TL;DR
On this page
ChatGPT will happily write you a cold email in four seconds, and it will sound exactly like a cold email written by ChatGPT. Polished, padded, faintly corporate, and instantly deletable. The problem is not that you used AI to write it. The problem is that the client can tell, and the clients you actually want are the quickest to disengage. The fix is not editing harder after the fact. It is a prompt that builds your own voice and real personalization in from the first draft. Below is the copy-paste prompt, then how to use it for the 10 to 30 pitches a week a solo freelancer realistically sends.
This is part of the complete guide to freelancing in the AI era. If you want the wider playbook on pitch structure, see how to write a freelance email pitch; this post is specifically about getting AI to draft one without the AI smell.
The prompt
Paste this into ChatGPT (or Claude), fill the brackets, and send. It is built for a one-person business with no CRM and no sales team, just a prospect's website and two minutes.
You are helping me, a freelance [your service, e.g. "brand designer"],
write a short cold email to a potential client.
First, here is how I actually write. Match this voice exactly. Do not
make it more formal, more polished, or more enthusiastic:
"[paste 2-3 sentences from a real email or message you have sent]"
The prospect: [name], [role] at [company].
The one specific thing I noticed: [a real detail from their site, e.g.
"their pricing page lists three plans but has no comparison table"].
What I do that is relevant to it: [one sentence].
Write the email in under 80 words. Rules:
- Open with the specific thing I noticed. No compliment, no "I hope
this finds you well," no "I came across your website."
- One clear idea. One soft call to action: a question, not a request
for a 30-minute call.
- Banned phrases: leverage, synergy, circle back, game-changer, reach
out, touch base, in today's fast-paced world, I wanted to.
- Plain text. No bullet points, no bold. Sound like one person
emailing one person.
Then give me three subject-line options, each under six words, based on
the specific thing I noticed.
That is the whole tool. The rest of this post is why each part is there, because the prompt only works if you feed it the right inputs.
Why ChatGPT cold emails get deleted
Recipients have gotten very good at spotting AI, and they hold it against you. In Hunter's 2026 State of Cold Email report, built on 31 million emails sent in 2025, 69% of US decision-makers said it bothers them when an email was written by AI. Not the use of AI in the abstract, the feeling of reading something templated, synthetic, or too smooth to be a real person. That feeling is the default output of a general-purpose model, and it is what gets you deleted before the second line.
A solo freelancer who has cold-emailed for years put it more bluntly:
If your message reads like it was sent to 500 people, it's going straight to the trash.
Source: Matt Olpinski, "How to Cold Email Freelance Clients"
The trap is that raw ChatGPT output reads like it was sent to 500 people even when you only sent it to one, because the model has no real detail to anchor on and fills the gap with generic enthusiasm. The same Hunter data shows the cost is measurable: manually edited emails out-reply fully automated ones by 18% (5.2% versus 4.4%). The edit is not optional polish, it is the difference between a reply and the bin. The cleaner move is to prompt the email toward sounding human in the first place, rather than rescuing a generic draft afterward.
As Hunter's own guide on the subject frames it:
It's not whether you use AI, but whether your email still sounds like it came from you.
Source: James Milsom, "How to Use AI in Your Emails Without Losing Your Voice", Hunter.io
The fastest way to make a draft sound human is to cut the phrases that only ever show up in cold emails. The prompt bans them for you, but it helps to see what they get replaced with:
| The AI tell (cut it) | What a real person writes instead |
|---|---|
| "I hope this email finds you well" | Nothing. Open with your actual point. |
| "I wanted to reach out because..." | The specific thing you noticed about their work |
| "I came across your website" | "Your booking page sends people to a form, not a calendar" |
| "leverage, synergy, circle back" | Plain verbs: use, combine, follow up |
| "Let's hop on a quick 30-minute call" | "Worth a short reply if this is on your radar?" |
None of these are about being clever. They are about sounding like one person who noticed something, instead of a template that went out to a list.
How to use the prompt
The prompt has six load-bearing inputs. Get these right and the output is roughly 80% sendable instead of 20%.
1. Prime your voice. This is the step almost every guide skips. Before you ask for anything, paste two or three sentences you have actually written, a real email, a Slack message, a line from your portfolio, and tell the model to match that voice rather than improve it. Without a sample, ChatGPT writes in its house style, which is the synthetic register recipients are trained to ignore. With a sample, it mirrors your rhythm and word choice.
2. Lead with one real detail. Personalization is the single biggest lever, and you have to supply it, because the model cannot see the prospect's website. Spend two minutes there and find one concrete thing: a missing comparison table, a recent rebrand, a job posting that hints at a gap. Hunter found that emails with two custom attributes replied at 56% higher rates than non-personalized ones (5.6% versus 3.6%). One genuine detail in the opening line does more than any amount of clever copy.
3. Cap it under 80 words. Length is where AI drafts betray you, padding a simple ask into three paragraphs. Instantly's 2026 benchmark, drawn from billions of cold emails, found the elite performers averaged under 80 words per first-touch email, against an overall average reply rate of just 3.43%. The word ceiling in the prompt forces the model to make one point and stop.
4. Ban the tells. The banned-phrase list is doing real work. Buzzwords like leverage, circle back, and I wanted to reach out are the verbal fingerprints of machine-written and mass-sent email. Listing them in the prompt steers the model away from its laziest phrasing before it ever reaches the draft.
5. End on a soft question. A hard ask for a 30-minute call from a stranger has almost no chance. A specific, low-commitment question tied to the detail you noticed invites a reply, which is the only thing a first email needs to earn.
6. Pair the subject line. The prompt returns three subject lines built from the same detail, so the subject and body stay consistent. That matters: Hunter found subject lines with two custom attributes got 14% higher open rates (40.2% versus 35.4%). For the deeper version of this step, see freelance cold email subject lines.
pro tip
Read the draft out loud before you send it. If a sentence is something you would never actually say to a person, cut it or rewrite it in your own words. The read-aloud test catches the last 10% of AI phrasing that the prompt's rules did not.
Where to find the one real detail
The one specific detail is the input that decides whether the email works, and as a solo freelancer you are finding it by hand, with no sales-intelligence tool. You do not need one. Five free places almost always hold something worth opening with:
- Their pricing or services page. A gap, a confusing tier, or a missing option you could fix.
- Their latest blog post or case study. Reference the actual content, not the fact that it exists.
- Their about or team page. A recent hire or a stated goal that maps to what you do.
- A job posting. If they are hiring near your skill, they have both the budget and the pain.
- Their product itself. Sign up, click around, and name one concrete thing you noticed using it.
Two minutes across these is enough. The detail does not have to be impressive, it has to be true and specific, the kind of observation that proves a real person looked before writing a word.
A realistic cadence for a solo inbox
The temptation with an AI drafting tool is to scale up: if one email takes two minutes, why not send 500? Because volume without personalization is what trips spam filters and burns your sending domain, and because the data does not reward it. With a 3.43% average reply rate and top performers above 10%, the gap is personalization, not volume. For a solo freelancer, 10 to 30 properly researched emails a week is the sustainable number, the count you can actually find a real detail for. Track replies, not sends.
Using AI well here is also just good business positioning. AI-skilled freelancers earn 44% more per hour than peers who do not use the tools, and outreach is one of the highest-leverage places to put that skill to work, as long as the output still sounds like you.
After they reply: send something that closes
The cold email has one job: start a conversation. Once a prospect replies and shows interest, the next document does the actual closing, because the email got you the conversation but the proposal gets you the project. That handoff is where a polished paper trail starts to matter.
A sharp proposal is what turns a warm reply into signed work. FreelanceDesk builds clean proposals you can send the moment a lead bites, and you can draft the first version with ChatGPT using the freelance proposal prompt. When a reply is positive but slow, a short nudge keeps it alive: see freelance proposal follow-up.