Writing Difficult Emails With AI: A Method That Actually Works

The instinct when facing a difficult email, the overdue invoice, the client who's gone quiet, the pushback on a price increase, is to ask AI how to write it. That produces advice: acknowledge their concern, stay firm but warm, set a clear deadline. All true. All still leaves you staring at a blank draft, because the hard part of a difficult email was never the principles behind it. It was committing to one actual version and sending it.

The more useful move is asking for the email itself, in more than one version, so the choice you were avoiding becomes a comparison instead of a blank page.

Why does asking AI for email advice never produce an actual email?

Ask for advice on a difficult email and you'll get good, general principles: acknowledge the issue, don't get defensive, propose a next step, keep it shorter than you think. It reads like it was compiled from every business-writing guide published in the last twenty years, because it more or less was.

None of it is wrong, and none of it solves the actual problem, which is that you still have to write the email. Principles don't draft sentences. The gap between "be direct but warm" and an actual sentence that is direct and warm is exactly the gap most people get stuck in, and it's the reason a difficult email can sit in drafts for three days even after you've read every piece of advice on how to write it.

Is AI actually good at difficult emails, or does it just sound like it is?

Unaided, it's genuinely strong at generating multiple complete drafts fast, which turns "how should I write this" into "which of these three should I send," a much easier decision. It's also good at spotting an unintentionally aggressive or unintentionally weak phrase you didn't notice in your own draft.

It's weak at knowing the specific person on the other end. A technically excellent draft can still misjudge tone for a recipient who responds badly to directness, or who needs more warmth than the AI's generic "professional" register defaults to. It has no relationship history with your client; you do.

What prompts actually produce a usable draft today?

Each of the four below runs fine in Claude, ChatGPT, or Gemini as written, no setup beyond a fresh chat window.

1. The outcome brief. "I need to write a difficult email about [situation]. Before drafting anything, tell me: what are the realistic outcomes I could be aiming for here, and what would each one require the email to actually say?" Why it works: most difficult emails stall because the writer hasn't decided what they actually want, preserve the relationship, force a decision, or escalate, and are trying to write one email that vaguely accomplishes all three. Naming the outcome first forces the decision the blank page was avoiding.

2. The multi-draft request. "Write me three complete versions of this email: one that prioritizes preserving the relationship, one that forces a clear decision by a deadline, and one that escalates without sounding like an escalation. Full drafts, not summaries." Why it works: comparing three finished options is a fundamentally easier decision than staring at one blank page and hedging, and it surfaces a register you might not have thought to try.

3. The tone check. "Read this draft as if you were the recipient, someone who is [describe them, defensive, busy, already annoyed]. What in this email would land badly, and what's the smallest change that fixes it?" Why it works: writers are bad at judging their own tone under stress, and asking the AI to read as a specific type of recipient, rather than a generic one, surfaces genuinely different problems than a generic "is this too harsh" question does.

4. The read-aloud pass. "Read this draft back to me exactly as written, sentence by sentence, and flag anywhere the phrasing sounds unnatural if spoken out loud." Why it works: written-sounding phrases that read fine silently often reveal themselves as stiff or over-formal the moment they're voiced, and this catches the specific artifact of AI-drafted prose that "sounds like an email" rather than like a person.

Where does prompting alone stop working?

Run all four and you'll produce a strong, send-ready draft faster than most people manage alone. What you still won't have is a consistent register for your own high-stakes writing, a voice that reads as you across a dozen different difficult emails rather than as a slightly different AI persona each time, one drafted a little too apologetic, the next a little too clipped, depending on the mood of that particular session. The gap isn't quality on any single email; it's consistency across many of them, which a fresh prompt every time can't provide by itself.

What does the draft look like once a documented framework replaces generic tone advice?

David Ogilvy built his direct-response career on a documented, unglamorous discipline: write for the actual reader, not to sound clever, and test rather than guess at what persuades. His most quoted line on the subject, that the consumer isn't a moron, she's your wife, is really an argument against condescending or overly clever copy in favor of respectful clarity. He backed it with rigorous, published headline-testing, treating persuasive writing as something to measure, not merely feel your way toward.

Before, generic tone advice: asked to soften an overdue-invoice email, a generic AI response adds hedges and apologetic language, "we understand things get busy, no rush, but," which reads as uncertain rather than warm, and buries the actual ask under padding.

After, Ogilvy's framework applied: the framework strips the hedging and replaces it with a single, clear, respectful statement of fact and a specific next step, "This invoice is six weeks overdue. Can we get it settled this week, or would a short payment plan work better for you?" No condescension, no apology-padding, and a real choice handed to the reader instead of a vague plea. Ogilvy's own discipline of testing over guessing shows up here too: the three-draft method above is a compressed version of his test-don't-assume principle, applied to a single email instead of a headline campaign.

A generic tone request has no reason to apply that standing rule, respect the reader's intelligence, state the fact plainly, give them a real choice, on its own. The framework does, every time, regardless of how tempting the softer, hedgier version feels in the moment of hitting send.

How do you actually put a framework to work in Claude, ChatGPT, or Gemini?

Claude takes it as a Skill. Upload the .zip under Settings → Skills → Add skill, and Claude will auto-invoke it whenever a question matches the skill's description; typing /david-ogilvy-framework ahead of your message overrides that and applies it regardless. Working from Claude Code instead, the same package drops into ~/.claude/skills to cover every project, or into a single project's own .claude/skills folder if you only want it there.

ChatGPT doesn't have a Skills equivalent, so the plain .md file becomes the Instructions for a Custom GPT: Create a GPT → Configure → Instructions, paste it in. Don't use the standard Custom Instructions fields for this; they max out at 1,500 characters, nowhere near enough for a full framework, whereas a Custom GPT's Instructions field holds roughly 8,000.

Gemini has no dedicated upload path for this, so the .md content goes straight into the system prompt if your interface exposes one, or into the first message of the chat with a short lead-in: "Here is a thinking framework to apply throughout this conversation. Read it carefully, then answer using this framework's approach."

What's the next move?

David Ogilvy's documented method, along with Hormozi's, Godin's, and Schwartz's, lives in the Marketing & Sales category as .md files ready for Claude, ChatGPT, or Gemini, worth building into a habit rather than reaching for once. One specific hard email due today, though, is better served by Strategic Email, a $49 tool that runs the three-draft comparison from this guide in minutes.