy |
references/components.md |
Native elements and headless primitives per pattern and stack; the token → component layering |
references/tells.md |
Default patterns paired with their replacements |
references/mobile.md |
Phone layout, touch and comfort rules for the 390px build |
references/restraint.md |
Minimal pages: when restraint is enough, the exact few values, the critic's checks |
../../reports/Mobile first UI taste.md |
The evidence behind the phone rules, with sources and confidence (in this repo only) |
scripts/probe.mjs |
Measures a live page into a profile; --check adds the design, cost and reduced-motion passes |
scripts/shoot.mjs |
Screenshots of page and reference at three widths, plus a blur copy |
scripts/lint.mjs |
Static checks on source for defects code can prove |
1---2name: taste3description: >-4 Art direction and visual taste for any web UI: landing pages, product and marketing sites,5 portfolios, app screens, dashboards, settings, lists. Grounds the design in the subject, measures6 real reference sites into numbers, writes a token plan before code, lints, and runs a screenshot7 critique loop against the reference. Use whenever you build or design a new web page or screen in8 any stack (HTML/CSS, React, Next.js, Tailwind, Vue, Svelte, Astro), when new output looks generic,9 templated or AI-made, and for small edits to a design this skill planned (design/DESIGN.md). To10 redesign, restyle, polish or review a UI that already exists, use taste-redesign. For motion11 polish after the layout is settled, use the animate or review-animations skills.12metadata:13 source: https://skillmd.com/skills/leocanela279/taste14---1516# Taste1718Left alone, a model builds the average of the web. Every choice in it is plausible, and none of them19was actually decided. This skill makes the decisions explicitly and grounds them in two things the20average doesn't have: the subject's own material, and measured numbers from real work. Then it looks21at the result next to its reference.2223Four beliefs behind every step:2425- **The user's taste outranks every rule here.** They are the only judge of whether it is26 distinctive. In this skill's first blind test, the owner chose a richer no-skill page over the27 skill's own sparser build, even though it broke several "rules". Ask what they love before28 deciding anything.29- **Familiar structure, exact surface.** People judge a page in about 50ms, and they like it when it30 looks simple and typical of its kind. Novelty helps only while the page still reads as typical.31 Keep the skeleton people expect, and spend novelty on the surface.3233- **Taste is calibration, not avoidance.** linear.app uses Inter, pills, gradients and borders, and34 it still looks expensive because its values are precise:35 - grounds 1–7 RGB steps apart36 - four ink greys at 18.7 / 13.6 / 6.1 / 3.5:137 - weight 51038 - −0.022em tracking on the headline39 - a spacing jump from 12 to 32px4041 Copy relationships between values, not ingredients.42- **Say what to do, with numbers, and make it fast by construction.** Naming what to avoid primes43 it, and performance retrofitted at the end means deleting effects. Every decision below is written44 positively, and its cost is chosen when it is planned. Then the page is seen next to its45 reference, by someone who didn't build it.4647`<skill>` is this folder. Working files go in `design/` at the project root. The scripts need Node 22+48and Chrome, Chromium, Edge or Brave (set `CHROME_PATH` if it isn't found).4950## Small edits: the fast path5152When `design/DESIGN.md` already exists and the request is a small edit to that design (copy, a53link, an icon, a colour, one component, one section's spacing), skip steps 1–4:54551. Use the plan's tokens as they are. A new value is allowed only if the plan has no fitting one;56 add it to the plan.572. Build the change (step 5's rules still apply).583. Lint, and run the probe at 1440 and 390 (step 6).594. Screenshot the changed area at both widths, with the previous version beside it, and compare them60 yourself. Run the critic (step 7) only if the change touches the first screen or the hierarchy.615. Add one line to DESIGN.md: the date, the request, what changed.6263Not a small edit, so run the full process: a new page or section, a change of direction, palette,64type or layout, or anything the user calls a redesign. If you are unsure, ask the user which one65they mean.6667## 1. Ground the brief6869Start `design/DESIGN.md` with:7071- **The user's taste, first.** If the user is there, ask in one message:72 - Which 2–3 sites or apps do you love?73 - Light or dark?74 - Anything you already know you want or hate?7576 Record the answers verbatim; they outrank everything below. If `design/critiques.md` exists, read77 it too.78- **Subject, audience, job.** The job is one of persuade, operate, read or experience. Add the one79 action the page exists for.80- **Mode.** Mobile-led (consumer, retail, local, single-action pages, traffic from social or QR),81 desktop-led (dashboards, pro tools, B2B and developer sites), or both (portfolios, mixed82 marketing sites). Every mode is complete and comfortable on a phone; the mode only decides which83 composition leads. Real analytics, if the user has them, beat this table.84- **Inventory.** 6–10 things that belong to this subject and to no competitor: its name and what the85 name means, its place, materials, numbers, constraints, rituals, the printed things of its world,86 its real data.87- **Real content.** Write the actual copy and data now, at full density. Use sample details where the88 brief has gaps, and list them for the user to confirm. Proof is never invented: no customers,89 quotes, logos or business metrics.9091If the brief carries nothing specific to the subject, ask at most two more questions, in the same92message as the taste questions. Otherwise state your reading in one line and continue.9394## 2. Measure a reference9596The style lives in numbers that a screenshot doesn't carry.9798- **The user named sites they love:** these are the references. For each, run99 `node <skill>/scripts/probe.mjs <url> --out design/refs/<slug>.json`.100- **No URL:** open `<skill>/library/README.md`. Pick the 1–2 profiles whose notes match this job and101 tone, then read each `tokens.json` and `notes.md`.102- **Only an image:** write your estimates (grounds, inks, families, sizes, spacing) as numbers in103 DESIGN.md and mark them as estimates for the user to correct.104105For each reference, write one line saying what you borrow (its ink staircase, spacing rhythm, type106calibration) and what you leave (its subject, imagery and composition). Borrow calibration, never the107whole layout.108109## 3. Diverge, then drop the most likely110111First fix the skeleton: the structure people expect from this kind of page, such as a nav, a clear112first screen, the work, about and contact for a portfolio, or a list with filters for an app screen.113It is not up for divergence. Directions differ on the surface instead: type, the one visual, the one114motion, colour, and how the idea shows.115116List five directions. For each, write one line plus your estimate of how likely it is that you would117have produced it for *any* similar brief. Discard the two most likely. From the rest, pick the one that118best carries an item from the inventory and fits the user's taste. Then write the idea sentence:119120> On this page, **[something only this subject has]** becomes **[the device]**.121122Then run the competitor test: put the nearest competitor's name in the sentence. If it still works,123it is a category idea. Go back to the inventory.124125If the user asks for a minimal, simple or text-first page, or loves one (emilkowal.ski,126Apple-like), read `references/restraint.md` now. It says when restraint is enough, and which127few values must then be exact.128129## 4. Write the token plan before any code130131Add these to DESIGN.md, all as numbers, derived from the reference profile:132133- **Grounds:** 1–3 of them, and how surfaces separate. Use a tone step (give its size), or hairlines134 (give their alpha).135- **Inks:** a 3–4 step staircase, each with its contrast ratio on the main ground. Body text uses the136 strongest or second ink and is at least 4.5:1.137- **Accent:** one hue (two at most), and the places it appears: the action, plus at most one more.138- **Type:**139 - Families: at most two; one is normal.140 - Weights: at most three.141 - Sizes: 3–5 in use, with one decisive jump. Display is at least 3× body on pages and at least 1.6×142 on app screens. On a phone, 2–3× body (about 32–44px at a 16px body), and the longest word143 fits.144 - Tracking and leading for each size. Large type gets tighter tracking (−0.01 to −0.04em) and145 leading (0.95–1.15).146 - Body: 15–20px on pages, 13–15px on app screens, line-height 1.4–1.6, lines of 45–75 characters.147 On a phone, running text is at least 16px (lines of 30–60 characters), nothing is under 12px,148 and weight 300 or lighter only appears at 20px and up.149- **Spacing:** a set of 5–7 values. The gap between groups is at least 2× the gap inside a group.150- **Radii:** one rule (square, small or full), decided by hierarchy. Never one radius on everything.151- **Boldness:** the one place it is spent: a scale jump, a field of colour, an image, a live element.152 Everything else stays quiet.153154Then review the plan line by line. Would you have written that line for any similar brief? If so,155revise it from the reference or the inventory, and note the change. Read `design/history.md` if it156exists, and don't repeat a past build's direction.157158The budgets above are defaults. A measured reference overrides them (Linear runs four inks); write159down the override and its source. Where neither the reference nor the user decides something (dark160surfaces, button sizes, tracking at each size, hover and focus, motion timing), take the value from161`references/defaults.md`. Don't invent one.162163Finally, write a **cost line for every planned effect**: which rendering stage it touches (composite,164paint or layout), and the cheap way to get it from `references/performance.md`. The signature165moment gets its own line, and the performance budget applies to it.166167## 5. Build168169- **Tokens first,** as CSS custom properties or a Tailwind theme. Components reference tokens only.170 No value outside the plan.171- **Fast from the first line:**172 - CSS compiled at build time, never a runtime compiler.173 - Every `<img>` has `width` and `height`.174 - The hero image is never lazy and gets `fetchpriority="high"`.175 - At most 2 font families, in WOFF2, with variable axes pinned to what is used.176 - No script in `<head>` without `defer` or `type=module`.177- **The first screen first,** at 1440 and 390 wide, until it is right. Then each section, each one178 carrying the idea. Neighbouring sections never share a layout. A section exists only if the content179 has a claim for it.180- **Two compositions, one priority.** 390 and 1440 share the content and its order; each is composed181 for its width. The desktop uses the width (columns, a side element, a wider figure), never a182 phone column floating in the middle. The phone gets one column of text, never two side by side.183- **The phone layout:** read `references/mobile.md` before building it: the first screen, the one184 primary action, visible navigation, forms, touch and comfort.185- **Rich, not sparse.** Every screen carries real content and enough parts to reward a look: meta,186 figures, real fragments of the work. Group things to lower complexity. Don't delete them. Empty187 space is spent once per page at most, on purpose.188- **Every data view ships its states:**189 - loading: a skeleton shaped like the result190 - first use: what this is, plus the next action191 - filtered to nothing192 - error: what happened, how to fix it, input kept193 - sparse and dense194- **Interaction:**195 - a real button for the action, labelled verb + noun196 - `:focus-visible` styles197 - targets of at least 24px, and 44px or more on touch (bigger at the screen edges)198 - `<a>` for navigation199 - `tabular-nums` on figures200- **Components: borrow the behaviour, own the look.** Dialogs, menus, selects, comboboxes, tabs,201 tooltips, toasts and date pickers come from native HTML or a tested headless primitive, never202 hand-rolled. They are styled only with the plan's tokens, and a kit's default theme never ships.203 Which primitive to use for each pattern and stack: `references/components.md`.204- **Product fragments:** a product fragment shown as a picture (a hero mockup of the app) gets205 `data-ui` and `aria-hidden="true"`. That keeps it out of the contrast and target checks. Everything206 a person can really use stays in.207- **Motion:** none yet. Get it right standing still.208209For app screens (dashboards, settings, lists, editors), read `references/craft.md` before building.210211## 6. Lint212213```bash214node <skill>/scripts/lint.mjs <source files or folder>215node <skill>/scripts/probe.mjs <url or index.html> --check216node <skill>/scripts/probe.mjs <url or index.html> --check --width 390217node <skill>/scripts/probe.mjs <url or index.html> --check --width 360218```219220The phone runs add their own checks: running text under 16px, form fields under 16px, touch targets221under 44px, a first screen that ends in empty space, a buried h1, centred paragraphs, thin small222text, overflowing headings and disabled zoom.223224`--check` runs three passes:225- **Design:** contrast, targets, type and colour budgets.226- **Cost:** the page loaded at 4× CPU slowdown, checking LCP, CLS, long tasks, render-blocking227 scripts, script and font bytes, blur and non-composited animations.228- **Reduced motion:** a reload with `prefers-reduced-motion` on, which must leave nothing moving.229230Fix every `ERROR`. A `WARN` stays only if DESIGN.md says why the reference, the subject or the user231earns it. The gate measures defects, not liking. A clean gate is the floor, not the verdict.232233## 7. Look2342351. **Shoot.** Run236 `node <skill>/scripts/shoot.mjs <page> --ref <reference url> --out design/shots/r1`. It captures the237 page and the reference at 390, 1024 and 1440 wide, plus a blurred copy for the hierarchy test.2382. **Critique.** Spawn a critic subagent with a fresh context, one that did not build the page. Give239 it `<skill>/references/critic.md`, the shots folder and DESIGN.md. If `design/critiques.md`240 exists, give it that file too: it holds the user's own example critiques and outranks the rubric.241 The critic returns ranked issues and one priority target. If your agent cannot run subagents, do242 it as a separate pass: re-read `references/critic.md`, look only at the shots and DESIGN.md (not243 the code), and write the critique before changing anything.2443. **Fix.** Fix the priority target and at most two more issues, without breaking what already works.2454. **Compare.** Shoot again into `r2`. The critic compares the new version with the best so far,246 pairwise. Keep the new version only if it is strictly better; otherwise revert it.247248Run one round by default. Run a second only if round 1 found a P0. Never run more than three. Ship the249best version, not the last one.250251## 8. Motion and finish252253Only now add motion, and keep it within these limits:254255- **Required:** motion that explains a change: things opening, closing, moving or confirming. Also256 press feedback, and tracking an item from a list into its detail.257- **Optional:** one signature moment per page.258- **Never:** on keyboard-triggered or high-frequency actions. No scroll-jacking, and no parallax on259 text.260- **How:**261 - Platform first: CSS transitions, `@starting-style`, same-document View Transitions.262 Scroll-driven animation only inside `@supports`.263 - JS only where the platform can't, and per-frame values never go through React state.264 - `transform` and `opacity` only. Timings from `references/defaults.md`: 70–150ms micro,265 150–250ms small, 200–350ms modals, exits shorter.266 - Ease-out, never from `scale(0)`.267- **Always:** a reduced-motion fallback (a cross-fade). Phone motion and comfort:268 `references/mobile.md`.269270Use the `animate` skill for anything beyond that. Then look in the mirror and remove one thing.271272## 9. Hand over273274Tell the user, briefly:275276- the reference and what you borrowed from it277- the idea sentence278- the plan in five lines279- the sample content to confirm280- what the critic still flags281- the runner-up direction from step 3, in one line282283Then append to `design/history.md`: the date, direction, grounds and accent, families, and what the284user said. The user is the judge of whether it is distinctive. Ask for their verdict in one question.285286## When it starts to look generic287288`references/tells.md` lists the defaults that generated pages fall into, and pairs each one with the289move that replaces it. Read it during critique, not before planning. Tells are warnings, never290targets to minimise. One is allowed when DESIGN.md says why the reference, the subject or the user291earns it. If the user loves mono labels, they're in.292293## Files294295| Path | Use |296| --- | --- |297| `library/` | Measured reference profiles (`tokens.json` + `notes.md`), indexed in `library/README.md` |298| `references/critic.md` | Brief for the critic (a subagent, or a separate pass) |299| `references/defaults.md` | Values for type, buttons, states, colour, dark mode and motion when nothing else decides |300| `references/performance.md` | Budget, the cheap way to get each costly effect, platform feature readiness |301| `references/craft.md` | App-screen rules: states, feedback, density, copy |302| `references/components.md` | Native elements and headless primitives per pattern and stack; the token → component layering |303| `references/tells.md` | Default patterns paired with their replacements |304| `references/mobile.md` | Phone layout, touch and comfort rules for the 390px build |305| `references/restraint.md` | Minimal pages: when restraint is enough, the exact few values, the critic's checks |306| `../../reports/Mobile first UI taste.md` | The evidence behind the phone rules, with sources and confidence (in this repo only) |307| `scripts/probe.mjs` | Measures a live page into a profile; `--check` adds the design, cost and reduced-motion passes |308| `scripts/shoot.mjs` | Screenshots of page and reference at three widths, plus a blur copy |309| `scripts/lint.mjs` | Static checks on source for defects code can prove |