Nearly every failed print I have ever had failed in the first 0.2mm. Everything above that is just the first layer's mistakes, repeated upward with confidence. So this is the thing I've spent the most effort getting right, and the thing I most resent having to think about, because when it works you forget it exists and when it doesn't, it ruins an eight-hour print at hour seven.
This is what I actually do now, after years of doing it badly.
the paper feel is a lie we agree to tell
Manual levelling starts with the famous "drag a sheet of paper under the nozzle until you feel slight resistance" routine. It works, sort of, but the amount of resistance you call "slight" is personal, inconsistent, and changes with how much coffee you've had. The paper is roughly 0.1mm thick, so you're not zeroing the nozzle to the bed, you're setting it about a paper's width high and then relying on squish to make up the difference.
The trick that made manual levelling tolerable was to stop chasing perfection at the four corners and instead do two passes. First pass: get all four corners roughly equal with the paper. Second pass: print a single-layer test square that covers most of the bed and adjust the knobs while it prints, watching the lines go down. You can see under-extrusion (gaps between lines, bed showing through) and over-squish (translucent, smeared, ridged) far better than you can feel it.
what good actually looks like
A good first layer line is slightly wider than its commanded width and just touches its neighbours, no gaps, no ridges where it's been mashed flat. Lay a torch across it at a low angle. Gaps mean nozzle too high, raise nothing and lower the bed (or increase squish). A surface like corduroy, with raised ridges between lines, means too low: the plastic has nowhere to go but up and sideways.
I print a 0.2mm single-layer patch as my calibration object. It takes ninety seconds and tells me more than any amount of paper-dragging.
the bed is not flat, and that's the real enemy
Here's the thing nobody warns you about: manual four-point levelling assumes the bed is a flat plane that's merely tilted. Cheap beds aren't. They're warped, usually bowed up or dished in the middle, by more than your layer height. You can level all four corners perfectly and still get a dish of bad adhesion in the centre, because the centre is physically further from the nozzle than the corners.
A sheet of borosilicate glass on top helps enormously, because glass is flat in a way a stamped aluminium plate simply isn't. For a long time that was my whole answer: glass, a wipe of glue stick, and accept the corner levelling. It got me a long way.
the BLTouch, and why I stopped fighting
Eventually I fitted a BLTouch (an inductive-style probe that physically taps the bed at a grid of points) and let the firmware build a mesh of the actual surface. This is the part that ends the war. Instead of pretending the bed is a tilted plane, Marlin measures, say, a 5x5 grid and then adjusts Z in real time across the print to follow the warp. The dished centre stops being a problem because the nozzle dips to follow it.
The setup that mattered:
M851 Z-1.40 ; probe Z-offset, the single most important number
G29 ; run the mesh probe
M500 ; save to EEPROM so it survives a reboot
The Z-offset (M851 Z) is the value that turns a measured mesh into an actual first layer. The probe triggers at some height above the nozzle tip, and that difference is the offset. I set it live with babystepping during a first-layer print: start too high, nudge down 0.025mm at a time until the lines join, then save. One number, set once, and most of my first-layer grief evaporated.
A few things that quietly mattered as much as any of the above:
- A clean bed. Skin oils are the silent killer of adhesion. Isopropyl alcohol on a kitchen towel before any print I care about.
- Probe at temperature. Aluminium grows when hot. Probing cold and printing at 60C bed means probing a different shape than you print on. Heat the bed first, then
G29. - Don't over-tighten the springs. A bed clamped solid can't be adjusted and transmits every vibration; leave them with a little give.
the honest conclusion
The war isn't really winnable, it's manageable. Hardware drifts, belts on the Z axis settle, a probe's trigger height changes a hair when you swap a nozzle. What changed for me was moving from "make the bed perfect" to "measure the bed and compensate". The first is a fool's errand on consumer hardware. The second is a fifteen-minute setup and a number saved to EEPROM, and then I get to forget about it again, which was always the goal.