object — one thing, built from parts

world/ is still the master workflow: step 0 (commit to a brief) and step 5 (finish honestly) are the same here. This skill replaces steps 1–4, because an object diverges from a scene in one way that changes everything: the iteration unit is a part, not the world.

A gramophone written as one 800-line main.js cannot converge. Changing the horn means re-running the whole world and squinting at one wide shot. Split it, and each piece gets its own closed loop.

The loop

decompose → arithmetic in params.js → build one part → inspect it alone → next part
                                          ↑______ sheet not clean? fix here ______|
                                                                    ↓ all parts clean
                                                    assemble → check joints → light → finish

Gate: main.js does not import a part until that part's sheet is clean. Downstream never consumes an unreviewed upstream. This is the whole reason the split pays for itself.

Steps

1. Decompose, and write the brief

Name the identity-defining features first — the handful of things that, if wrong, make it not that object. Cap it at five. A gramophone is a brass flare, a wooden box, a spinning record and an arm reaching over it; the moulding profile is not on that list.

Split by how it is made and how it moves, never by what your eye groups together. See parts.md for the contract, assembly.md for what breaks once they are together.

Put the reference in worlds/<name>/refs/ in the same sitting (git-ignored — a reference can never become an asset). Name the identity-defining features off the photograph, not off your idea of the thing: what your memory hands you is a generic member of the category, and the two features that make it that object are exactly the ones it drops. From here on, inspect --ref and capture --ref put it in the same sheet as your frames — see review.md, and docs/principles.md workflow 2 for why this is scar tissue rather than advice.

2. Do the arithmetic before any geometry

Write params.js first, with named values in metres, and check the relationships numerically. On the gramophone the tonearm post sits 0.184 m from the platter centre and the arm is 0.185 m — so at the wrong swing angle the head lands off the back edge, and with the wrong drop angle the needle passes through the record. Both were arithmetic, not aesthetics:

postH − armLen · sin(drop) = record surface height + needle drop

Reach, clearance and contact are algebra. Do them in params.js where they stay checkable, not by nudging numbers until a screenshot looks acceptable.

3. Build one part, review it alone

node harness/inspect.mjs <world> --part horn

Four viewpoints plus the numbers a screenshot cannot tell you. Read the sheet before writing the next part — see review.md. Do not batch five parts and then look.

4. Assemble, and check what only exists once it is assembled

Three classes of defect belong to the assembly and to no part, so nothing upstream has looked at any of them: two parts in the same place, joints that read as glue, and — for anything with an inside — whether the thing can be seen at all. See assembly.md.

This is where the time actually goes on an object with an interior, and it is the step the loop is most tempted to rush, because by then every part sheet is clean. Stand the parts up in their real positions as soon as they exist rather than at the end: these checks cost seconds now and a rebuild later.

node harness/probe.mjs <world> --rows                     # repeated parts inside each other
node harness/probe.mjs <world> --scout x,y,z --dist 0.3   # where can this point be seen from?

5. Light and finish

Metals are black without an environment; runtime/studio.js generates one (no downloaded HDRI — D4 holds). Wood under that environment turns to plastic below roughness ~0.7.

Then finish per world/ step 5, plus the honesty rules in review.md. A sheet that passes is not the same as the brief being delivered — say which one you have.

What this skill is not

Not manufacturing. We build form, not parts you could machine: no STEP, no tolerances, no wall-thickness checks. If the user actually wants to print it, that is a different tool (docs/architecture.md D3).

And when matching a reference photo: we copy the form, never the pixels. Projecting a photograph onto the mesh is the highest-fidelity trick in the image-to-3D world and it is exactly what D4 forbids. Reference images are for comparison only.