Building a Second Brain for My Creative AI Workflow

AI didn't speed up my creative workflow until I stopped trapping my context inside it.

My biggest unlock recently wasn't a new model. It was a folder structure that holds my creative brain.

Every tool had its own memory. I'd build meaningful context with Claude, then open Perplexity or Codex and start from zero. My best brainstorming was scattered across chat histories I could never find again. The tools were powerful, but I was the bottleneck, re-explaining myself every single session.

As a creative, I switch tools constantly based on the problem. Claude for heavy thinking and writing, Codex for coding, DeepSeek for research, OpenDesign or GPT Image 2 for asset generation. Each one is great at its job. But bouncing between them is a pain when every tool forgets who I am the second I switch.

The asset side made it worse. Getting AI to generate work that doesn't look like slop takes real direction, references, and a point of view. That context is the most valuable thing I have, and it was the first thing lost every time I changed tools.

There was a practical reason too. With token prices climbing on the major models, I didn't want to be locked into one provider just because that's where all my context lived. So I built a memory layer that lives with me, not inside any one platform.

I didn't invent this from scratch. I pulled from a few places and then adapted hard:

Karpathy's thinking on personal AI. The idea that your context is an asset worth owning, separate from any model.

The YC conversations on personal "second brains." Treating durable memory as infrastructure for working with AI.

The PARA method. Organizing by Projects, Areas, Resources, and Archives.

A lot of trial and error. The parts that survived are the parts I actually used.

It wasn't a one-shot solution. At first I over-structured it. Too many folders, too many rules, all built before any of it had earned its place. It looked organized and was miserable to maintain. The structure was serving the diagram, not the creative work. Version two got simpler as it got more useful.

It splits into a root operating layer and a brain.

The key rule: the brain is canonical, everything else is temporary or downstream. Chats are disposable. The brain is not.

Separate active memory from raw source. 01_Source_Material is evidence. Everything a tool reads day to day is distilled context, not raw exports or half-baked drafts. This one split fixed most of my early mess, because the tools stopped reading the wrong things.

Context packs for portability. The packs in 09_Context_Packs are what let me swap tools freely. When I move from writing in Claude to generating assets in GPT Image 2, the direction comes with me. Same taste, same references, same point of view. The tools change. The creative brain behind them doesn't. That's what broke the single-tool lock-in, and it's why the output stops looking like generic AI slop.

Here's the part I care about most. Building this didn't make me less of a designer. It did the opposite. My voice, my taste, and my judgment are now the layer that steers every tool I touch. The models are powerful, but they work inside my structure, on my direction.

That's the whole game. As these tools advance, everything starts to look like the same regurgitated thing. The creatives who keep their voice and learn to direct this stuff won't be replaced. They'll have more range than they've ever had.

We don't lose our role in product development to these tools. We keep our voice and use them as leverage. The folder structure is small. The point underneath it is not: own your context, direct the tools, and you gain range instead of losing your seat.

Still a work in progress. But it works noticeably better than the first version, and the difference shows up in the output. The AI sounds like me, remembers my projects, and follows my rules without me babysitting it.

The real lesson: the tools didn't replace my judgment. They got better the moment I gave them structure to work inside. AI as creative infrastructure, not a shortcut. The point of view, the curation, and the decision about what's worth keeping are still the human job. The model just needed somewhere to put it.