How to Reply to a Difficult Email with AI

Anyone staring at a rude, vague, or political inbox thread can reply to a difficult email with AI in 15 minutes: extract facts, name one ask, draft cold, then do one human pass. Run it in ChatGPT, Claude, or Gemini before you type a sentence you cannot unsend.

The model is not your therapist. It is a clerk. Ask it to "make this nicer" and you get fog. Ask it for facts and one ask, and you get a reply someone can act on.

This loop is a skill, not a vent. Start with what a skill prompt actually is, then run the steps until they are boring.

How to reply to a difficult email with AI in five steps

1. Do not draft first

Paste the incoming email. Instruct ChatGPT, Claude, or Gemini: "List only facts, dates, names, and asks. No tone notes. No advice. If something is implied, mark IMPLIED. Do not draft a reply yet."

Helpful models skip this. They hear conflict and start composing. Hold them to a mirror.

If the paste is a ramble, treat it like a dump. The same discipline that turns a messy note into a one-page brief applies: raw first, job second, shape third.

2. Name one job and one ask

One sentence. Verb plus object. "Confirm the Friday ship date." "Decline the extra scope." "Ask for the invoice number." Not "clear the air."

Send that sentence back with the fact list. Instruct: "The job is this sentence. If the incoming email contains more than one ask, list them and stop. I will pick one. Do not merge asks into a 'comprehensive' reply."

Merging is the silent failure. A merged reply sounds thorough and answers nothing. Park the extra ask.

3. Draft cold, with a banned list

Now, and only now, ask for the draft. Hard rules:

• Three short paragraphs, or fewer.
• First line answers the ask or states the decision.
• No apology unless you named one in the job.
• No "I hope this email finds you well."
• No new facts. If a date is not in the list, write UNKNOWN.

Banned words: circling back, as per my last, going forward, let's take this offline, happy to hop on a call used as a dodge, per my understanding.

Tell the model: "Draft the reply in my one-word tone. If a banned phrase appears, rewrite the whole draft. Do not swap in a synonym."

4. One human pass at speaking speed

Read the draft out loud. If you would not say a line on a call, cut it.

Check three traps:

• A decision buried in paragraph three.
• A concession you did not authorize.
• A CC list the model invented "for visibility."

Fix those in your own words. Send the edited draft back: "Replace only the lines I changed. Do not re-expand. Do not add warmth."

One pass. If you need a place to keep the banned list, save it with the Promptcrates tools for reusable writing skills.

5. Send, or park the second ask

If a second ask survived the fact list, park it. One line at the end is enough: "I will answer the budget question in a separate note."

Then stop. The skill is done when send is the only remaining work.

Reusable skill card

Trigger: A rude, vague, or political email just arrived and you want to reply before the feeling writes it.

Input: The incoming email or thread, one job sentence (verb plus object), one tone word (plain, firm, or brief), plus your banned-phrase list.

Output: A fact list with IMPLIED marks, then one reply of 80–160 words, first line is the decision or the ask, no new facts, UNKNOWN where the thread is silent.

Stop: After one read-aloud pass and one fact check. No second rewrite. No merged second ask. No invented CC.

Takeaways

• Facts first. Job second. Draft third. Never in one prompt.
• One ask per reply. A merged reply is a meeting in disguise.
• UNKNOWN is better than a date ChatGPT, Claude, or Gemini guessed.
• One human pass at speaking speed. Then send or park.