返信の攻略ガイド · AI
How to reply to AI posts on X without adding to the hype: post types, reply angles grounded in real use, six sample replies and mistakes to avoid.
AI on X moves faster than any other niche. Researchers post papers, developers post demos and benchmarks, product people share prompts and workflows, and a large group reposts every launch with a thread of excitement. New model releases bring thousands of replies within an hour, most of them saying the same thing.
The replies that stand out come from people who actually used the thing. A reply saying what broke when you tried the new model on your own codebase, or which prompt change fixed an agent loop, is worth more than any reaction. Builders and researchers answer concrete reports because they want the feedback, and their followers want to find people who test things.
01
A new model, a chart where it beats the others. The replies are a flood of reactions. A short report from a real task you ran is the reply that gets read.
02
"I built an agent that does X in a weekend." Ask how it handles the failure case, or share how you solved the same problem.
03
System prompts, coding setups, tool chains. Add the tweak you found or the case where the tip stops working.
04
Jobs, timelines, open versus closed models. These get heated. A calm reply with one concrete observation does better than a big opinion.
01
Say what you ran, on what kind of input, and what happened. "Tried it on our support tickets, it handled X and missed Y" is exactly the feedback this niche values.
02
Demos show the happy path. Asking what happens with long inputs, ambiguous instructions or a tool error shows you have built something yourself.
03
Latency, cost per call, context handling, how it behaves on your language or domain. Name the dimension that matters for real use and is missing from the chart.
04
A one-line prompt change, a retry rule, a way to structure output. Small practical fixes are the most saved and quoted replies in this niche.
05
Weekly "this changes everything" threads are easy to tease. Keep it light and about the pattern, not the person who posted.
このページのために書いた投稿例です。実在のアカウントではありません。
ポスト
New model just dropped and it tops every coding benchmark. This is a huge leap.
返信
Ran it on a real refactor this morning. Better at reading the codebase, still renamed a function it was told to leave alone.
ポスト
I built an AI agent that handles my whole inbox. It reads, sorts and replies to emails automatically.
返信
what does it do with an email that needs a yes or no from you? that's the part I never managed to automate safely
ポスト
Prompting tip: tell the model to think step by step and your results get way better.
返信
Helped a lot on older models for us. On the newer ones, giving it two worked examples of the output did more than the instruction.
ポスト
In two years nobody will write code by hand anymore.
返信
Maybe. Today the slow part for my team is reviewing the code, and that got slower when we started generating more of it.
ポスト
Hot take: RAG is dead now that context windows are huge. Just put everything in the prompt.
返信
Works until the bill arrives. We tried it on our docs and paid for the full manual on every question about password resets.
ポスト
Every week: "this new AI tool changes everything." Every week I go back to the same three tools.
返信
my bookmarks folder of tools that changed everything is now longer than my list of tools I open