Three-Body Spirograph
April 2026A scientific-instrument redesign of the three-body problem — monochrome graphite UI, cold/warm mode accents, orbit trails carrying all the color. Canvas 2D, velocity-Verlet physics, seven presets from Star & Planet to Moore's figure-eight to the full eight-planet Solar System. More than a simulation: a proof-of-concept for chaining specialized AI tools — a design model produces a build-ready handoff, a planner/reviewer holds the spec, a coding model writes the diff.
- 01Claude Design — Generated a pixel-perfect HTML prototype, TypeScript reference scaffolds (Stage, Slider, Toggle, DivergencePlot), design tokens, and a 16-point acceptance rubric — delivered as one build-ready handoff package.
- 02Claude Code — Director — Received the handoff, planned a 5-chunk breakdown (tokens → physics → rails → stage → intro), wrote tight specs for each chunk with file paths, hard constraints, and per-chunk acceptance criteria.
- 03Codex CLI — Coder — Landed chunks 1 and 2 via the codex-rescue subagent, returning a diff per run. A spec deviation on preset initial conditions (Keplerian-circular instead of README-verbatim) was caught by QA and corrected in a single loop.
- 04Claude Code — QA — Ran a 27-point rubric after every chunk: build + lint, runtime probes, acceptance checks against the handoff, and regression tests to confirm V2 and V4 still rendered unchanged.
- 05Claude Code — Fallback Coder — When Codex's sandbox locked read-only on chunks 3–5, the director took the keyboard directly under the same spec/QA discipline rather than blocking the build — orchestration as a pattern, not a dependency.
