Producer handbook — the rig, the chain, the guests, and the gate before air
The show page tells you what the seat is. This tells you how to run it. Read it end to end once; after that it is a reference you come back to under time pressure, which is why every section is short and every rule says why it exists.
The production model. Participants use their own computer and their own phone — we do not ship hardware. That is what makes twelve seats plus twelve standby affordable, and it is why you are a support desk as much as a broadcast operator. Everything a participant needs to do on their own machine lives in the setup sheets, so you send a link rather than retyping it twenty-four times.
Settings that are wrong here fail quietly: the stream goes out, looks acceptable to you, and degrades for everyone watching on a worse connection than yours.
Setting
Value
Why
Canvas = output
1920×1080 both
Base and output resolution identical. Rescaling spends GPU on nothing and softens the picture.
Encoder
Hardware (NVENC/QSV/AMF)
Dedicated silicon, so encoding does not fight the machine for frames. x264 only makes sense on a dedicated streaming box.
Stream rate control
CBR
Platforms need a predictable ingest. Variable bitrate on a live ingest is what produces buffering on the viewer's end.
Record rate control
CQP / CRF 18–20
Constant quality, not constant size. The local file is the edit master; let it spend bits where the picture is complex.
Keyframe interval
2 seconds, never auto
Ingest pipelines split segments on keyframes. Irregular intervals trigger stability warnings and hurt seeking.
Bitrate
6000 kbps @ 1080p60
And the whole stream — video plus audio — stays under 80% of measured upload, not advertised upload.
Codec
H.264 unless told otherwise
AV1 is materially more efficient but only where the destination actually accepts it. Confirm before switching.
Sample rate
48 kHz everywhere
Device, OBS and destination. Mismatched rates are the usual cause of audio drifting out of sync over a two-hour show.
Our box has one encoder, not two. You will read everywhere that you should run the stream on one NVENC chip and a lossless local record on a second. That needs a dual-encoder card. The studio machine is an RTX 4070 — single encoder. Plan for one encode path and take the local record from the same one, or accept the frame cost. Do not design a show around a capability this hardware does not have.
Running OBS as administrator does not give it GPU priority. That claim is everywhere and it is wrong. Admin affects whether OBS can capture elevated windows and how the process is scheduled — nothing about GPU allocation. It is worth doing for capture reliability. It is not a fix for dropped frames, and reaching for it as one wastes the ten minutes you needed to find the real cause.
BThe audio chain
Audio is what makes a stream feel amateur, and it is the thing viewers leave over. Gain-stage before you filter anything. Raw microphone peaks belong at −12 to −6 dB. Filters applied to a signal that is too quiet just amplify the room.
1 · Noise Suppression→2 · Compressor→3 · Limiter
That order is not a preference. OBS runs filters top to bottom. Clean the signal, then shape it, then protect it. Compress before you suppress and you have amplified the noise floor into the same range as the voice, where suppression can no longer separate them.
Filter
Solves
Noise Suppression
Fans, air conditioning, room hum — the constant bed you stop hearing after ten minutes and the audience never does.
Compressor
Inconsistent level. Brings the quiet up and the loud down so nobody rides their volume control.
Limiter
The ceiling. Catches the laugh, the desk bang, the sudden lean into the mic — the spikes that clip and cannot be repaired.
Alerts get Monitor Only (Mute Output). An alert source left on Monitor and Output is broadcast twice — once as the browser source, once through your monitoring path. You will not hear the fault; only the audience will.
Everything above is per source. The bus architecture for twelve to twenty-four microphones open at once is not yet built. Routing, who hears whom, how the public/private split behaves under load, and what the mix does when eight people talk together — none of it is yet measured, and it is the first thing the pilot broadcast exists to settle. Treat a clean chain on one microphone as evidence about one microphone.
CCameras and inputs
Participants bring a computer and a phone. The phone is the second angle, and it is usually the better sensor — a modern rear camera beats almost any webcam.
Rear lens, never the selfie camera. Bigger sensor, better in poor light.
Choose the smooth/low-latency profile over maximum resolution. On a live show, delay costs more than sharpness.
Set 1920×1080 in two places — the source itself and the OBS browser source. Setting one and not the other silently gives you an upscaled 720p.
Lock the camera down: manual focus (auto-focus hunts every time hands move), shutter at double the frame rate, lowest usable ISO, and a fixed white balance matched to the room. Auto-anything drifts on camera.
Sync offset. USB and browser video arrive late; microphone audio does not. Expect to delay audio by 100–300 ms (200 is a good first try) and confirm with a clap.
Phones get hot and phones are needed. A handset running a camera feed for two hours is hot enough to matter, and it is also the device that person takes calls on. Treat it as a second angle, not the primary.
Do not treat passcode masking as privacy. Phone operating systems often blank secure fields when mirroring — not always, not on every OS version, and not for every app. If a participant is sharing a phone screen, they log in before they go on, or you cut away while they do. Never let "it usually hides it" be the control.
DGuests
Every participant is a remote guest, so this is the section you will use most. Guests join a room in the browser; you take each one into OBS as a source.
Individual links per guest. Never one group link. A group capture arrives as a single audio fader. When one of eight people is twice as loud as the rest — and someone always is — you cannot fix them without quieting everyone. Individual sources give you a fader and a mute per person, and separate tracks in the recording for the edit. This does not change how guests talk to each other; they are all still in one room.
Create the room as director. Keep the invite link and the per-guest view links separate — you send the first, you paste the second.
Add each guest to OBS as a browser source at 1920×1080.
Tick "control audio via OBS" on every one. Without it the audio bypasses your mixer entirely and you have no fader on a live person.
Send your programme output back into the room as a virtual camera. Guests then see what is actually on air — a real tally light, and they stop asking whether they are live.
Echo — the one you will fight every week
Echo is caused by speakers, essentially always. A guest on speakers has your voice coming out beside their microphone, which sends it back to you. There is no software fix that beats the physical one.
Headphones are required, not encouraged. Any wired pair works; the cheapest wired earbuds solve this completely. See the setup sheets for what we tell them to buy.
Echo cancellation on only for someone who genuinely cannot wear headphones — it degrades their voice, so it is the fallback, not the default.
Wired network over wireless wherever it is possible at all.
Mute anyone not speaking. Desk knocks and chairs are picked up long before anyone notices they are transmitting.
EScenes and graphics
Sixty-five episodes means the layout gets built once and reused. Build it so a change happens in one place.
This section is how you operate the scenes. It is not why they are shaped the way they are. The layouts are computed rather than drawn, against three ceilings you cannot argue with — nameplate legibility, bitrate divided across every moving region, and how many feeds one box can decode while it is also encoding. That reasoning, the sixteen-face cap, the virtual set and the colourway live on the broadcast architecture page. Read it once before you rebuild a layout, because a layout that crosses a ceiling degrades quietly rather than failing.
One source scene per feed, nested into every layout. A colour correction or a resolution fix then applies everywhere at once. Duplicating a camera into six scenes guarantees six versions that slowly stop matching.
When one feed genuinely needs different treatment in different layouts, clone the source rather than rebuilding it.
Animated assets are media sources with loop enabled; static ones are images. Picking the wrong type is why an overlay "will not import".
Data-driven graphics — scores, names, tickers, gate verdicts — are HTML browser sources that read the system's own state and re-render in place. They are version-controlled and styled from brand tokens, and they need no plugin and no second machine. You do not type into them.
NDI is local-network only and is relevant only if a real-time 3D engine is adopted, which is still an open decision — see the architecture page. If it is adopted, the plugin has to be installed and OBS restarted before the source type appears. Until then there is nothing to install.
Lock every layer before you go live. The most common self-inflicted on-air fault is dragging a source while reaching for something else.
Every graphic is a cost. Browser sources and filters render on the same GPU that is encoding the show. Heavy animated overlays can take a serious bite out of headroom. Watch the stats panel, not your impression of how it feels.
FRecording while streaming
The live stream is not the deliverable. The recording is, because it becomes the episode, the clips and the archive.
Format
Behaviour on a crash
Use when
MKV
Playable right up to the moment of failure
The safe default. Remux to MP4 afterwards — it is near-instant and lossless, and can be automatic.
Fragmented MP4
Also survives — written in chunks as it goes
You want a directly usable MP4 with no remux step. A real option, not a compromise.
Plain MP4 / MOV
Total loss. Not finalised until you press stop
Never, for a two-hour live show.
Record each source separately as well as the programme mix. A clean camera with no overlays burned in is what makes a good clip later; a composited recording is stuck the way it aired.
Separate audio tracks per person. One person coughing over the answer is fixable on separate tracks and permanent on a mixed one.
If vertical cuts are wanted, produce them in the same session rather than re-cropping later.
GSimulcast — the destinations
Every episode goes to all of them, horizontal and vertical at once. Each destination is there for a different reason, and the reason drives what you do with the feed.
Destination
Role
Audience
YouTube Live
Anchor VODs · long-form archive
Subscribers · VOD search
Twitch
Interactive build-along · trial screens
Dev audience · clippers
Kick
Raw, unfiltered late builds
Culture audience
X
Rapid takes · live debugging
Authority · founders
LinkedIn Live
Enterprise strategy · case studies
High-ticket enquiries
TikTok / IG (vertical)
Vertical canvas simulcast
Reach · younger audience
HPre-flight — the stream does not start until this passes
This is a gate, not a checklist. If an item fails, the show starts late. Starting on time with a known fault is the more expensive choice every single time, because it becomes sixty-five episodes of the same fault.
Test broadcast, 60–120 seconds, to the real destination. Confirms ingest and that your bitrate survives your actual upload right now.
Test recording, 20–30 seconds — and play it back. Recording it is not the test. Watching it is.
Clap on camera and check the sound lands on the frame where hands meet. If not, set the sync offset now.
Stats panel: no encoding overload, no missed frames, no dropped frames. All three are zero, not "low".
Every participant confirmed — camera up, level in range, headphones on, name correct on screen.
Release status verified for everyone on camera today. Unsigned means not on the stream. This is the one item with no discretion attached to it.
IWhen it breaks live
It will. What is being judged is not whether something broke, it is how long the audience watched it break.
Failure
Move
One guest drops
Cut to a layout without them and keep going. Do not hold the show while somebody reconnects; bring them back when they return.
Their audio dies, video fine
Cut away immediately. A silent talking head reads as your fault, not theirs.
Encoder overload mid-show
Drop the heaviest browser sources first — they are usually the cost. Resolution comes down before the stream comes down.
A destination stops accepting
Let it go and keep the others up. Note the time; it matters for the report.
Your machine dies entirely
The recording is the fallback. This is why the format choice above is not academic.
Log every incident with a timestamp, during the show. Not afterwards from memory. Whether a participant was unable to work because of their own setup or ours is a pay question, and it gets settled from your log. Reconstructed timings are worth very little in that conversation.
JElimination night
Every fourth Saturday, and it is a different programme from a weekday episode. Somebody is losing a job on camera.
Nothing about the outcome is improvised. The decision is read from the system — work executed, gates passed, work published. Your job is to make sure it is legible on screen.
Each participant's counterpart speaks about them on air. Build a layout that holds both faces; this is the moment the audience is actually watching.
Record everything separately. This is the most-clipped night of the cycle and clips need clean sources.
Have the exit handled before you go live — access, the closing shot, and what airs after. Discovering it live is a bad look at somebody else's expense.
The run of show, the rounds and the dates are on the programme page. They are not repeated here on purpose — two copies of a schedule become two different schedules.
KAfter the stream
Remux if needed, and confirm the file plays before you close anything.
Check the per-source recordings exist and are not zero-length. This is the failure people find a week later.
Hand off to the Technical Operator in the agreed format and location.
File the incident log, including a clean night with nothing on it — "nothing went wrong" is data.
Open item. The handoff format, naming and destination are not yet agreed. Settle it with your Technical Operator in week one and write it down, because "I sent it over" is not a location.
LEscalation — who decides what
You produce. You do not approve. The Executive Manager holds the brand sign-off; the Executive Chairman holds the show sign-off. Neither of those is yours, and neither should be discovered mid-season.
Decision
Whose
Anything about the broadcast chain — rig, layout, cut, when to cut away
Yours, live, without asking.
Whether an episode is right for the brand
Executive Manager
Whether it is a good episode of the show
Executive Chairman
Whether someone gets paid for a disrupted session
Not yours — but it is decided from your log
The five-minute rule. Under time pressure before air, anything that is purely broadcast is your call and you make it. Anything touching brand or on-air consequence goes up. If you genuinely cannot tell which one it is, cut away first and escalate after — cutting away is reversible and broadcasting the wrong thing is not.