AI cold email copy gives itself away. It is too polished, slightly eager, and it reads like every other AI-written email in the inbox. The gap between that and copy that sounds like you is almost entirely in how you brief the tool.

Most people treat a prompt like a search box: a few words, hit enter, hope. Here is the structure we use instead, and the review layer that sits underneath it.

Why the output feels generic

Ask for a cold email and you get something professional and forgettable. The problem is not the model. It is what you handed it.

We were writing personalised opening lines for a client selling into aviation and aerospace. The first batch came back like this:

Have you considered how a custom booth design could showcase your engineering innovations more effectively at industry events?

That could have been written for any company in any industry. Nothing in it knew who the client was, what made them different, or what the prospect cared about.

So we added the client’s business model, their value proposition, and the competitive landscape. Better, and still flat. The model had the information but no idea which parts of it mattered.

Then we added research on each prospect’s own business and what they actually sell. Same model, same campaign:

Have you considered how to make your precision NXT sensor technology’s complexity feel intuitive and unmatched to aerospace buyers?

That one is about the prospect rather than about us. The shift is to stop writing prompts as commands and start writing them as briefs. You are not giving orders to an intern. You are briefing someone who has every piece of information you have and no judgement about which of it matters.

Work in projects, not one-off prompts

Both ChatGPT and Claude let you create a workspace that holds context across conversations. Every client gets one, with separate threads for research, personalisation, copy, and strategy.

Inside it we upload the knowledge files:

• A client brief covering who they are and what they do

• An ICP document naming exactly who we target

• Past emails that performed

• Framework docs describing how we work

• Performance data from previous campaigns

Before we organised this way, every new request meant re-explaining the client, the voice, and what had already been tried. Now that context lives in the project, several people can work the same account without colliding, and nobody digs through chat history looking for a decision made two weeks ago.

The four elements of a prompt that works

Good prompts are not clever. They are specific to the point of being tedious.

1. Context

Explain what the email is for, who it is going to, and what you want them to do after reading it. Without the bigger picture the model guesses at tone, angle, and approach, and it usually guesses corporate.

2. The outcome, exactly

Not “write something personal” but “suggest one specific, actionable idea this prospect could implement based on their business, under 20 words, natural in conversation, and give them a reason to reply.” Include word counts, structure, and reading level. Being specific upfront costs less than revising afterwards.

3. Constraints

Tell it what not to do. No generic compliments. No claims we cannot verify. No phrasing that sounds like it is trying too hard. The list of things to avoid is usually longer than the list of things to do, and it is what stops the model failing in the same predictable ways every time.

4. Examples, not adjectives

Describing a tone rarely lands. Showing three or four pieces of copy that hit it does. The model picks up sentence length, question use, formality, and how much jargon you tolerate far faster from samples than from description.

Why human review is not optional

AI does not write our final anything. It drafts, it researches, it suggests angles. Everything that goes out is reviewed by a person first.

We learned that the hard way. The model once proposed an auction featuring community leaders for a Black community nonprofit, and a chopsticks-themed campaign for an Asian cultural organisation. Both would have been serious failures if they had gone out unchecked.

Approval happens at the template level, not per email. A sequence might have three steps with four variations on the first, three on the second, and one on the third, and we review every variation. Nobody reads every individual send.

The real check sits earlier, during enrichment. Keyword filters flag words we avoid, a safety model reviews the output, and anything either one flags gets rewritten by hand. Cultural sensitivity and tone are what we watch hardest.

What still needs a person

• Strategy. What to run, who to target, how to position it.

• Anything touching tone or cultural context, where a technically correct suggestion can be completely wrong.

• Client relationships. AI helps you prepare. It does not replace understanding what someone actually needs.

• Creative judgement. Knowing when a draft is good enough, and when an idea is clever but would not survive a real prospect.

One hard line: never let AI reply to interested prospects or book calls. Once someone is in a conversation, a person handles it from there.

Where to start

Write down how your business works before you open any tool. Two or three pages covering your value proposition, how you operate, who you sell to, what has worked, and what has not. Garbage in, garbage out applies here more than anywhere.

Create a project, upload that document along with anything else relevant, and pick one use case. Research summaries, or first-draft openers, or subject lines. Get that one right before expanding to the next.

Your first prompts will be mediocre. The quality comes from iterating on the brief, not from finding a better model.