mcpField
A
art-director

art-director

github.com/hcaianoDesign & UI54 installs
Repository

About art-director

A design skill for exploring visual directions with a user and selecting one before implementation. It structures intake, image review, refinement, and handoff.

Key features

  1. 01Inspect existing brand and design context
  2. 02Interview to establish design constraints
  3. 03Develop three to five visual directions
  4. 04Generate and curate comparative images
  5. 05Refine a selected direction
  6. 06Test and hand off the visual system

Use cases

  1. 01Explore visual concepts for a new product
  2. 02Choose a direction within an existing brand
  3. 03Refine a design using image variants
  4. 04Validate a direction on an adjacent surface

Install

Copy the command and run it in a terminal.

$npx skills add https://github.com/hcaiano/skills --skill art-director

After installing, type/art-directorin Claude Code to confirm it loaded.

Bundled filesagents/openai.yaml

Details

From the project README

Art Director

Explore visual directions, curate them with the user, and converge on one defensible choice. The deliverable is the chosen direction; implementation starts after this skill ends.

Image generation

In Codex, use the built-in imagegen skill: its generation, editing, validation, and save-path rules own the image work. Anywhere else, delegate each generation brief with delegate_task to a Codex provider instance chosen with t3-capacity. The delegate generates with imagegen, always saves the images in this workspace, and returns their absolute paths and exact prompts. Each brief stands alone: absolute reference image paths, the save directory, and for a refinement the prior exact prompt and the one trait to change. Open every returned path before curating it.

Run intake as a grilling conversation: investigate facts yourself, ask decisions one at a time with your recommended answer, and wait for the user's reply.

Workflow

  1. Inspect. Determine whether this is blank-slate or constrained by an existing brand. Read available references, product context, design tokens, and existing surfaces before asking questions.
  2. Intake. Resolve the audience, surface, communication goal, required content, format, references, anti-references, hard brand constraints, and decision deadline. Ask only what materially changes the exploration.
  3. Frame directions. Name three to five concrete visual premises. Each direction needs a different underlying rule for composition, hierarchy, imagery, and mood. "More options" is not a direction.
  4. Generate breadth. Generate one image per direction. Keep the functional content and format constant so the visual ideas are comparable. Preserve an established identity; for blank-slate work, explore a genuine range. Keep the exact prompt for every output.
  5. Curate. Inspect every image. Identify it by filename or stable ID and explain what works, what fails, and which reusable trait it contributes. Maintain a shortlist and a rejected-pattern list. Ask the user for the next decision one question at a time.
  6. Refine. Generate two or three variants of the strongest direction. Change one named trait at a time instead of remixing everything.
  7. Prove the system. Stress-test the winner on an adjacent page, state, or asset. For identity work, generate one coherent system board from the selected direction: mark, palette, type, and one applied surface.
  8. Hand off. Return the selected asset paths, exact prompts, visual rules, rejected patterns, and remaining trade-offs. Save project-bound outputs in the workspace according to imagegen; preview-only outputs may remain inline.

Stop condition

Stop when the user has selected a direction and it survives one relevant system proof. Do not start an unattended loop, create arbitrary batches of dozens of images, or create persistent BRIEF.md or curation files unless the user asks for them or the exploration belongs in the project.

Comments

Markdown supported