Art Director

The Art Director is the senior designer who directs fleets of AI agents instead of executing every detail by hand. In practice that means you write a CLAUDE.md file that reads like the best design brief you have ever received. It spells out your voice your tokens your components exactly how naming works and what clean work looks like in your codebase. Then you hand goals to Claude Code the same way you used to hand specs to a junior designer three years ago. The agent explores the repo opens files runs tests and opens PRs. You review the output with an art director eye. You say this focus ring is too heavy or these three variants are redundant or the hero section needs more breathing room. Your value sits entirely in taste standards and review rigor not in syntax or Figma skills. The designer who shipped a full design system refactor in three days last month did it by becoming this kind of Art Director. She never wrote a line of real code. She simply refused to accept sloppy output from the model and updated her CLAUDE.md with every lesson so the next run got sharper.

The Art Director is not a fancy title for someone who likes prompting AI image generators on the side. It is not the designer who still manually hunts down every hardcoded 16px value across ten repositories every quarter. It is not a passive observer who accepts whatever the model ships on the first try without critical feedback. It is not someone who cannot read a git diff or explain why a particular border radius feels dated in 2025. It is not a replacement for human illustration work that requires soul or for product strategy calls that require deep user empathy. If you treat the agent like an autocomplete tool instead of a long running junior with its own initiative you are not an Art Director. You are just a slower more expensive engineer with better taste and a terminal tab you do not understand.

Take the concrete example of Mia at FinFlow in February 2025. Their design system had ballooned after two years of feature velocity. It included seventeen variations of their primary Card component. Half of them were unused leftovers from a 2023 campaign page. Mia spent one focused hour writing a rich CLAUDE.md. She listed every existing component with one line descriptions so the agent would not reinvent them. She defined forbidden patterns like inline styles or magic numbers for spacing or px values anywhere in the codebase. She wrote what clean work looked like in this codebase: small focused PRs that update the Storybook stories when components change include before after screenshots for any visual update and never introduce new dependencies without asking first. Then she gave Claude Code this goal. Audit every Card usage across our monorepo including the eight packages in our Turborepo setup. Identify which variants are dead code. Propose a streamlined API with a feature flag for the deprecated paths. Implement it cleanly. Update all stories and tests so nothing breaks. The agent returned a detailed plan in plain English within minutes. Mia approved with one note about preserving a specific hover interaction from their 2024 brand guidelines that lived in their tokens package. Forty seven minutes later she had a perfect PR that passed typecheck linting and visual tests. The work that used to eat an entire sprint for two people now took an afternoon. She repeated this process for their Button Hero Nav and DataTable components over the next ten days. The system went from bloated to crisp and the whole team felt the difference in velocity.

A second example comes from the marketing site refresh at Vercel in late 2024. Their Art Director wired up the Figma MCP server so Claude Code could read exact frame measurements layer names and connected design token links directly instead of guessing from a screenshot. She gave it the new hero design and said build this using only our existing components from the design system package. Do not invent new ones. Match the spacing tokens exactly. Respect our voice guidelines listed in the CLAUDE.md. The first pass was close but the typography scale felt off by one weight and the padding around the CTA felt tight. She replied with precise feedback referencing their exact type scale document and the spacing rules already written in the brief. The second PR matched pixel for pixel and passed every test. What normally took three days of back and forth with engineering took four hours total. The designer spent the saved time on higher order problems like whether the new layout actually converted better in user testing.

Use the Art Director role when your foundation is solid and you want to move fast on repetitive high leverage work that used to drain design ops capacity. Fire it up for design system maintenance sweeps that replace old spacing values like 8 16 and 24 pixels with the proper --space-2 --space-4 and --space-6 tokens across every repo. Deploy it when you need to consolidate overgrown component APIs like the twelve variant Button problem that every design system develops by year three. Bring it in for Figma to code workflows where you want production ready PRs that respect your existing system instead of throwaway prototypes from tools like v0 or Lovable. Call on it for brand token propagation when marketing updates the primary blue in Q3 and you need it everywhere from the docs site to the electron app without introducing regressions or missed instances. Run it for visual reviews that catch overflow bugs and spacing inconsistencies your eye misses after the tenth page because the agent can spin up the dev server take screenshots and compare them systematically. These are the five workflows where it earns its keep every single week and compounds in value as your CLAUDE.md gets smarter with every lesson you feed it.

Do not use it when the work requires original taste that cannot be fully briefed in text or when the output demands human emotional intelligence the model simply lacks in 2025. Real brand illustration falls squarely in this bucket. If you need a new mascot character that feels exactly like the hand drawn figures you launched with in 2022 then brief a human illustrator at Prismlight or Runesmith instead. The agent will produce something generic even with perfect prompts. Skip it for purely subjective layout calls where the question is not whether the grid aligns but whether the composition feels exciting or emotionally boring. The model can implement a correct layout but it cannot feel that the hero needs more air or that the weight decision makes the whole page feel corporate and safe. Never use it for first principles product decisions like whether that new dashboard should exist in the app at all or if the feature addresses a real user pain. It will happily build whatever you describe even if the entire direction is wrong. That judgment call is yours alone and it is what you actually get paid for. And avoid it entirely if your codebase is a disaster zone with no tokens no consistent naming conventions no tests and no CLAUDE.md file. The agent will simply accelerate your existing chaos and generate technical debt at lightspeed. Fix the foundation first or you will spend the next month cleaning up AI generated spaghetti code.

The junior is not a person anymore. The junior is a model with unlimited patience and zero ego. You are the Art Director.