Not the arc. The things on this page are at DC or at mains frequency, they are inside the cabinet, and they do not announce themselves. The site's own safety guide puts the same ordering in one line: the supply, the primary circuit and the capacitors are far likelier to kill you than the discharge everyone is watching.
The usual explanation for why a streamer hit is survivable is wrong, and it is worth killing before the rest of the page. The story is skin effect, that hundreds of kilohertz travels on the surface and misses the organs. Skin depth goes as one over the square root of conductivity times frequency, and the primary is an impedance carries the anchor for it, 0.103 mm in copper at 400 kHz. Tissue conducts something like a hundred million times worse than copper, and the depth goes as one over the square root of that, so at 400 kHz the skin depth in it is between about 0.8 m and 2.5 m for conductivities of 1 down to 0.1 siemens per metre [derived by scaling the copper figure]. Neither end of that bracket is a shield. The conductivity is an assumption rather than a figure this site holds, and the answer does not turn on it: the current goes through the middle of you at every value in the range.
What makes the difference is excitation, not depth. At mains frequency a current through the body drives nerve and muscle directly. That is the grip nobody can let go of, and it is what puts a heart into fibrillation. As the frequency rises the same current excites less and less, and what is left of it is heat. So a hit at hundreds of kilohertz burns rather than fibrillates, and because nothing is keeping the current on the surface the burn can be deep and can be worse than it first feels. The Tesla Coil Safety document on pupman, contributed by Chip Atkinson, Ed Phillips, Mark S. Rzeszotarski and R.W. Stephens, states the mechanism in one line: "Higher frequencies are less able to cause these involuntary contractions." The same document warns that the heating from a large coil is mostly internal, which is the opposite of what a surface-current story would predict, and a few lines earlier it repeats the skin effect claim anyway, with a figure of about 500 kHz attached and an admission that it has little data between 50 and 250 kHz, which is exactly the band coils run in.
And the arc is a wire back to everything on this page. The pupman document's sharpest warning is not about the streamer's own current at all: if the coil strikes something on the mains side while somebody is in contact with a streamer, the ionised path connects that person to the mains, and it is the mains that kills them. A grounded strike rail lowers the odds of that and is not proof against it. Being in contact with a running coil is not made safe by any of the physics above.
The bus bank
Thousands of microfarads at a few hundred volts. On 230 V mains that is about 325 V behind a plain bridge and about 650 behind a Delon doubler, and the bank on the machine this site works from is 9900 µF; every figure in this section comes from the bus bank, which sizes the part properly. That is a lethal store, and it does not go away when you switch off.
The size of it is not a matter of taste either. Across that 9900 µF, 6 kΩ takes two and a half minutes to fall to 50 V and dissipates 70 W; 2.4 kΩ takes a minute and dissipates 176. Both are serious resistors, the wait is long, and that wait is the reason the machine also needs an indicator.
Indication, and its limits. A bright LED with a dropper resistor tells you the bank is up and bleeds it slowly at the same time. A neon is a useful backup.
Precharge, separately, is about protecting the machine rather than you: a cold bank at switch-on is a short circuit across the mains. A resistor in the AC line before the rectifier, shorted out by a relay once the bank is up. How long that takes is a time constant and not a habit. Sixty ohms into a single 9900 µF bank behind a plain rectifier is 0.59 s, and the same sixty ohms behind a doubler is 2.38 s, because the doubler's capacitors are in series and each side is fed by its own diode for half the cycle, so the two effects multiply. A doubler charges four times as slowly, not twice. Three seconds of relay delay is only 1.26 time constants of the doubler case, which leaves the bank 72 per cent charged when the resistor is shorted out and the rest of the charge arrives with nothing limiting it. Set the delay from the machine in front of you.
The frame under the primary
A metal frame under the primary is a one-turn secondary of that primary.
The cure is to break the loop: nylon bushes, washers and Kapton, so that each upper rail is grounded at one end only, and the strike rail taken to RF ground on its own wire rather than through the frame. The numbers are in grounding a Tesla coil, including how much gap that break actually needs, which is more than the two millimetres it looks like.
The secondary's ground
A counterpoise instead, and a short lead to it, which is its own piece.
The current transformer nobody thinks about
Which of the two burdens you can put an ordinary probe across, and which one needs a differential probe, is on what a DRSSTC current transformer is, and what it senses.
The tank, while the coil is off
The tank capacitor sits in a resonant circuit that has been at kilovolts, and it holds that charge like any other capacitor. Treat it as charged, short it deliberately, and do it with something insulated.
A bleeder does not spoil the tank, so there is no argument for leaving it off. The site's safety guide asks for roughly 1 to 10 MΩ across every high-voltage capacitor, and on this part that costs nothing measurable. At its own working frequency the tank capacitor's reactance is the primary circuit's characteristic impedance, 37.3 Ω on the coil this site works from, from a 16.96 µH primary and a 12.2 nF tank (the primary is an impedance). One megohm in parallel with 37.3 Ω of reactance is an added series resistance of 1.4 mΩ [derived, X²/R], against a tank loop that is measured in tens of milliohms: about 90 mΩ on the coil fitted in SPICE in dead time as a fraction of the period. The same resistors drain the bank fast. That 12.2 nF is 99 parts of 10 nF in nine series by eleven parallel (the MMC is sized by current), and 1 MΩ across each part is a 10 ms time constant on one string and on the whole bank alike [derived, since the resistance and the capacitance divide the same way]. Ten milliseconds is not a reason to skip the shorting stick, because a cracked resistor and a lifted lead look exactly like a working bleeder.
And the thing that catches people who know all this
The bench, not the coil. The measurement setup is where most equipment and a fair number of people get hurt, because a scope's ground is bonded to mains earth and a coil's power section is not. Scoping one without killing anything is a separate page and it is worth reading before the first power-up rather than after.
This page is a floor, not a safety course. The safety guide is the site's own, and none of it substitutes for the habit of assuming every capacitor in the cabinet is charged until you have personally shorted it.