Mesh like a pro.
One asset, five modules, real reps.

A mesh is three things and nothing more: vertices (points in space), edges (a line between two verts), and faces (a filled surface bounded by edges). A face with four verts is a quad, three is a tri, five-plus is an n-gon. Quads are the currency of modeling — they subdivide predictably, bevel cleanly, and shade without surprises. Tris are fine on flat, final geometry; n-gons on a curved or soon-to-be-subdivided surface are a bug waiting to happen.

Before shape, though: scale. Treat one world unit as 1 meter and never drift from it. A colonist is ~1.8 m. A doorway is ~2.1 × 0.9 m. A work table sits at 0.75 m. Model the Pod against those and it drops into the game correctly on the first try. Model it "some size that looks fine on screen" and Module 05 becomes a rescue operation.

Then block out: rough the entire form in gray proxy primitives, judging only proportion and silhouette — the shape you'd recognize as a black cutout at thumbnail size. Detail is seductive and cheap to add later; a wrong silhouette is expensive to discover after you've beveled it.

Extrude pulls a face (or edge) outward, growing new connected geometry from it — the workhorse for building form. Inset shrinks a face inward, leaving a border ring; it's how you carve panels, frames, and recesses. Loop cuts insert a ring of edges around the mesh to divide a surface exactly where you need control. And bevel replaces a sharp edge with one or more small faces — a chamfer.

Here's the part people miss. Turn on smooth shading and the renderer interpolates surface normals across faces to fake a curved surface. On an unprepared box, that makes crisp corners go soft and pillowy. The fix is not to give up and leave it faceted — it's support loops (holding edges): a pair of edge loops placed close to a corner. They pin the shading tight to that corner. The distance from the corner sets the look: loops hugging the edge give a tight, milled highlight; loops set back give a soft, cast-metal roll. Same geometry, different soul — controlled entirely by where those two loops sit.

So the rule of this module: you never add a hard edge without immediately deciding how it shades. Bevel, then support. Extrude a panel, then hold its edges. This is what separates a model that survives a close-up from one that falls apart under light.

A boolean (CSG) combines two meshes: subtract to punch a hole or port, union to fuse volumes, intersect to keep only the overlap. It's the fastest route to mechanical form — a cylinder subtracted from a face is a bolt hole in one operation.

But look at what it leaves at the intersection: n-gons, stray edges, sometimes coincident faces. Smooth-shade that and you get black pinching around the cut; try to subdivide it and it explodes. So the discipline is non-negotiable and immediate — after every boolean you resolve the aftermath: add support loops around the new cut so it holds under shading, put a small bevel on the cut edge so it catches light, delete any interior or duplicated faces, and confirm the result is still watertight (no non-manifold edges). Cut, then clean. Never "later."

Two force-multipliers make this fast. Panel lines — shallow inset-and-beveled grooves — imply that a smooth surface is really separate riveted plates, turning one boring face into believable hull. And kitbashing: model a small library of reusable greebles (a hex bolt, a louvered vent, a port) once, then instance them across the model. Instances are cheap and edit-once. The trick to a convincing sci-fi surface isn't uniform detail — it's density contrast: pack greebles onto a hero area and leave calm negative space beside it, so the busy parts read as purposeful machinery, not noise.

Lathe / revolve: draw a 2D profile — a half cross-section — and spin it around an axis. Out comes a perfectly radial solid: canisters, tanks, nozzles, valve caps, turbine housings. The profile is the design; a clean, well-proportioned profile with its bevels already in gives a clean revolve.

Sweep: take a profile and run it along a path (a spline). This builds anything with a consistent cross-section that follows a route — rails, bumpers, trim mouldings, cable runs. Tube is the special case where the cross-section is a circle: hoses, conduit, wiring — the connective tissue between mechanical parts.

The one judgement call throughout is resolution versus budget. More radial segments = rounder = heavier. Match it to how big the part reads on screen: a thin hose is convincing at 6–8 sides; a big tank rim near the camera might want 24–32. And because generated geometry won't automatically share your hard-surface quad flow, you always finish by converting to mesh and cleaning the join where curved meets built — support loops, merged verts, no pinch.

First, the audit — one deliberate sweep of the whole asset for the errors that accumulated while you were busy making it look good: leftover n-gons (booleans and lathes love to leave them), doubled or loose verts, non-manifold edges, interior faces hidden inside the mesh, and inconsistent normals. Use selection-by-trait and statistics rather than eyeballing — the machine finds these faster than you.

Now a distinction that trips everyone up: face orientation versus smoothing. Face orientation (facing normals) is which way a face points — outward or inward. A face pointing the wrong way renders invisible or inside-out in engine; you fix it with recalculate outside. Smoothing / hard-edge data is different: it's the information telling the renderer which edges are sharp and which belong to a continuous smooth surface, so it knows where to split the shading.

And here's the trap beginners fall into. Smoothing marks are related to the support loops from Module 02 but not the same — they're two solutions to the same problem (controlling how an edge catches light), with different costs. A support loop is permanent geometry: it exports like everything else and it costs polygons against your budget — in exchange it gives a real, dimensional chamfer highlight that holds up under a close camera. A hard-edge / smoothing mark is normal metadata: it's nearly free on the poly count, but it needs a normal map (or a UV seam) to back it, or the edge smears and pinches. So the axis is not editor-versus-exported — both export — it's geometry versus normal-data, and you pick per edge: real bevel where the camera gets close, cheap hard-edge where it doesn't. Get the smoothing wrong and the model looks perfect in Mod Kit and faceted or smeared in-game.

Finally, budget. Poly count is a resource and each asset type has a ceiling — a worldmap building targets around ~80k faces; a prop like the Pod lives far lower, a few thousand up to maybe ~10k. Where you're over, you reduce with intent: dissolve redundant support loops on calm, far-from-camera areas, delete faces that are never seen, and trim segment counts on background curves — protecting the silhouette and every hard edge. Blind auto-decimate is a last resort on hard-surface; it shreds exactly the crisp edges you worked for.