In June 2025, I wrote a 600-word system prompt and gave Claude the same brief I would have handed a new junior creative director. I was not testing AI. I was testing whether my own thinking was clear enough to articulate to a machine.
The Brief I Actually Wrote
The brief was structured exactly the way I write briefs for human creatives: client overview, creative challenge, three reference directions, constraints around budget and timeline and output format, and a clear definition of what done looks like. Nothing fancy. The same format I have used for eight years at Intelligent Operations
What surprised me was Claude's first response. A junior CD would take the brief, disappear for a day, and come back with questions. Claude came back instantly with stated assumptions: "I am assuming X, Y, and Z. If any of these are wrong, let me know before I proceed." That format turned out to be more useful than open questions. It showed me exactly where Claude might go wrong before any work was done.
What Actually Worked
Brand voice consistency at scale. Claude maintains tone consistency across long documents better than I do at 11pm after a full shoot day. The key was a 400-word voice document with specific examples of what my brand sounds like and equally important, what it does not sound like. Without those negative examples, the outputs were competent but generic.
Intake to brief in under ten minutes. A 30-minute client intake call transcript turned into a full creative brief in under ten minutes. The brief quality was consistently good and honestly better than the briefs I used to write in 45 minutes after a call because I was tired and rushing to get to the next task.
Content repurposing without losing the thread. One of my chronic problems was having a great long-form observation that I never turned into LinkedIn content. Claude's repurposing approach, extracting three angles each with a different hook type, captures the idea before it fades from memory.
What Failed
The first three weeks of AI-generated LinkedIn posts all sounded the same. Not generic exactly. They sounded like a competent version of me, but not the version I actually want to project. The fix was not a better prompt. It was reading my own best posts and annotating what made them work, then encoding those annotations into the system prompt.
The second failure was anything that required judgment about what was not said. Claude excels at analyzing and extending what exists in context. It is 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 the gaps.
The System That Emerged
Six months in, I use Claude in three modes. About 80 percent of my usage is augmentation: AI accelerating work I would have done anyway like drafting, structuring, and reformatting. About 15 percent is delegation: tasks I have defined well enough to hand off entirely like brand audits and intake-to-brief. And about 5 percent is genuine collaboration: multi-turn conversations where I am discovering something through the exchange.
The net result is roughly a 40 percent reduction in time spent on non-creative work. Some of that saved time went back into the system itself, building better tools and writing better prompts. Most of it went into the client work and creative thinking that I was not making enough time for before. That shift, from operational overhead to creative capacity, is the real outcome of this experiment.