Claude as a Creative Director:
Six Months of Real Workflows

In June 2025, I wrote a 600-word system prompt and sent Claude Sonnet 4.6 the same brief I would have handed a new junior creative director. I wasn't testing AI — I was testing whether my own thinking was clear enough to articulate.

The answer, it turned out, was both yes and no. Claude revealed the gaps in my own systems as much as it filled them. Six months later, here's an honest account of what changed, what stayed the same, and what I genuinely couldn't have anticipated.

The Brief I Gave Claude

The brief was structured exactly like the briefs I write for human creatives: a client overview, a creative challenge, three reference directions, constraints (budget, timeline, output format), and a clear definition of what "done" looks like.

What surprised me was what happened next. A junior CD would take the brief, go away, and come back with questions — the questions they ask reveal how well they understood the brief. Claude's first response was essentially a set of clarifying questions formatted as assumptions: "I'm assuming X, Y, and Z. If any of these are wrong, let me know before I proceed."

That response format — stated assumptions instead of open questions— turned out to be more useful than the open-question format I was used to. It showed the AI's reasoning before it acted, which meant I could course-correct before any work was done, not after.

Six Months In: What We Actually Did

Over the following six months, I used Claude in a formal "creative director" capacity across four categories of work:

• Brand audits — Analyzing client briefs and existing brand materials against a 12-attribute framework, identifying gaps, and producing positioning recommendations.
• Content direction — Developing editorial calendars, LinkedIn content strategies, and article outlines across NR's four content pillars.
• Creative briefs — Drafting client-facing briefs for Motion Co. projects based on intake calls (I provide the transcript; Claude drafts the brief).
• Voice and copy review — Running drafts through a brand voice scoring system with written feedback, not just a score.

What Actually Worked

Let me be direct: a lot of this worked remarkably well. More than I expected going in. The things that worked weren't the glamorous use cases people imagine — it wasn't Claude generating brilliant original creative from nothing. It was Claude doing the disciplined, systematic work that I always knew was important but didn't have enough time to do consistently.

What Failed (Or Required More Work Than Expected)

The failures were instructive. They mostly came from one source: I hadn't given Claude enough context about what I was trying to achieve, or I'd given it the wrong kind of context.

The second area of failure was anything that required judgment about what wasn'tsaid. Claude is excellent at analyzing and extending what exists in context. It's much weaker at identifying when something important is missing entirely. Brand strategy work, especially for clients with established-but-problematic positioning, still requires a human director who can see what the brand isn't saying.

Third: real creative inspiration. The moments in a pitch where you introduce an idea that nobody asked for and it changes the direction of the whole project — Claude doesn't generate those. What it does is give me more capacity to think about those moments by handling the systematic work that would otherwise fill my schedule.

The Evolving Paradigms

Six months of daily use has clarified how the relationship between a creative director and AI tools actually works in practice. There are three modes, and which one you're in determines what you should expect from the system.

Before vs After: The Numbers

I tracked time-on-task for six representative workflows before and after implementing the IO system. Here's what the data showed:

The System Prompt That Changed Everything

Everything above — the brand audits, the content, the briefs — runs through a single master system prompt. I've iterated on it over six months and it's now 580 words. The core structure has stayed the same. Here's the pattern (not the full prompt — that's in the IO Knowledge Base):

# WHO YOU ARE WRITING AS
You are writing as [Name], a [role] based in [location].
Your work focuses on [3 areas]. Your clients are [ICP description].

# VOICE ATTRIBUTES (pick 8-12)
- Direct without being blunt
- Specific and example-driven, not abstract
- Confident but not arrogant
- [Your specific attributes here...]

# WHAT YOU NEVER DO
- Never use the word "delve" or "leverage"
- Never open with "In today's..."
- [Your specific avoidances here...]

# OUTPUT FORMAT
For LinkedIn posts: Hook (1 line) → Body (4-6 lines) → CTA
For articles: Strong opener → 3-5 H2 sections → Close
[Specify your preferred formats here...]

# CONTEXT (injected dynamically)
Current brand docs: [Notion MCP retrieval]
Recent outputs: [From Supabase]

Tools & Editors

The technical stack is simpler than it looks. Everything runs on HTML/CSS/JS pages that make API calls to Claude. No frameworks, no build steps for most tools. The ones that need slightly more (the Portfolio Builder, the Style Tastemaker) use vanilla JavaScript that most creatives with basic coding exposure can read and modify.

The development environment is almost entirely Claude.ai itself — I describe what I want, Claude generates the code, I test it in a browser. For a creative director with no formal engineering background, this is the most important thing to understand: you don't need to write code from scratch to build these tools.

Team & Scaling

I built this system for myself, but I've started thinking about how it scales to Motion Co. as a team. The honest answer is that scaling a personal brand OS to a studio OS is a different problem. The architecture is similar but the data layer gets more complex — you need client-specific context that doesn't contaminate other clients, role-based access to different tools, and a review workflow that accounts for multiple people's voices.

Deploy Options

Everything in the NR IO system is deployed on GitHub Pages, with a Cloudflare CDN in front. Zero infrastructure cost, near-instant global delivery, and the deploy workflow is a single GitHub push.

What's Next

Six months in, I'm more committed to this approach than I was when I started — not because it's perfect, but because it keeps improving. Every month, the system prompts get more calibrated, the tools get more connected, and the outputs get closer to what I would have produced myself.

The next major evolution is real-time context awareness — tools that surface relevant knowledge base entries as you work, without you having to ask. I'm prototyping this now and the early results are promising.