Chat moderation is generally treated as something a friend does while watching. Having done a few hundred hours of it, and having watched several streams go badly wrong because nobody had thought about it, I would treat it as a role with a brief.
What the job actually involves
More than removing abuse, which is the part everyone thinks of.
Surfacing questions the broadcaster should answer, which requires judgement about what is worth interrupting for and what is not.
Answering things that do not need the broadcaster — where the link is, when it started, whether it will be available afterwards. On a busy stream this is most of the volume.
Watching the room's temperature and telling the broadcaster when something is going on that they cannot see, because they are talking and cannot read chat.
And handling the technical failures, which is the thing nobody plans for. When the stream drops, the moderator is the only communication channel, and if they do not know what is happening they cannot say anything useful.
The rules need to exist before you need them
The failure mode I have seen most often is a moderator making a judgement call under pressure with no guidance, being second-guessed afterwards, and either overreacting or freezing for the rest of the stream.
What prevents this is a short written statement of what is not acceptable and what happens when it occurs. Three or four lines is enough.
It needs to specify the escalation. Delete the message, or time the person out, or ban. And whether a ban is permanent or whether it is reviewed.
Publishing it to the audience helps considerably, because it converts an apparently arbitrary decision into the application of a stated rule.
Automated filtering does the volume
Every platform provides some automated moderation and the default settings are usually too permissive for a stream of any size.
Worth configuring: a blocked term list, rate limiting so no individual can flood, a delay before new accounts can post, and link restrictions.
The link restriction in particular. Almost all scam and spam activity in live chat involves links, and restricting posting to established accounts eliminates most of it without affecting genuine participants.
Automated filtering handles the obvious cases and frees the human to do the judgement work, which is the only part a human is needed for.
Slow mode is underused
A setting that limits how often each person can post. It sounds restrictive and it changes the character of a busy chat entirely.
At high volume, chat becomes unreadable — messages scroll past faster than anyone can follow, and it stops being a conversation and becomes noise.
Slow mode reduces the rate to something readable, which means people actually read each other, which means the conversation improves.
Regulars complain when it is introduced and generally stop within about ten minutes.
The broadcaster cannot moderate
Something that needs saying because a lot of people try.
Reading chat while talking is genuinely difficult and does both things badly. The broadcaster's attention drops, the delivery becomes hesitant, and moderation decisions get made badly because they are being made by someone who is mid-sentence.
It also means the broadcaster reads the abuse, which is the thing a moderator exists to prevent.
The workable arrangement is that the moderator filters everything and passes on a small number of things, ideally through a separate channel the audience cannot see.
Looking after the moderator
The part that gets no attention at all.
The moderator reads everything, including the material the audience never sees because it was removed within seconds. On a stream that attracts hostility, that is a genuinely unpleasant few hours.
Practical things that help. More than one moderator so nobody does a long stream alone. Explicit permission to remove things without justifying them. Someone to talk to afterwards if the stream was bad.
And not treating it as a favour that is owed indefinitely. People burn out of this quickly and the ones who last are the ones whose role was defined and appreciated.
What I do for anything with an audience
Two moderators minimum, briefed in advance, with the rules written down.
A private channel between the broadcaster and moderators for anything technical or urgent.
Automated filtering configured properly, not left at defaults.
Slow mode on from the start rather than introduced when it gets bad, because introducing it mid-stream reads as a reaction and introducing it at the start reads as normal.
And a short line at the beginning about what the chat is for. Setting the expectation early does more than any amount of enforcement afterwards.
Chat replay and what it leaves behind
Something people forget. On most platforms the chat is retained alongside the recording and plays back with it.
Which means anything not removed during the stream stays visible to everyone who watches later, indefinitely.
Moderation during a live stream is done under time pressure and things get missed. Reviewing the chat afterwards and removing what was missed takes a few minutes and nobody does it.
It is also worth deciding in advance whether to retain the chat at all. For some kinds of stream it adds value to the recording. For others it is a permanent record of the worst thing anybody typed, attached to your video.