AI has not replaced craft, but it has changed where the hours go. I work alone across creative direction, design, motion and full-stack engineering, so I have had to get very practical about where AI genuinely saves time and where it quietly makes work worse. This is the workflow I actually use, with the rules and checklists that keep quality where clients expect it.
The real problem AI solves for a one-person studio
When one person directs the visuals and also builds the product, the hardest part is not any single skill. It is switching between them. In one afternoon I might move from refining a logo in Illustrator to a motion test in After Effects to a database change in a Next.js project. Every switch costs time to reload context.
That is where AI earns its place for me. It is very good at shortening the distance between "I know roughly what I need" and "I have a working first draft". It is not good at deciding what is worth making, what feels right, or what will still hold up in a year. Those decisions are the job.
So my rule is simple: AI drafts, I decide.
Where AI genuinely helps
Exploration and ideation
- Generating many divergent directions quickly: naming options, taglines, campaign angles, layout ideas described in words.
- Turning a vague brief into sharper questions I should ask the client.
- Rough mood and lighting references to talk through a direction before anyone draws anything.
The value here is quantity. I ask for twenty options knowing I will throw away eighteen. The two that survive usually get rewritten, but they start a better conversation than a blank page would.
Engineering boilerplate
- Scaffolding repetitive code: form validation schemas, TypeScript types from an API response, test skeletons, config files.
- Explaining unfamiliar code, or an error message from a library I rarely touch.
- First drafts of one-off scripts: a data migration, a bulk image rename, a regular expression I would otherwise spend ten minutes on.
- Drafting documentation, READMEs and commit messages from changes I have already made.
Motion and post-production
Much of the AI in motion work is not a chatbot at all. It is machine-learning features now built into mainstream tools: assisted rotoscoping and masking, content-aware fill for removing objects, smarter tracking, upscaling and noise reduction. These remove hours of mechanical work, and I treat them like any other tool: a first pass that I check frame by frame.
Where I keep AI away from the final call
Brand marks and identity
A logo has to be distinctive, ownable and drawn with intent. Generated imagery tends towards the average of everything it has seen, which is the opposite of distinctive, and the legal ownership of purely generated images is still unsettled in many places. I may use AI to explore a mood, but every mark I deliver is built by hand as clean vector artwork.
Security-sensitive code
Authentication, payments, permissions, file uploads and anything that touches user input get reviewed line by line, whoever wrote the first draft. AI-generated code is often plausible, confident and subtly wrong, which is the most dangerous combination in security.
Performance-critical paths
Shaders, animation loops and anything that runs every frame. Assistants happily produce code that works but allocates memory on every frame, or recalculates something that could be cached. On a site with live WebGL effects, that is the difference between smooth and stuttering on a mid-range phone.
Voice and final copy
AI can suggest structure, but the words a client publishes should sound like a person, ideally the client. I edit every line for tone and strip out the filler phrases assistants love.
My six-step workflow
- Write the brief by hand first. Before I prompt anything, I describe the problem in my own words: who it is for, what success looks like, what the constraints are. If I cannot write that clearly, no tool will rescue the project.
- Diverge with AI. I ask for many options, not one answer. Variety is the point at this stage.
- Converge with judgement. I filter options against the brief, not against what looks impressive. Most get cut.
- Build with assistance, then own it. Generated code gets read, renamed to match the project's conventions, and trimmed. If I cannot explain a line, it does not ship.
- Verify as if a stranger wrote it. Type checks, tests, a Lighthouse run, and testing on real devices, not just my development machine.
- Polish by hand. The last ten per cent, such as easing curves, kerning, the timing of a transition or a single line of microcopy, is where craft lives, and it is still entirely human work.
A review checklist for AI-generated code
I run through this before merging anything an assistant helped write:
- Does it handle failure? Empty states, network errors, slow responses, invalid input.
- Did it add dependencies? A new package for something the platform already does is a cost, not a convenience.
- Is the API real and current? Assistants sometimes invent functions or reach for deprecated patterns, especially in fast-moving frameworks like Next.js.
- Did it move work to the browser? In the App Router, one unnecessary client component can pull a whole tree into the JavaScript bundle.
- Is it accessible? Labels, focus states, keyboard support, and respect for reduced-motion settings.
- Are secrets safe? No keys in client code, no sensitive data in logs or URLs.
- Does it match the codebase? Naming, folder structure, error handling. Consistency is what keeps a project maintainable a year later.
Treat generated visuals as sketches, not deliverables
For visual work, I treat generated images the way an art director treats thumbnail sketches. They are useful for:
- agreeing on a mood, palette or lighting direction with a client early;
- storyboarding a sequence before committing to animation;
- exploring texture and material ideas for 3D scenes.
Then the real asset gets made properly: vector artwork, a lit and rendered 3D scene, or footage composited by hand. I am also upfront with clients about what was generated and what was made, because trust is worth more than a shortcut.
What this means if you are hiring a designer or developer
If you are commissioning creative or technical work, a few things are worth knowing:
- Ask how they use AI. A good answer is specific about which tasks and which checks. "I never touch it" and "I use it for everything" are both warning signs.
- Drafts are cheap now; judgement is not. What you are really paying for is direction, taste and accountability for the result.
- Speed should buy quality. Time saved on boilerplate should go into more exploration, more testing and better polish, not just a faster, thinner delivery.
This is how I approach both sides of my work, from creative direction to full-stack web development: AI as a capable assistant, with a human responsible for every decision that reaches the user.
Key takeaways
- AI is most valuable for reducing context-switching: first drafts, boilerplate, explanations and exploration.
- Keep identity design, security-sensitive code, performance-critical code and final copy under direct human control.
- Write the brief yourself before prompting, ask for many options, then filter hard against the brief.
- Review AI-generated code with a checklist covering failure handling, dependencies, real APIs, bundle impact, accessibility and secrets.
- Use generated visuals as sketches, build the real asset properly, and be transparent about which is which.
Planning a project? Book a free discovery call — I'll tell you honestly what it needs.