CommunityImagegithub.com

artist

Use when making or directing visual assets such as illustrations, hero images, icons, logos, brand marks, textures or diagrams, or writing prompts for an image-generation model. Covers art direction, reference boards and critique against a brief. Not for page layout; use `designer`. Not for palettes and tokens; use `visual-design`.

What is artist?

artist is a Claude Code agent skill that use when making or directing visual assets such as illustrations, hero images, icons, logos, brand marks, textures or diagrams, or writing prompts for an image-generation model. Covers art direction, reference boards and critique against a brief. Not for page layout; use `designer`. Not for palettes and tokens; use `visual-design`.

Works with~Claude Code~Codex CLI~Cursor
npx skills add https://github.com/heliacon/claude-skills/tree/HEAD/skills/artist

Installed? Explore more Image skills: steipete/songsee, affaan-m/frontend-design, mattpocock/improve-codebase-architecture · View all 6 →

Ask in your favorite AI

Open a new chat with this agent skill pre-loaded.

Documentation

What does artist do?

Making the images a design needs, and making them belong to it. The designer skill decides where an image goes and what job it does. The visual-design skill owns the palette and type the images must agree with. This skill directs the asset: the brief, the references, the tool, the prompts, the critique and the delivery.

The rules that matter most

  1. Write the brief before generating anything. Job, medium, palette, light, mood, composition, what it must not look like. A prompt with no brief behind it produces the model's average image.
  2. Choose the tool for the job first. Generate, draw in a vector tool, render from code, or use a photograph. An image model is the wrong tool for exact geometry and text.
  3. References set the ground, palette and light, not a texture to paste. Read what a reference shows (its colour range, its light direction, its contrast, its density) and write those down as brief values. Copying a surface pattern from it misses the point.
  4. Critique every output against the written brief, item by item, not against how impressive it looks. A striking image that breaks the brief is a reject.
  5. Keep an iteration log. Prompt, settings, seed, output file, verdict, and what to change next. Without it you cannot reproduce the good one or explain the bad one.
  6. The person who owns the brand gives the art direction. Propose options and name the trade-offs. Do not invent a house style and present it as settled.

When not to use

  • Deciding page layout, image placement or crop on a page: designer.
  • Choosing brand colours, type or tokens: visual-design.
  • Charts and data graphics: the frontend skill's references/charts.md. Their colours encode data.
  • Purely decorative code art (generative patterns in a canvas): treat it as code, with its own render and look loop.

Choose the tool

AssetDefault toolWhy
Mood, hero, spot illustration, textureImage-generation modelFast exploration of look and light
Icon set, logo, wordmarkVector tool, explored first with the image modelNeeds exact geometry, a shared grid and clean scaling
DiagramRendered from code or a diagram toolNeeds exact labels and editable structure
Real people, real products, real placesPhotographyGenerated images of real things mislead
Text inside an imageSet it in HTML or a vector toolImage models garble letterforms

Prefer a generated, drawn or photographed image for illustration. Hand-written SVG or ASCII art rarely matches their quality next to them. Keep hand-written SVG for geometry (icons on a grid, diagrams, charts).

The brief

Job:          what the asset must do (set a mood, explain a structure, mark a brand)
Where:        placement, size and crop on the page; what sits on top of it
Medium:       photographic | illustrative | abstract | diagrammatic
Palette:      2 to 4 values taken from the system's tokens, with proportions
Light:        direction, quality (hard or soft), time of day, contrast range
Mood:         three words, and three words it must not be
References:   2 to 5 images, each with what to take from it
Composition:  focal point, negative space for text, aspect ratio
Avoid:        the look it would drift to by default
Done when:    the checks it must pass (see Critique)

Reference boards

  • Collect 2 to 5 references. For each, write the one thing to take ("the low, warm side light", "the large flat shapes with no outline").
  • Extract measurable values: dominant colours (sample them), light direction, value range (how dark the darks, how bright the lights), density of detail.
  • A reference that shows a quality you cannot name is not yet useful. Name it or drop it.

Prompting an image model

  • Order: subject, style, composition, palette, light, finish. Specific nouns beat adjectives. "A brass pressure gauge on a workbench" beats "a beautiful industrial scene".
  • Composition for use. If text sits on top, ask for a calm, low-detail region in the right place ("plain dark upper third"). For icons, ask for a centred subject on a flat ground.
  • One variable at a time. Fix the seed and change one clause, so you know what each change did. Change the seed only to explore.
  • Sets share a style clause. Write the style clause once and reuse it verbatim; vary only the subject. If the model supports a reference image or style input, use the same one for every item.
  • Generate several, choose a direction, then refine. The first batch is for choosing, not for shipping.
  • Avoid asking for: readable text, exact logos, hands holding small objects precisely, tables, pixel-exact alignment, counts of objects above about five. Do those in code or a vector tool and composite.
  • Check the model's licence and terms for commercial use before the asset ships.

Iconography

  • One grid, one stroke weight, one corner radius across the set. A set on a shared grid reads as a system; a mixed set reads as clip art.
  • One clear metaphor per icon. If it needs a caption to be understood, change the metaphor.
  • Balance optically, not mathematically. A circle must be slightly larger than a square to look the same size.
  • Test at the size it ships at, on light and dark.

Brand marks

  • Simple enough to draw from memory.
  • Works in one colour, and in pure black and white, before any gradient.
  • Test at 16px (favicon) and at hero size. If it fails at either end, it is not the mark.

Diagrams

  • One idea per diagram. Two ideas, two diagrams.
  • The shape carries the argument: a fan for one-to-many, a line for sequence, convergence for aggregation. Remove the labels and ask whether the shape still says something. If not, the layout is arbitrary.
  • Use the real names (real endpoints, real event names, real fields), never "Input, Process, Output".
  • One visual language: the same shape always means the same kind of thing; colour means something or is not used.
  • Render, look, fix: check for clipped text, overlaps, crossed arrows and uneven spacing. Two to four passes is normal.

Critique

Score each output against the brief, line by line:

  • Palette within the brief's values (sample the image, do not eyeball).
  • Light direction and contrast range match the references.
  • Composition leaves the space the layout needs, at the real crop and size.
  • In a set: grid, weight and style match its siblings when placed side by side.
  • Legible at the size it will ship at.
  • No artefacts: extra fingers, melted text, broken symmetry, seams in tiles.
  • Earns its place. If removing it loses nothing, remove it.

A measure (a palette distance, a sharpness score) checks one property and can reward the wrong thing. Place the asset in the real page and look.

Delivery

  • Format: photographs and rich illustration as AVIF or WebP with a JPEG fallback; flat art and icons as SVG; PNG only when transparency and raster are both needed.
  • Export at the sizes used (1x and 2x), and compress to the smallest file that still looks right.
  • Write alt text that says what the image communicates in context. Decorative images get an empty alt="".
  • Record the source of each asset (prompt and seed, or file, or photographer and licence).

Gotchas

  • "Not like that" is not a direction. When feedback says what the image must not be, ask or propose what it should be before regenerating. Several different targets satisfy the same "not".
  • Prompting harder to fix a wrong tool. Ten attempts at a crisp logo from an image model is a sign to switch to a vector tool, not to add adjectives.
  • A filter is not a mark. Asked for visible brush strokes or layered paint, a smooth image with a texture overlay is not the same thing. Each stroke is a placed object.
  • Faking a real object from primitives. A real shape traced or referenced from its source beats a plausible approximation made of circles and noise.
  • A set made one image at a time with a new prompt each time. It will not match. Fix the style clause and the reference first.
  • Judging the asset alone. An image that looks good full-screen can fail at the 320px crop behind a headline. Judge it in place.

Verification

  • Written brief exists and every delivered asset maps to it.
  • Iteration log records prompt, settings, seed and verdict for the chosen output.
  • Critique checklist passed with the asset placed in the real page, both themes.
  • Exported at shipping sizes and formats; file sizes noted.
  • Alt text written; source and licence recorded.

Individual skills in this repo

This repo contains 4 individual skills — each has its own dedicated page.

Related Skills