How to answer your LinkedIn inbox without sounding like a bot
Short answer: the problem is almost never the model. It is that the draft was written from the shape of a reply instead of from the thread. Feed a model your own sent messages, make it answer the actual question, cut it to half the length, and read every one before it goes.
Six things that give an AI reply away
These are the patterns people recognize instantly, in the order they usually appear.
- The opening compliment. "Great to hear from you, and congratulations on the new role." Nobody who is actually replying starts here.
- Restating the message. A first paragraph that summarizes what the other person just wrote back to them. They know what they wrote.
- The contrast sentence. "I would love to help, but I want to make sure this is the right fit." Two clauses hinged on a turn. Real messages are usually one clause.
- The closing question nobody asked. "Would you be open to a quick fifteen minutes next week?" appended to a reply about something else entirely.
- Wrong length. Four paragraphs answering a question that took one line to ask.
- Wrong hour. A thoughtful reply timestamped 3:47 in the morning their time.
Fixing all six is the whole job. None of them requires a better model.
Give it your writing, not adjectives
Telling a model to sound casual and human produces a draft that sounds like a model trying to sound casual and human. Telling it nothing and showing it fifteen messages you actually sent produces something much closer.
Your own sent messages carry things you would never think to specify. How you open. Whether you use the person's name. Whether you use contractions. How long your sentences run. Whether you sign off at all. A model reads all of that off examples in a way it cannot read off instructions.
Pull fifteen replies out of your own LinkedIn history. Include short ones. Include the ones where you said no. The declines teach tone better than anything else.
Sort before you write
Most inboxes are mostly noise. Writing a careful reply to each thread in order is the slowest possible way to work.
Classify first. Recruiters reaching out. Sales pitches. Warm intros from people you know. Threads where someone is waiting on you. Threads that have gone quiet and need a nudge. Once the pile is sorted, most of it needs nothing, and the ten that matter are visible.
What to never automate
- Sending without reading. A reply goes to one person who said one specific thing. A misread thread produces a message that reads as careless, and that is worse than a slow reply.
- Anything about money, a job offer, or a decline. Draft it if you like. Rewrite it yourself.
- Replies to people you actually know. They will notice, and the relationship is worth more than the four minutes.
Questions people ask next
Can I just auto-reply to everything?
You can, and it will cost you more than it saves. LinkedIn threads are one-to-one and people compare notes. The tools that work keep drafting automatic and keep sending manual. That way the time saved is the writing time, which is most of it, and the risk stays at zero.
How long should a LinkedIn reply be?
About as long as the message you received. A one-line question gets a one-line answer. Length mismatch is the single most reliable tell, and it is the easiest thing to fix. Cut every draft by half before you send it.
Does the send time really matter?
Yes. Working hours in their time zone is the whole rule. Draft at midnight if that is when you have time, then set the send for nine the next morning in their timezone.
How do I handle recruiter messages?
Decide the answer first and let the draft follow. Interested, not now, or never. A short honest reply to the first two and nothing at all to the third handles almost every recruiter thread. The mistake is writing a polite paragraph that commits to nothing, because it invites three more messages.
Can a tool tell me if a draft sounds robotic?
Some can. Scoring each draft for how templated it reads, and dropping the low scorers before you ever see them, keeps a review queue worth reviewing. A queue full of drafts you would never send trains you to stop reading carefully.
How Iridium does it
Iridium classifies your inbox first, so you see what is actually there before you write anything. It drafts replies from the full thread and from your own example messages, scores each draft for how robotic it reads, and drops the weak ones before they reach your queue.
Nothing sends until you approve it. You can rewrite a draft, edit it yourself, skip it, or schedule it. You choose the send time, so a draft written at midnight can go out at nine the next morning. Every message that goes out is written to a log with its exact text.
It runs as a web app and as an MCP server, so you can work the inbox from a dashboard or by asking Claude to do it.
Get drafts that sound like you
7-day free trial. No credit card. Cancel anytime.
Start free → Watch the demos