We analyzed 500 cold DMs last month. AI-generated messages got under 2% replies. AI-assisted messages (where someone actually knew how to prompt) got 11%.
Same AI tools. Same prospects. 5x difference.
The gap? Your prompt. Garbage in, garbage out. Nice.
The Problem: You're Prompting Like Everyone Else
Here's what 90% of people type into ChatGPT:
❌ Generic prompt:
"Write a cold DM to a potential client about my marketing services"
And here's what they get back:
❌ Generic output:
"Hey [Name], I hope this message finds you well! I came across your profile and was impressed by your work. I help businesses like yours grow through strategic marketing. Would you be open to a quick chat to see if there's a fit?"
You know what your prospect sees when they open that DM?
The same message they've received 47 times this week. (This is exactly why X DMs get ignored.)
We talked to a founder with 50k followers. He showed us his DM requests. Literally 50+ variations of that exact message. Every single day.
The Fix: Prompts That Actually Work for X Outreach
Four parts make a good prompt:
Element
What It Does
Example
Prospect Intel
Makes the message actually personal
"They just tweeted about struggling with X algo"
Specific Angle
Gives the AI a direction
"Lead with a question about their recent thread"
Voice Examples
Stops the LinkedIn-speak
"Match this tone: [paste your tweets]"
Hard Constraints
Kills the AI cringe
"No 'hope this finds you well,' max 3 sentences"
The X Outreach Prompt Template
This is the exact template we use. It works for DMs, reply-to-engage, and connection messages. (Want ready-made templates instead? See our cold DM templates.)
CONTEXT: I'm reaching out to [specific person] who has [follower count range] followers. They recently [specific action - tweeted about X, launched Y, complained about Z]. Their content focuses on [their niche]. My offer: [what you actually help with, specific outcome]. THEIR RECENT CONTENT: [Paste 1-2 of their recent tweets that are relevant] TASK: Write a DM that [specific approach]: - Option A: Asks a genuine question about their recent tweet - Option B: Offers a specific, actionable piece of value - Option C: Points out something I noticed in their content CONSTRAINTS: - Maximum 3 sentences - No "I hope this finds you well" - No asking for calls/meetings in first message - No compliments that could apply to anyone - Sound like a peer, not a salesperson - Match this tone: [paste 2-3 of your own tweets] OUTPUT: Give me 3 variations to choose from.
That's a lot more work than "write me a cold DM." It's a pain in the ass.
It also gets 5x the replies. Your call.
Example: Before and After
Let's say you're reaching out to a SaaS founder with 25k followers who just tweeted about their content not getting reach anymore.
❌ Generic prompt produces:
"Hey! Love what you're building. I noticed you're in the SaaS space and thought we might be able to help each other. I specialize in helping founders like you get more visibility. Would love to connect if you're open to it!"
Now with our template:
✓ Specific prompt produces:
"Saw your tweet about reach dying. We tested something with 3 SaaS founders last month - posting threads between 7-8am EST instead of 9am got 40% more impressions. Algo seems to favor early momentum right now. Might be worth testing?"
The second one references their specific problem, offers actual value, includes a real number, and doesn't ask for anything. It sounds like someone who's been in the trenches, not someone running a spray-and-pray outreach campaign.
Hot Take: Your Constraints Matter More Than Your Instructions
Most people tell AI what to write. Backwards.
Tell it what NOT to write.
AI has default patterns. If you don't explicitly ban them, they'll show up. Every time.
X Outreach Banned Phrases (add these to every prompt):
- • "I hope this message finds you well"
- • "I came across your profile"
- • "I was impressed by your work"
- • "Would you be open to a quick chat"
- • "I help [vague category] achieve [vague outcome]"
- • "I'd love to connect"
- • Any compliment that could apply to 1,000 other people
- • Questions ending with "?"that are actually pitches
We've tested this. Same prompt with and without banned phrases. The version with explicit constraints got 3x better outputs on the first try.
Less editing. Better messages. More replies.
The Voice Problem (And How to Fix It)
Even with good prompts, AI defaults to LinkedIn-speak. Professional. Polished. Boring.
X doesn't work like LinkedIn. The accounts that win on X sound like actual humans. Slightly unpolished. Opinionated. Sometimes a bit spicy.
Getting AI to match your voice:
MY VOICE (match this exactly): Example tweets I've written: 1. "Threads are dead. Single tweets with hot takes are back. The algo rewards controversy, not length." 2. "You don't need 10k followers to get clients on X. You need 50 real relationships. Stop optimizing for vanity." 3. "Just spent 2 hours in someone's DMs. No pitch. Just helped them fix their funnel. That's how you sell without selling." Voice characteristics: - Short sentences - Occasional single-word reactions - Specific numbers when possible - Slightly contrarian - No emojis - No hashtags - No questions at the end of posts
Now the AI has patterns to follow. Your DMs will sound like an extension of your feed, not like a different person suddenly got formal in the inbox. For more on this, see how to make AI DMs sound human.
Prompt Templates for Different X Outreach Scenarios
For Reply-to-Engage (Warming Up Prospects)
CONTEXT: I want to add value to this person's thread before ever DMing them. Their tweet: [paste the tweet] My expertise: [what you actually know about] TASK: Write a reply that: - Adds a specific insight they didn't mention - OR respectfully disagrees with one small point - OR shares a relevant personal experience CONSTRAINTS: - Max 2 sentences - No "great post!" - No just agreeing with everything - Must add something new to the conversation
For Cold DMs (First Touch)
CONTEXT: First-time DM to someone I've never interacted with. They have [X] followers and post about [topic]. Their recent tweet that caught my attention: [paste it] What I could actually help them with: [specific thing] TASK: Write an opening DM that leads with value, not a pitch. Either ask a genuine question OR share something useful. Do NOT ask for a call or meeting. CONSTRAINTS: - Max 3 sentences - Must reference something specific they posted - No "I help people like you..." - Must sound like it was written for this one person
For Follow-Up DMs (After Engagement)
CONTEXT: I've been engaging with this person's content for [X] days/weeks. They've liked/replied to [X] of my comments. Their main challenge based on their content: [what they complain about] How I can help: [specific solution] TASK: Write a DM that transitions from content engagement to conversation. Reference our previous interaction OR their ongoing struggle. Offer one specific piece of help. CONSTRAINTS: - Max 4 sentences - Must feel like natural progression, not sudden pitch - One clear call to action (not "hop on a call") - Could be: "Want me to send you [specific resource]?"
The Numbers That Actually Matter
We track everything. Here's what we've learned from 2,000+ X outreach messages:
Metric
Generic AI DMs
Prompted AI DMs
Reply Rate
1.8%
11.2%
Positive Replies
0.4%
7.3%
Meetings Booked
0.1%
2.1%
Time to Write (per DM)
30 seconds
3 minutes
Yes, prompted DMs take 6x longer to create.
They also convert 21x better.
Do the math. 20 well-prompted DMs beat 200 generic ones. Every time.
The Automation Trap (Don't Fall For It)
Someone's going to read this and think: "Great, I'll automate my prompts too."
Don't.
X's algorithm detects automated DMs. The patterns are obvious: same structure, same timing, same send velocity. You'll get shadowbanned within a week.
We've seen accounts with 30k+ followers get their DM privileges revoked. Permanently. Because they tried to scale what shouldn't be scaled.
The right workflow:
- Spend 15 minutes researching 10 prospects
- Use AI to draft 10 personalized messages (3 min each = 30 min)
- Review and tweak each one (1 min each = 10 min)
- Send manually, spread throughout the day
That's ~1 hour for 10 high-quality outreach messages. If even 1 converts to a call, it's worth it.
Frequently Asked Questions
Why do AI-generated DMs get ignored on X?
Because they sound like every other AI-generated DM. When you prompt with "write a DM to a potential client," the AI produces the same template 10,000 other people are sending. Your prospect has seen this exact message 50 times this week. The fix: prompt with specifics about the person, their recent content, and the exact problem you solve.
What's the best AI prompt for X outreach DMs?
Start with context about your prospect: paste their recent tweet, mention their industry, note their follower count range. Then specify the angle: are you complimenting, asking a question, or offering value? Finally, include constraints: no "I hope this finds you well," no asking for calls in the first message, no more than 3 sentences. The more specific your prompt, the less editing you'll need.
How do I get AI to write in my voice for X content?
Paste 3-5 of your best-performing tweets into the prompt and tell the AI to match the tone exactly. Include specifics: "I use short sentences, I'm slightly sarcastic, I never use emojis, I reference specific numbers." Without examples, AI defaults to generic LinkedIn-speak that will tank your engagement.
Should I use AI to automate my X DMs?
No. Automated DMs get you shadowbanned and destroy your reply rates. Use AI to draft messages faster, not to send them automatically. The best use case: AI generates 10 personalized DM drafts based on prospects you've researched, you review and tweak each one, then send manually. Volume without automation.
Bottom Line
AI doesn't write bad DMs. Bad prompts do.
Give the AI prospect intel, specific angles, voice examples, and hard constraints. You'll get messages that actually sound like they were written for one person, because they were.
The 5 minutes you spend on a good prompt will save you from sending 100 messages that go nowhere.
Your move.