visual-assets
/visual-assets
Zamienia dekoracyjną grafikę, tło lub ilustrację w wyrazistą warstwę dla stron wejściowych, pustych stanów i ekranów marketingowych.
- Autor
- cdeweyx ↗
- Repozytorium
- conor-os ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 12.11.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 10
- Skrypty
- 2
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 12.11.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
- 2 skrypty w paczce. Przejrzyj je przed instalacją, bo Claude może je uruchomić.
Skrypty: scripts/fetch-glyph.py, scripts/linefield.py
Co robi i kiedy
Zamienia dekoracyjną grafikę, tło lub ilustrację w wyrazistą warstwę dla stron wejściowych, pustych stanów i ekranów marketingowych.
Wyzwalacze z opisu
Oryginalny opis (EN)
Turns a sketchy decorative visual (backdrop, hero graphic, illustration, collage, texture, generative field) into a striking layer the surface can use. Use when a surface is a front door rather than a working screen (home, empty state, marketing, onboarding), when the user says "boring", "cooler", "striking", "make this visual better", "illustration", "backdrop", "wallpaper", "hero visual", "collage", or when hifi-design opens a treatment axis. Do NOT use for charts (dataviz), UI layout and hierarchy (hifi-design), motion (motion-design), or brand identity (brand-design).
Jak działa
- Frame the job
- Research before drawing
- Explore on one axis
- Make it striking
- Critique, then stop
- Hand back
- Things that go wrong
- Skill pause
- Checklist
Podgląd SKILL.md
---
name: visual-assets
description: >
Turns a sketchy decorative visual (backdrop, hero graphic, illustration,
collage, texture, generative field) into a striking layer the surface can
use. Use when a surface is a front door rather than a working screen
(home, empty state, marketing, onboarding), when the user says "boring",
"cooler", "striking", "make this visual better", "illustration",
"backdrop", "wallpaper", "hero visual", "collage", or when hifi-design
opens a treatment axis. Do NOT use for charts (dataviz), UI layout and
hierarchy (hifi-design), motion (motion-design), or brand identity
(brand-design).
---
# Visual assets
A decorative visual earns its place by saying something the copy can't, in the product's own vocabulary, without competing with the job of the screen. The first draft of one is always sketchy, because the agent draws from memory and memory holds the average. What separates striking from competent is a loop: research the specific kind of visual before drawing it, explore on one axis, then spend on the techniques that are expensive but read as crafted. The loop is the skill; the fan, the collage, or the contour map is just this run's instance.
Read `references/preferences.md` every run, before the brief; it holds the user's picks from past rounds, and the first direction set starts from them. Read `references/techniques.md` before making anything (`references/explainers.md` too when the visual explains how something works) and `references/scorecard.md` before judging it. `references/sourcing.md` has the models, open sets, open-collection pictures, and licenses; `references/inspiration.md` has where to look and how to annotate what you find. `scripts/linefield.py` is a worked example of the generative path: seeded random walks rendered as compact SVG.
## Frame the job
Start from the surface, not the visual. Say in two sentences what the screen is for and whether it's a front door (home, empty, onboarding) or a working screen (results, tables, editors); a visual on a working screen stays quiet. Then name what the visual has to say (proof, breadth, the domain, the data, the payoff), the product's own vocabulary it can draw from (its chart forms, its objects, its numbers), and the constraints it inherits: DESIGN.md's palette and deny list (the user's ask for this piece outranks the Brand section, by the precedence line in brand-design's Brand section; say when you depart from it), the medium (Paper renders inline HTML and SVG, ignores stroke opacity and probably filters, and drops `color-mix()` fills silently, so derived colors go in as literal hex and every export is pixel-checked; the web can run shaders), and whether raster is on the table this round. Raster costs usage and tends toward slop, so it's off until the user says on.
When a surface is a front door in one state and a working screen in another, the brief names which one gets the treatment and says the other stays quiet; whether any element carries across (a strip, a mark) is decided at the states step, not here.
Write that as a one-paragraph brief and get a yes. Adjectives don't go in it; "striking" resolves to the default.
Then classify the visual before choosing a path, because the class decides the approach: a **glyph** (an object standing for a state: a whisk, a pot on a burner, a soufflé that fell) is sourced first and constructed only when the object is product-specific; a **field or backdrop** (walks, contours, dots) is generated by a seeded script; an **illustration** (a scene or a figure) is composed from an open set or primitives, never freehand; **type as the visual** is set, not drawn. Say the class in the brief. The sketchy first draft almost always comes from treating a glyph like a field and drawing it from memory.
Before the brief goes out, run the kill test: remove every decorative element in your head. If what's left still says the thing, or the field itself is the thing said (walks drawn from the product's own data), there's a concept; if it's a headline on a ground, the visual was wallpaper, and a field of dots or nodes behind it doesn't change that. For an illustration of a real thing (the product's device, a place, how a feature works), keep an evidence ledger beside the source, one row per feature: supplied, estimated (from a photo, and marked so), omitted, or invented and labeled as concept.
Empty and error states are front doors at a small scale and get the same loop shrunk down: a glyph, not a scene; one hue; it explains the cause (a soufflé that fell, a route that ends in nothing) rather than showing a sad object; it keeps the frame the results would have taken; and it sits centered with the sentence and the one action as a block. The first attempt is usually a miniature of the surface's big visual, and that reads as decoration; the domain's own objects read as the product. Stay inside the domain the user names.
## Research before drawing
Two lanes, previewed and confirmed before they run, because a visual type has its own craft literature and the agent doesn't hold it. Each lane is a `researcher` agent (Model: researcher role):
- Techniques for this visual type. Contour maps, line fields, halftone, ticket stubs, line-drawn objects, type-as-hero, each has a canon: the algorithm, the smoothing, the weight and lightness rules, the libraries, which models produce it well, and what makes it look cheap. Ask for rules an agent can apply, not moods.
- Inspiration and competitors. Who does this kind of visual well, in this category and in adjacent ones; what the incumbents do on the equivalent screen; the domain's own visual tradition (a map, a recipe card, a receipt) that the product can borrow legitimately. Mobbin's `search_sections` is the first stop for marketing and front-door visuals when the MCP is connected (hifi-design's reference pass has the install line and the report format); the galleries in `references/inspiration.md` and Playwright otherwise. The two or three that set the bar get read through explain-interface: the layers, their order, and the values doing the work.
Each state gets its own short lane (first-run empty, error, and zero results are different moments with different canons); a report written for one doesn't cover the next. Raw reports follow product-research's same-day rule.
Screenshot the public references with Playwright in one pass, tile them into a contact sheet, and annotate each with the technique it uses, not the feeling it gives: line weights, number of hues, density, structure versus randomness, negative space, what carries meaning. Distill to three principles and one move: the single decision the best references share. The principles and the move go into the brief.
## Explore on one axis
Directions differ in what the visual is, never in polish. Two fidelity tiers: layout ideas (a ghost grid, a backdrop, text only) can be sketched; object ideas must be shown with a sourced icon or a constructed glyph, never freehand paths, because a badly drawn whisk tests whether the agent can draw, not whether a whisk is the idea. The set leans toward what the user has picked before (see `references/preferences.md`) and spends its width on what's new for this surface; re-exploring a placement or a style the user already chose wastes the round. The brief has the same three parts as wireframe-explorations' brief: the job, why it's worth exploring instead of defaulting, each direction's assumption. Get a yes, then build the set once, wide (eight to twelve is fine when the user asked for range), every one on a duplicate of the locked surface with the form untouched and the treatment as one layer behind it. When the user says "try one", build one probe first; it calibrates the brief for a tenth of the cost, and the wide set follows if they want range. Get every idea on the table before anyone judges: the user's additions go in as directions, the ones they cut come out, and the research lane can add two or three of its own. One axis per round; a combination the user asks for ("the abstract one plus the strip") is the next axis, not a variant of this one. Vector only unless raster was approved. Name each with the surface, a letter, and the direction. Before the set reaches the user, walk hifi-design's steer-yourself checklist (`references/brief-and-review.md` there) on the real exports. Present with a lean and the reasons, then let the user pick, cut, or combine.
## Make it striking
This is where to spend. The cheap version of any visual is drawn from memory in one pass; the striking version is generated, rendered, or sourced with the technique the research named, then curated.
- Generative fields (walks, flow fields, contours, dot fields): a script, not hand-written paths. Seeded, so it's reproducible; simplified then smoothed to cubic curves; depth by pre-blended lightness at full opacity, never stacked transparency; three line weights at most with the hero at three times the field; one terminal marker with air around its exit; envelope curves for structure; the text zone kept clear by bending the walks under the form in the generator, not by a mask (Paper ignores SVG clipPath, and a ground-colored rect shows as a hard cut and a box behind the input); twenty seeds, pick one by eye, not by a number. A field that would ship as about a thousand DOM or SVG nodes or more renders to `canvas` instead (Interfere's dot pattern). `references/techniques.md` has the numbers.
- Glyphs for empty and error states: sourcing gate first. Search Iconify across every set (`scripts/fetch-glyph.py search <object>`), check coverage and license, and pull a stroke-based icon (Lucide and Lab, Tabler, Hugeicons, Iconoir) re-weighted to a true 1.5 to 2px line at the target height; baked-outline sets (Phosphor, Material, Font Awesome) only at 32px. Recolor to the token, record the license in the assets folder's ATTRIBUTIONS file. Construct a custom glyph only when the object is product-specific or nothing stroke-based exists, and then by the recipe in `references/techniques.md`: named anchors on a 4px grid, primitives and computed arcs only, one fixed stroke, render at 1x and 50%, check against the checklist, three passes, keep the script. One hue, a ground line, the cue that makes the object readable in three strokes (the handle on a mug, the tines on a fork, the flame under a pot). Composing two sourced parts (a pot set over a two-stroke flame) is usually the answer between sourcing and drawing.
- Illustration: primitives composed in three named layers, five hues, one stroke width; or an open set (Open Peeps, Humaaans) recolored to tokens instead of a figure drawn from scratch.
- Objects and collages (tickets, cards, stubs): real content from the product's data, one silhouette path plus one perforation line, tilt sparingly.
- Texture and atmosphere: Paper Shaders on the web, grain tested once on the canvas and replaced by a dot grid if filters don't render.
- Raster, only when approved: a style-anchor paragraph with fixed and fluid parts, hex-bound palette, ten candidates, a vision critique, duotone onto the paper color, a 2x upscale at most, alpha recovered from a white-and-black pair.
- Type as the visual: the question at display size, a rotating or cycling line of real examples, no picture at all. Often the answer for a monochrome product.
Every new color the visual needs becomes a token in DESIGN.md with its rule, so the asset doesn't carry hex codes the system doesn't know. Placement follows the job: the field in the lower third with the form above it, the strip at the edge, the glyph above its sentence. Centered, louder, quieter, and minimal variants of a placed field all lost to the placed one.
## Critique, then stop
Screenshot at 2x and run hifi-design's scored critic loop on the layer in place, not a loop of its own: three screenshot-only `critic` agents in fresh contexts (Model: critic role), `/impeccable critique` plus two scoring against `references/scorecard.md` for taste and DESIGN.md for compliance; the prompt is in hifi-design's `references/critics.md`. From inside the loop you stop seeing the composition, which is why the critics never see the brief or the seeds. Fix what fails, re-render, score again. Same cap and early stops as the surface; the fourth pass is taste drift. Log the rounds to the surface's `mockups/hifi/critique-<surface>.md`. Called from hifi-design or landing-page, its comparison pass or blind critique covers the layer. Run on its own, run hifi-design's comparison pass here on the surface with the layer in place, gold screens from this run's references, and take any finding about the layout to the user, since this skill owns only the layer. Keep the losing seeds and directions on the explorations page as the record.
## Hand back
This closes the treatment round; the surface's states are cloned after it, so the layer has to be final before they are. Port the pick into the state artboard yourself and clean the layout around it (center the block, fix the gap), then screenshot; the user's next comment is about the layout, not the visual, if you don't. Report which steps ran (frame, research, explore, make, critique, hand back) and which were skipped; "did we run the full skill?" is the question that follows a partial run. The surface owns the layer: place it at the back of the artboard under the locked form, name it with the seed and parameters ("<Surface>/<visual> (seed <n>, <parameters>)"), and write a hand-back note in the mockups folder (what it says, the script, the parameters, per-state behavior, the critique record) so the build can regenerate it. Export at 2x. Note in the spec what the visual says, how it's generated, and what changes per state (thin data, empty, error), since a decorative layer that ignores states is the first thing to break.
## Things that go wrong
- Drawing what a library already has. A whisk, a bell, a house exist in a dozen stroke sets; a hand-drawn one reads amateur next to them.
- A visual that competes with the form. The job of the screen wins every time.
- A raster hero. The tells are known and the usage is real; vector first, raster on request.
- A miniature of the big visual as the empty state. Field lines under "no recipes match" meant nothing; a soufflé that fell did.
- Presenting a scorecard fail as a defect when the user picked it. Their pick on the canvas wins; log the miss as deliberate. A named like counts the same: when a critic contradicts it, the like stays and the fix is softened.
## Skill pause
Run authoring-skills' Closing a skill update (the skill pause) at the end of every run. What goes where in this skill: a rule into the body or `references/techniques.md`, a pick into `references/preferences.md`, a scoring gap into `references/scorecard.md`.
## Checklist
- Brief written from the surface's job, with the medium and raster decision stated; user said yes.
- Visual classified (glyph, field, illustration, type), kill test passed, and the path chosen from the class; an evidence ledger for anything real; for glyphs, the Iconify sourcing check recorded before any drawing, licenses in ATTRIBUTIONS.md.
- Two research lanes previewed and run; references pulled from Mobbin `search_sections` or screenshotted, tiled, and annotated by technique; three principles and one move in the brief; raw research on the project's research page.
- Directions built once on duplicates, form untouched, one layer each, named; user picked or combined.
- The winner made with the named technique: generated or sourced, seeded, simplified and smoothed, lightness not opacity, three weights, twenty seeds and one pick.
- New colors added to DESIGN.md as tokens.
- SVG mechanical checks clean before the critics; scored through the shared critic loop (three screenshot-only critics, scorecard plus DESIGN.md, median), three rounds at most, logged to the surface's critique file; comparison pass run if this ran on its own.
- Layer placed at the back, named, seed and parameters recorded, exported at 2x, state behavior noted in the spec.
- The skill pause ran.
Pliki w paczce (10)
- SKILL.md 16 KB
- references/explainers.md 3 KB
- references/inspiration.md 4 KB
- references/learnings.md 1 KB
- references/preferences.md 3 KB
- references/scorecard.md 2 KB
- references/sourcing.md 6 KB
- references/techniques.md 10 KB
- scripts/fetch-glyph.py 5 KB
- scripts/linefield.py 6 KB
Wersje
| Repozytorium | Wynik | Licencja |
|---|---|---|
| Ta strona · cdeweyx/conor-os | 100 | MIT |
| KILN-skills/kiln | 100 | MIT |
| wlehd00000-cell/document-studio | 91 | MIT |
Migawka bazy z 13.11.2026 · wersja paczki 9e844971aab1 (skrót SHA-256 całej paczki)
Pobrania w ostatnich 30 dniach: 0 · ranking
Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.
Podobne skille
- 100brand-design
Zamienia minimalny plik DESIGN.md produktu w pełną markę: ideę, typ, kolor, znak i zasady ruchu, a potem wstawia ją do tokenów produktu. - 100hifi-design
Przenosi ustabilizowane szkice na ekrany wysokiej wierności w Paper przez MCP, od DESIGN.md do przekazania do budowy. - 100ideagram
Zamienia koncepcję, opis funkcji, wpis lub pitch w jedną spójną z marką ilustrację, dopasowując ją do prawdziwej ilustracji unDraw z lokalnej biblioteki i przekolorowując na kolor akcentu marki. - 100banana-web
Generuje zasoby do projektowania stron: tła hero, wizuale landing page, obrazy kart, separatory sekcji i klatki kluczowe motion design, z obsługą trybu ciemnego i jasnego. - 100brand-illustration-systems
Tworzy, rozbudowuje i audytuje ilustracje marki i ich systemy, proponując trzy kierunki wizualne, gdy kierunek nie jest ustalony. - 95frontend-dev
Łączy projekt UI, animacje, grafiki i teksty sprzedażowe przy budowie stron marketingowych, landing page i dashboardów.