Most of the geometry I work from was drawn by somebody else, and almost everything I draw is cut by somebody else. Between those two ends a file gets exported, emailed, imported and re-exported three or four times, and every one of those steps is allowed to quietly throw something away. A part that is correct in my CAD and wrong on the router is rarely a modelling failure. It is a handover failure, and the cheapest kind of mistake to prevent.
Three kinds of geometry, and only one of them is a dimension
Before arguing about file extensions, be clear about what is inside them.
A solid knows its own inside. It has volume, so you can ask it for a wall thickness, cut a section through it and trust that a hole is a cylinder with an axis. This is what a fabricator actually needs.
A surface set is a skin: the right shape, no declared inside. It behaves like a solid right up to the moment a gap between two patches stops it stitching and the CAM refuses to offset it. Surfaces are normal in signage, where letter faces and returns start life as curves, but a surface model is only finished once it is watertight.
A mesh is an approximation made of flat triangles. However dense it is, a curve is a chain of facets and a hole is a polygon, not a circle, so there is no centre for a drill to find and no radius to measure. A mesh is a picture with depth: it answers "what does it look like" and never "how big is it, exactly".
Most bad handovers come down to somebody sending one of these when the receiver needed another.
What arrives from the architect, and what each thing is honestly for
On a fit-out or a shopfront the geometry lands in one of four forms, each with a job it does well and a job it cannot do.
- PDF. Layout intent, sequence, finishes, how the thing is meant to read. Its dimensions are worth what the person who typed them knew, and scaling off it with a measuring tool is guessing with extra steps.
- DWG. Still the currency of construction in both Ireland and Poland. It is a drawing though, not a product: layers, blocks and external references, drawn to whatever discipline the office keeps. If the references did not travel with it, half the plan is missing and nothing warns you.
- A building model, exported neutrally. Coordination data: grids, levels, storeys, walls as objects carrying properties. Good for knowing where a soffit sits relative to a column, and not fabrication geometry. The tolerance behind it is a building's, not a bracket's. Expect an older schema version than the current one, because the older ones are what every package reads reliably.
- A presentation model from a visualiser. A mesh, usually, sometimes beautiful. A mood board with coordinates.
None of these is a survey. Geometry from a designer tells me what is intended, geometry from a tape measure tells me what is there, and I only cut metal against the second kind.
STEP for 3D, and what it leaves behind
When another party's CAD has to open my part, I send STEP: the neutral solid format everyone can read, carrying exact faces, real edges, true radii, assembly structure and part names. There is an older generation and a newer one; both carry the shape properly, and the newer one can also carry dimensions and tolerances attached to the faces they belong to rather than as loose annotation. For signage and fit-out either is fine, and if the shop asks for a specific one, give them that and stop thinking about it.
What does not travel is the part you spent the time on: no history, no sketches, no parameters, no constraints, no assembly mates. A STEP file is deliberately dumb geometry. That is right for a handover and wrong as a master. Once a model has been round-tripped through STEP it is a frozen shape, and editing that instead of the original throws the parametric work away for good. Send STEP, keep the native file.
Two more traps. Threads usually arrive as plain cylinders, so a threaded hole has to be called out on a drawing and not only modelled. And face colour transfers inconsistently, which matters when a panel's visible face must be unambiguous.
DXF for the cutter, and why the old version is the safe one
For anything cut flat, DXF is still the format, and the part that surprises people is that the oldest widely supported version is usually the most reliable one to send. That old version has no spline entity at all, so on export every spline and ellipse is written out as a polyline: a chain of short straight segments.
That sounds like a downgrade and it is precisely the point. The curve gets flattened inside my CAD, where I set the tolerance and can look at the result, rather than inside an unknown importer on a machine older than the job. Send a newer DXF full of live splines to a controller that dislikes them and the first place you learn about it is the cut part. Ask the shop what their software wants, and when there is no answer, default to the oldest sensible version.
The check worth running on every cutting file before it leaves:
- every profile closed, with no gap that only shows up at high zoom;
- no duplicate geometry stacked on itself, or the machine cuts the same line twice;
- text converted to outlines, always: live text is a font dependency on somebody else's computer;
- cutting layers carrying geometry only, with dimensions, hatches and title blocks removed;
- an outer contour that really is outer, and cut-outs that really are interior;
- material and thickness in the file name, not only in the covering email.
The boring failures: units, origin, orientation
These cause more re-cuts than curve fidelity ever will.
Units. Millimetres, inches and metres look identical on screen. Mesh formats are the worst offender, because the common ones store no unit at all: the coordinates are bare numbers and the receiving software decides what they mean. A metric-to-imperial error is large enough to be obvious. A factor of ten is the dangerous one, because the part still looks plausible.
Origin. Put it where a human would have put it: a corner of the sheet for a cut file, a datum face for a machined part, the agreed setting-out point for anything site-related. An origin a thousand metres away, inherited from a site coordinate system, makes viewers behave strangely.
Orientation. A flat part belongs in the horizontal plane, good face up. Handed parts are the classic loss: a mirrored bracket costs the same to fabricate as the correct one and nobody notices until assembly, so if a part is handed, say so on the drawing in words.
And model at full size, always. Scaling is something a drawing does, not a model.
What to send to whom, and what to check before it goes
| Receiving it | Send | Why |
|---|---|---|
| Metal shop, welder | STEP of the assembly plus a dimensioned PDF | the solid gives shape, the drawing says what is critical |
| Laser, sheet metal | Flat-pattern DXF per part plus a PDF | the bend allowance is resolved before it leaves my desk |
| CNC router, plotter, signage shop | DXF, old version, outlines only, full size | their software is older than anyone admits |
| Architect, main contractor | PDF plus DWG | they coordinate a drawing, they do not make a part |
| Visualiser, renderer | Mesh, or STEP if preferred | nothing here will be machined |
| Prototype or 3D print | STEP where accepted, otherwise a mesh with the unit stated in writing | mesh formats will not state it for you |
Three habits carry most of the benefit. Open every export in something other than the software you drew it in. Name files so the material, thickness, revision and handedness read without opening anything. And when you are unsure what the receiving end can read, ask before you export rather than after they cut.
None of it is glamorous and all of it is cheaper than a return visit to site. If the handover is the part you would rather hand off, that is what production support covers: the pack leaves in the format the shop asked for, checked before it is sent.
