Working well

Best practices

None of this is difficult. Most of it is the small number of things that, done wrong, waste an afternoon — and done right, are invisible.

Prepare the base mesh

Computational tools amplify whatever they are given. A pattern applied to a bad mesh is a bad pattern, only more so, and no amount of parameter tuning fixes it.

Apply your transforms: Ctrl+A → All Transforms. Non-uniform object scale is the single most common cause of a result that is subtly wrong: components come out stretched, thickness is inconsistent, and nothing in the panel explains why.
Quads, evenly sized: Most tools here are happiest on an even quad mesh. Triangles double your copy count; ngons get triangulated anyway; wildly varying face sizes give wildly varying results.
Recalculate normals: Inconsistent normals produce components pointing inward on some faces. Free, instant, and worth doing on anything imported.
No poles: A UV sphere crowds every pattern into a knot at each pole. Use the library Blob or Dome, or a Conway result, which have no poles by construction.
Density before pattern: Set the face count you want first. Changing it afterwards changes cell size, component count and field scale all at once, which means retuning everything.

Component discipline

Model on the XY plane, growing up in +Z. That is the unit cell convention. Rhea normalises whatever you give it, but a component modelled this way behaves predictably.
Count the vertices. Every one is multiplied by the face count. A four-hundred-vertex component you thought was cheap is four million vertices over ten thousand faces.
Boundary vertices on the cell edge. Seams cannot weld if the vertices do not actually meet. If merge seams appears to do nothing, this is usually why.
Test on a grid first. Ten faces, flat. If the component behaves there, curvature is the only remaining variable.

Working with fields

Check the active object. Fields write to whatever is active. Generating one with the component selected is the mistake everybody makes at least once.
Turn on the heatmap overlay. Tune the field visually rather than by inference. It is roughly ten times faster.
Leave normalise on. Most consumers assume a 0–1 range, and raw curvature spans whatever it happens to span.
Blur two or three passes. Computed fields are noisy at the vertex level. A little blur keeps the structure and loses the speckle.
Cavity while iterating, AO when finished. They read similarly and one is instant.

While you work

Start coarse: Develop at a low face count. Everything is faster, the pattern scale relative to the surface is roughly preserved, and what you tune transfers when you raise the density.
Change one thing: These modules interact. Changing three parameters and getting a worse result tells you nothing about which one did it.
Try five seeds: Before concluding an effect does not work, change the seed a few times. It costs nothing and a surprising fraction of bad results are just unlucky arrangements.
Duplicate before applying: Shift+D carries the whole live stack. Apply the copy, keep the original. One second, and it has saved a lot of afternoons.
Read the result line: Vertices, faces, ngons and time, after every run. It tells you whether you are about to have a performance problem before you have one.

Save versions, not just files

Computational design is exploratory: you will produce things you like and then change them past recognition. Save incrementally — Ctrl+Alt+S — rather than trusting undo to reach back far enough. It will not.

Getting output you can use

Turn on strict topology for anything going into a game engine, a subdivision workflow or a slicer. The result line confirms zero ngons, so it is a fact rather than a hope.
Leave merge seams on. Unwelded seams look fine in the viewport and fall apart at every downstream step — shading, boolean, 3D print, export.
Use angle-based shading. A freshly built mesh has no sharp-edge data, so without it every hard corner renders rounded and every result looks like the same soft blob.
Check scale before exporting. Particularly for fabrication, where the export scale multiplier decides whether you cut something the right size or a thousand times too small.

Design advice, not software advice

Vary something. A perfectly uniform pattern reads as manufactured, which is fine if you meant it. Nature varies; a field driving thickness or size is what makes a result look grown.
Vary it with signal, not noise. Randomness makes things irregular; a field makes them responsive. Struts that thicken where the form bends look structural. Struts that thicken at random look broken.
Two scales beat one. A coarse pattern with a fine one over it is how almost every natural surface is built.
Let the form show through. If the pattern is so strong that the underlying shape is unreadable, you have made a texture rather than a design.
Stop earlier than you want to. The most common failure in generative work is one effect too many.

Things that will not work

Tessellating with an unapplied non-uniform scale. Fix the scale.
Running Conway on an open or non-manifold mesh. It is not that Rhea refuses out of caution; the operators are undefined there.
Expecting a hundred Voronoi cells on a two-hundred-face mesh to have clean boundaries. The partition cannot be finer than the faces.
Stacking two heavy cellular effects. The second sees a mesh that is mostly holes.
Raising iterations to fix a solver that is oscillating. Lower the relaxation instead.
Using noise where you wanted randomness. Perlin-family noise is smooth and clusters at mid-range; it will give you a suspiciously even result.