Directed

Every element of a chart, made twice or three times over: once by Claude Code working from the dataset alone, then again with a named skill or a real tool loop run on top. Same model, same data, same module contract. The method is the only thing that changes.

14 of 48 rows built · 34 of 107 cells · every cell generated by actually running its method, hashed at generation, and re-rendered on every build · 6 published as null results.

Method: what contaminates this comparison, and every cell that shipped without moving its lever →

Claude default
The dataset and the module contract, nothing else. Three runs; the median by SVG element count ships.
With a skill
A named skill invoked explicitly, one run, on the default arm's own output.
With a tool
A real loop — screenshot compare-and-fix, an image model, a charting library. One run.
refine · from-scratch
Each row declares one. refine runs the directed arms on the default arm's output, so the lever is the only difference; from-scratch generates each arm independently, because there the lever is the generation decision.