Creative Assembly Loops
Loop v1.6 · 4 shipped ·
One person, AI agents and a codified process, making small games in hours and publishing the clock.
What if we rebuilt Denki's creative assembly lines as agentic loops?
An ongoing project to remake how playthings get made, one loop at a time. The loop is the product; the playthings are the evidence.
Play 002b Piggy in the Middle →A remix of a remix: Pass Master became 002 Rondo, then Rondo got a new look because it was struggling to find the fun in its basic form. On your phone or desktop.
01The bet
The goal: tiny teams making playthings to a Denki standard in hours, not weeks, with agents filling the gaps.
Denki solved creative assembly lines 20 years ago. A team of three (Product Architect, Coder and Artist), working inside a network of context, standards and process, took six to ten weeks to make a Denki-quality product for interactive television.
My bet is that with an agentic exoskeleton I can make a Denki-quality plaything in hours, inside one coherent, consistent production process: Creative Assembly Loops.
The Product Architect role (scout, shaper, showrunner, editor, storyteller) is the part agents can't do, and it's the part I've done for 30 years.
02Scoreboard
Every plaything is an experiment on the loop. Each one adds a module to the process, and the clock is published with it.
| Plaything | Seed to ship | Human time |
|---|---|---|
| 001Pass Master | 4 days | untimed |
| 002Rondo | 1h 42m | 1h 02m |
| 002bPiggy in the Middle | 43m | ~19m |
| 001bPass Master (Unity) | 2h 21m | ~32m |
| Next: Amp! | – | – |
All times in hours and minutes. Seed to ship is elapsed time on the clock; human time is my share of it, the rest is Claude's. Remixes and ports run from kick-off, because they reuse a shipped game's design. ~ marks an estimate from timestamps.
Success signals
- Seed-to-ship time falls loop over loop, even as modules add complexity. Too early. One full loop has been timed so far (002 Rondo, 1h 42m). Remixes and ports aren't like-for-like.
- Amp! holds its own in a blind comparison with the sealed original Denki prototype. Not yet run. Amp! is up next.
- My hands-on time shrinks loop over loop while my judgement carries further: fewer rounds to each sign-off, and nothing signed off gets rebuilt. Early sign. Human time was 1h 02m on the last full loop, and under 35m on the remix and the port. Nothing signed off on 002 Rondo had to be rebuilt.
Failure signals
- Five loops in, ship times aren't falling.
- Five loops in, quality is dropping as speed rises or as modules are added.
- Five loops in, my hands-on time isn't shrinking, or it is but my judgement isn't carrying: sign-offs take more rounds, or signed-off work gets rebuilt.
If any of these fire, I'll change the experiment rather than push harder.
03How it works
A human-AI co-production loop for petit playthings: tiny, instantly playable browser games. Small enough to make in an afternoon, finished enough to play on your phone, satisfying enough to snack on. Each loop ships one thin slice, a polished first playable, and leaves the process better than it found it.
| Stage | Proves |
|---|---|
| 0 · Seed | The intent, and everything risky or unknown about it |
| 1 · Interactive Spec | Risky assumptions proven or killed in toys, before anything is built |
| 2 · Play 1.0 | Core play feels good on the target device, with placeholders |
| 3 · Game 1.0 | The rules and structure work |
| 4 · Thin Slice | It looks, sounds and feels finished |
| 5 · Ship | The public build works live |
| Postscript | What we learned, written up, and fed back into the loop |
Loop v1.6, correct as of 12 November 2026 (the process updates with each release). A remix or port of a shipped game takes a fast track: Seed, Play 1.0, then Ship.
- Me
- Vision, direction, shaping, balancing, playtesting, editing and sign-offs.
- Claude
- Questions, specs, toys, builds, tuning, testing and deployment.
Front-load risk, back-load polish.
Nothing moves without a sign-off.
Every post-mortem iterates and evolves the loop.
04Why
Making is cheap now, so the constraint has moved: choosing what to make, proving it early and polishing it properly. Denki solved that 20 years ago with process, not headcount: creative assembly lines that took games from idea to live on Sky in six to ten weeks (TV as a Games Platform). Agents let a tiny team run that process again as augmentation, not automation, in loops that get better with every pass.
The full argument: Creative Assembly Loops (Substack).
05Roadmap
For now, I'm focused on Version 1, Creative Assembly Loops: one loop, with a layer added for every plaything it ships, and plenty to do inside it. V2 and V3 are where one loop becomes a network of loops.
| Version | Adds | The question |
|---|---|---|
| V1 · Creative Assembly Loops (now) | Make it | How fast can a tiny human-AI team make playthings to a Denki standard? |
| V2 · + Attract Mode | Validate it | How might we put each plaything in front of a real audience and find out whether it pulls them in? |
| V3 · + Attention Arbitrage | Scout it | Which cultural opportunity, audience, mechanic and distribution surface should we combine in the first place? |
Attention Arbitrage identifies what might deserve to exist.
Creative Assembly Loops makes and iterates it.
Attract Mode discovers whether the audience agrees.
06Learning logs
What each loop taught, as it happened.
- Creative Assembly Loops (Substack: the bet, and how you'll know if it's working)
- The live learning log (a thread on X, updated weekly)
- The making of, part one (Substack, end of November)
- The making of, part two (Substack, mid-December)