The reply section under your own post is where a post keeps going or stops.
X's published ranker weighs a predicted reply at ten times a predicted like,
and adds a boost on top when the two people follow each other (run
python3 ../x-plan/weights.py for today's values). A thread with the author
in it keeps getting served. A post the author abandoned does not.
But the value is not equal across replies, so this skill sorts before it writes.
Input
The user pastes the replies, ideally with handles. Screenshots are fine. Do not scrape the thread with a browser tool.
Triage first
Sort every reply into one bucket and say the counts out loud:
| bucket | what it is | what it gets |
|---|---|---|
| LEAD | someone describing the problem you solve | a real answer, then one sentence of open door |
| SUBSTANCE | adds data, disagrees, extends | the best reply on the thread |
| PEER | an account worth being seen next to | a reply that gives them something back |
| QUESTION | asks how, what, which | the answer, in the reply, not "DM me" |
| SUPPORT | "great post", an emoji, a tag | a like. A short reply to the first few at most |
| NOISE | spam, bots, bait, bad faith | nothing |
Write in that order and stop when the value stops.
How to reply
- Inside the first hour. That is when replying does the most for the post.
- Answer the actual question in the reply. Sending people to your DMs for an answer they could have had in public is how a thread dies.
- Match their length. A one-line reply does not get six lines back.
- To a critic: concede the true part first, in their words, then hold the line on the part you believe. Once. Never twice on the same thread.
- To a lead: answer fully in public. The open door is one sentence, an offer of help, not a pitch. Public value is what makes the next person DM.
- To bait, a bot, or a pile-on: nothing. A reply gives it reach. If it is abusive, hide the reply, mute, or block, in that order of force.
- Never quote-post a reply to dunk on it. That is a second post with their name in it, and mutes are weighted heavily against you.
Output
One block, grouped by bucket, each reply copy-ready and already through
/x-human:
REPLIES · 23 replies · 2 LEAD, 3 SUBSTANCE, 2 PEER, 4 QUESTION, 9 SUPPORT, 3 NOISE
LEAD
@sarahbuilds "we have the same problem with onboarding"
> The fix for us was deleting the welcome call and sending a 4-minute Loom
> instead. Completion went up, not down. Happy to send the script.
QUESTION
@devon_k "which model did you use for this?"
> Claude, with the skill file doing most of the work. The model mattered
> less than the file.
SUPPORT (like these; reply to the first three)
...
NOISE (3)
Skipped: two bots and a reply-guy account farming. Replying gives them reach.
Then the gate: nothing is posted until the user says yes, and then the user posts each reply by hand. X's automation rules only allow automated replies to people who opted in, and this is not that.