If someone is interested... here you go.
The basic necessity - you must exactly know what you want to design, otherwise it usually doesn't end too well. The best is to have three(or more)view orthographic drawings, well-detailed and in exact scale. Or when the subject is so small that you can measure it with a caliper.
There are also less-ideal situations when we have no more than a few pictures, photos, screenshots or comic panels and a slight idea about the overall size (which is quite frequent with sci-fi subjects). My favourite solution is to take a pencil and draft ortho views on a paper, with no ruler, compass or scale. Paper accepts anything, doesn't need exact dimensions or coordinates and one has better idea about the proportions than on a screen.
The drawings are somehow imported into the computer and next phase begins:
For simple things I use just a cross-section as a reference or draw unfolded parts directly, but complex shapes, intersections etc. are out of my skills without 3D solid modeling. I work in an old student version of AutoCAD 2002 on even older computer (166 MHz processor, about 90 MB of RAM, several GB disk and Windows 98), which runs it quite smoothly.
I usually start with pinning the orthos in the modelling space, align positions and scales and then use them as a direct reference. The particular process varies. I usually create solid by extrusion or rotation, then slice them down to the final shape. Intersection shapes are determined by boolean operations. Of course, besides solids, you also need lines, curves and circles. Snap-to-something is a very useful utility too, there are many options: snap to endpoint, midpoint, centrepoint of circle, perpendicular to something, parallel to something and so on.
Disadvantage is that the basic AutoCAD is not parametric, so if you want for example move a hole, you must manually fill it and then drill at the new location. In a parametric editor, you just rewrite a coordinate.
This step occurs everytime. Unfolded parts (in vector environment) have to be drawn, placed on pages, given tabs and so on.
AutoCAD has no built-in function for this (or at least I don't know about it), so I do it manually: I copy edges of a solid and then unfold it face by face using "3D rotate" and "align". Quite tedious, but with total control over everything, possibility to add or delete something here and there (e.g. to compensate paper thickness) and choose what faces to add or skip.
If you are looking for a program for both modelling and unfolding, I'd recommend Rhino - it is said to be in all ways better than AutoCAD and much less expensive. Blender and Sketchup are for free and can unfold via plugins and scripts. Another interesting combination is Metasequoia + Pepakura designer. But I have very little experience with these programs.
The parts are printed and test-built. I usually discover lots of bugs during that: missing, redundant or overlapping tabs, missing parts, inconvenient cut/fold lines, inaccuracies caused by paper thickness, errors in markup etc. Then I return to the virtual drawing board and fix everything. Most things can be solved on the unfolded parts level, sometimes I have to get back to 3D and remodel something.
Export from AutoCAD works as a "print" on a virtual plotter PublishToWebPNG.pc3, resulting in PNG files (which is the only lossless and decently editable format from the available set). I usually use bitmap of 1500×2100 pixels at 200 DPI (i.e. 1 pixel = 0.127 mm), which fits nicely on an A4 paper with safe margins.
AutoCAD's virtual plotter lacks editable settings. It badly needs to regenerate arcs more smoothly, small circles are very rough and imprecise. Also, rounding errors can build up and some intersection points run away by a pixel - that must be corrected in some bitmap editor afterwards.
Theoretically, one could use Acrobat of PDF Creator and print to a PDF, but there's no point in using such a hard-to-edit format.
First step: number the parts. That's a few minutes' work in CAD.
Second step: create drawings. This is a bit more work. First I tried to use photos of the alpha prototype with arrows and numbers added. Of course, a nice and clean model in final configuration is needed for that, which is not the most usual case. Sometimes a screenshot of virtual 3D model or some drawing or cross-section directly from CAD can be used, but again, you need the model to be neat and complete. For more complex situations, manual drawing works best for me. I can draw exactly what I want, in any position or state of folding. Earlier I would scan it, decrease colour depth to two colours and clean it up manually. But it's a lot of work and needs a precise hand drawing. Now I just sketch it up somehow, photograph with a camera, rescale, redraw final line art and texts in MSPaint and then filter out the original sketches by a special program.
Third step: write text part of the instructions. There's no point in enumerating "glue part 2 to part 1" because everyone can figure that out from the drawings. There are more important things to write, like "seam of part 5 goes down, part 8 must align with part 3, and don't close part 10 until you finish assembly 11+12", which would otherwise have to be figured out by trial and error.
By now, the kit should be buildable by normal people too, not just me, and sometimes it's good to have several volunteers try it. Then I tweak the kit accordingly to their remarks.
This step gets omitted quite often - let everyone apply their favourite colours themselves.
First I draw outlines of surface details onto the unfolded parts in CAD. If possible, then directly. If not, I draw them by hand onto some of the built prototypes (shapes, proportions and distances are easier to eyeball on the real thing than behind a screen) and then measure them with a caliper and transfer back into the computer.
Then final export to bitmap and the colouring begins. I use a combination of Windows' Paint, shareware Paintshop Pro 3.12 and freeware Gimp. MSPaint is best for quick copy/paste and moving, Paintshop works well for most painting operations and effects. Gimp sucks at fine line and pixel art, but is the only one which can do semitransparent layers, resizing and rotating by any angle, curved texts and other details.
Or escape with a bloody trail left behind. Depends on the situation.