Co-Creative Loops
Framework
One loop per job.
Between May 2025 and August 2026 I had 12,000 interactions with an LLM while making a musical, its music, a website, a subway installation, and a film. This framework is what survived: a way to see exactly where the machine was in the work — and to keep your authorship where you want it.
01How a Loop Works
Each project has multiple loops. Each loop corresponds to one job within a project. A musical carries loops for writing, design, music, choreography, film, publishing, direction. Each loop breaks its job into three phases — Generation, Feedback, and Iteration — and each phase has a ratio of human vs. AI involvement.
Size — the averaged ratios set the size of a loop; the bigger the loop, the more human. Speed — how many times the job’s output was iterated across the project’s timeline; more iterations, faster loop. Drag the ratios below and watch both.
Every loop cycles the same three phases. Here’s just what each one is — the principles that guide them come in section 03.
making original work — ideas and first drafts, your authorship behind it
gaining insight into what you've created
changing existing work on new insight; planning what's left to make or cut
02My Three Projects
Each row is one project; its loops run along the shared production pipeline — Pre → Production → Post, read left to right. The same design loop runs fast and AI-heavy inside the website, slow and hand-made inside the subway piece. Click a project to open its full map.
The Techno-Apocalypse Musical
Thesis Site
Hacking the Subway
accent = human · grey = AI · halves = Generation | Feedback · outer ring = Iteration · disc size = human share · hub = iterations. 24 of 24 loops specified — the rest render as [TODO] until their ratios are filled in (lib/loops.js). Nothing is estimated.
0313 Co-Creative Principles
making original work — ideas and first drafts, your authorship behind it
- 1.Begin with the AI-last approach
- 2.Your honest answer is more important than AI's correct answer
- 3.Treat yourself like a generative algorithm
- 4.Always cite the models you use
- 5.Learn the tools of the craft, but try building your own tools
gaining insight into what you've created
- 1.Human–AI overlapping signals tend to be critical
- 2.Use both coordinate-based & coordinate-less feedback
- 3.Novelty is not the metric — use AI to appreciate the obvious
- 4.AI emotional support isn't strange
changing existing work on new insight; planning what's left to make or cut
- 1.Personalize your scale of an iterative loop
- 2.If you don't know how it works, you're in a bad place
- 3.Become an expert in cross-skilling & agent management
- 4.Decide what is the dirty work
Rules that guide you — and break — in each phase. Full principles →
04Automation vs Personal Growth
Looking across the loops, you might ask why I chose to automate some jobs more than others. It comes down to two things: how much personal growth I wanted in a domain — where it felt meaningful to do the work myself and keep authorship — and how capable I thought AI was at the job without me.
The horizontal axis is how it actually turned out — left is automated, right is human-authored — and the vertical axis is how much personal growth I wanted. Jobs I cared to grow in sit up top, and mostly over to the right where I kept my hands on them; the bottom-left is what I was happy to automate. Each dot’s size is how capable AI was at that job — bigger = more capable — so a big dot on the left is something I automated because the model could carry it. A hollow dot means I haven’t rated its capability yet.
05My Year in Data
Read across the built artifacts, the loop signatures diverge — the website a knot of large fast loops, the mini-musical a few small slow ones, the music in between. The same loops recur at different settings, and that comparison is the clearest thing the data does that prose alone could not. See the year →