Quick answer
To write email newsletter using AI that sounds human, skip tone adjectives entirely. Paste two of your own past issues into the prompt as style anchors, add an explicit banned-word list, and put this week's raw notes underneath. Store all three permanently, then spend five minutes editing the draft before sending.
Your draft came back in thirty seconds and it read like a press release from a company you have never heard of. Search for how to write email newsletter using AI and you will find a hundred prompt lists, nearly all of them telling you to ask for a warm, conversational tone. That advice is the reason your draft sounds the way it does.
Adjectives are the problem. When you tell a model to be “friendly and professional,” it has no idea what those words mean to you — so it falls back on the average of every friendly, professional email ever published on the internet. That average is a marketing brochure. What actually works is showing it two things you already wrote.
- Adjectives describe. They leave the model guessing at your rhythm.
- Samples demonstrate. They hand it your sentence lengths, your vocabulary, your habits.
- Bans protect. They strip the giveaway words before they reach your subscribers.
That three-part fix is the Show-Ban-Store Method we teach at PromptPeakAI: show the AI two real samples of your writing, ban the words that give it away, then store both so you never rebuild them again. Setup takes about twenty minutes once. Every issue after that takes five.
Will this work for you?
Why AI Newsletters Sound Automated
AI newsletters sound automated because the model is optimising for a reader who does not exist — a generic professional who wants balanced, hedged, evenly weighted prose. Your subscribers signed up for a person, not an average.
Two prompts, two very different outputs — and the only difference is what you fed in.
The tells your readers notice
Most people cannot articulate what gave it away, but they feel it. The paragraphs are all the same length. Every sentence runs fifteen to twenty words. There are exactly three items in every list. Nothing is ever slightly awkward, which is itself the giveaway — real writing has texture.
✕ Adjective-prompted draft
As we move into the second half of the quarter, it is important to remember that our onboarding process continues to evolve. We remain committed to delivering a seamless experience, and we will continue to leverage feedback to drive meaningful improvements across the board.
✓ Sample-anchored draft
Onboarding changed on Monday. New hires now get their laptop before day one instead of on day three, which was the single loudest complaint in the last survey. Three people have gone through the new version. All three said it was less chaotic.
Why the fix is not editing harder
You can absolutely rewrite the first draft into something human. Most people do, and it takes them longer than writing from scratch would have. The point of this method is to move the fix upstream, into the input, so you are polishing rather than rescuing.
Show Samples, Not Tone Adjectives
Few-shot prompting means giving the AI two to four real examples of your writing alongside your instruction, so it copies observable patterns instead of guessing from adjectives. It is the single highest-value change most professionals can make to their input, and it takes about four minutes to set up.
Three moves, done once. The weekly loop that follows takes five minutes.
How many samples you need
Two is the working minimum. Three is better. Past four you are mostly adding noise, and on a free plan you will run into storage limits before you see any benefit.
Pick issues you were happy with rather than your most recent ones. If you have never sent a newsletter, use two emails you wrote to a colleague you like — the register is close enough, and it is far better than nothing. The mechanics behind this are covered properly in our guide to few-shot prompting for professionals.
Where the samples go
- Near the end of the prompt, not the top — models weight recent instructions more heavily on longer generations.
- Wrapped in a clear label so the AI knows this is reference material, not content to summarise.
- Whole issues, not excerpts. Openings and sign-offs carry more of your voice than the middle does.
Not sure what your own patterns actually are? Ask. This takes one minute and the answer becomes the shortest version of your voice file.
Below are two newsletters I wrote. Describe my writing patterns as observable rules someone else could follow. Cover: - Average and range of sentence length - How I open and how I sign off - Words and phrases I use repeatedly - Formal writing rules I break on purpose - What I never do Give me 8 to 12 short rules. No praise, no summary of the content. [Paste past issue 1] [Paste past issue 2]
This is the Context element of the Logic System doing the heavy lifting. Most people write a strong Role and Goal and then leave Context nearly empty, which is exactly the field that determines whether output sounds like you. Our guide on giving AI usable context goes deeper on that imbalance.
How to Write Email Newsletter Using AI
To write email newsletter using AI that sounds like you, paste your instruction, two past issues, this week’s raw notes, and an explicit banned-word list into one prompt — then edit the draft for five minutes. Here is the full sequence.
Gather this week's raw material
Open a blank note and dump everything you might mention, unedited. Bullet fragments are fine. Voice-to-text is fine. This should take four minutes and look messy — the AI is going to structure it, so structuring it yourself is wasted effort.
Pick your two anchor issues
Copy two past newsletters into the same note. Strip out anything confidential before you go anywhere near a chat window; there is a full section on that below.
Assemble the master prompt
This is the block that does the work. The instruction sits at the top, the constraints in the middle, and your samples at the bottom where the model will weight them most heavily.
You are drafting my weekly newsletter. You are not writing it from scratch — you are converting my raw notes into my existing voice. STEP 1: Read the two issues inside <samples>. Note my average sentence length, how I open, how I sign off, the words I actually use, and where I break the rules of formal writing. STEP 2: Turn the material inside <raw_notes> into a [300]-word newsletter that matches those patterns. STRUCTURE: - One-line hook that states the point, not a wind-up - The main item in [2-3] short paragraphs - [2] secondary items, two sentences each - A sign-off in my usual style RULES: - Maximum 3 sentences per paragraph - Vary sentence length deliberately: mix 4-word sentences with 20-word ones - Active voice throughout - Never invent a fact, number, name, or quote that is not in <raw_notes>. If something is missing, write [GAP] and move on. BANNED WORDS AND PHRASES — do not use any of these: delve, tapestry, testament, pivotal, seamless, robust, leverage, utilise, synergy, landscape, realm, navigate, unlock, elevate, game-changer, in today's fast-paced, it is important to note, moreover, furthermore, in conclusion, dive in, at the end of the day <samples> [Paste past issue 1] [Paste past issue 2] </samples> <raw_notes> [Paste your messy notes] </raw_notes>
The banned list is the part people skip, and it is doing more work than the samples on the first pass. If you want the reasoning behind constraint-style prompting, we covered it in the guide to negative constraints as a prompting technique.
Order matters. Samples sit at the bottom because that is where the model weights them hardest.
Read the draft against your samples
Put the draft beside one of your anchor issues and read both openings aloud. If they sound like the same person, the input worked. If they do not, the fix is almost always a thinner banned list or samples that were too polished to be representative.
You should now have a draft that needs editing rather than rewriting — roughly five minutes of work, not forty.
Where to Store Your Voice File
Everything above assumes you plan to write email newsletter using AI more than once. Storing the instruction and the samples somewhere permanent is what turns a clever prompt into a system. Otherwise you are pasting the same eight hundred words into a fresh chat window every week, and eventually you stop bothering.
Every major tool has a place for this, and the differences matter more than the marketing suggests.
| Where it lives | What it holds | Best when |
|---|---|---|
| Account-wide instructions field | Short instruction plus a trimmed banned list | You use one AI for everything |
| A project or workspace | Full instruction plus uploaded past issues | You send weekly and have real samples |
| A saved custom assistant | The whole block as a reusable helper | You want one-click drafting |
| A shared team workspace | One voice file several people draw on | More than one person sends it |
ChatGPT: the instructions field
Settings, then Personalization, then Custom Instructions. According to OpenAI’s Custom Instructions documentation, the feature is available on every plan — including free — but the storage limit differs: 1,500 characters on Free and Go, 5,000 on the paid tiers. That cap is the real constraint here, because 1,500 characters is roughly a paragraph and a half. On a free plan, store the instruction and the banned list, and paste your samples per session.
The ChatGPT Custom Instructions panel with a short voice instruction and banned-word list typed into the field.
Crop to the settings panel only. Hide the account name, email address, sidebar chat history, and any workspace name.
File: chatgpt-custom-instructions-voice-samples.png · Alt: "ChatGPT Custom Instructions field holding a newsletter voice instruction and banned word list"Here is a version trimmed to fit inside the smaller limit. Paste it once and it applies to every new chat.
When I ask for newsletter copy, you convert my notes into my voice. You never compose from scratch. Always: - Maximum 3 sentences per paragraph - Vary sentence length deliberately, short next to long - Active voice, first person - Lead with the point, never a wind-up - Mark anything missing as [GAP] instead of inventing it Never use: delve, tapestry, testament, pivotal, seamless, robust, leverage, synergy, landscape, realm, unlock, elevate, game-changer, moreover, furthermore, in conclusion, it is important to note. If I paste past issues, treat them as the style reference and match them.
Our full walkthrough of setting up ChatGPT custom instructions covers the two-box structure and what belongs in each.
Claude: a project workspace
A project holds both standing instructions and uploaded files, which is the cleanest home for a voice file with real samples attached. Claude’s official Projects documentation describes project instructions and a knowledge base you upload documents into, and notes that the expanded project-knowledge capacity is a paid-plan feature. Practically: upload your last five issues as files, put the instruction and banned list in the project instructions, and every chat inside that project starts primed.
A Claude project sidebar showing the project instructions field and several past newsletter files uploaded to project knowledge.
Crop to the project panel. Hide the account initials, organisation name, other project names, and any file names containing client or company identifiers.
File: claude-project-knowledge-newsletter-samples.png · Alt: "Claude project knowledge panel with past newsletter issues uploaded as voice reference files"We cover the setup end to end in the guide to using Claude Projects to organise your AI work.
Gemini: a saved Gem
A Gem is a saved assistant with its own standing instructions, which suits people already working inside Google Docs and Gmail. Paste the whole master prompt into the Gem’s instructions, leaving only the raw-notes slot to fill each week. The guide to building Gemini Gems for recurring tasks walks through it.
Pick the storage route by how often you send and how many people write the thing.
Whichever you pick, the discipline is the same one described in our piece on scaling content production without losing your voice: the voice file is a living document. Every time a draft slips out a word you hate, add it to the banned list.
The Five-Minute Human Polish Pass
No stored prompt gets you to send-ready. The whole reason to write email newsletter using AI is speed, and speed collapses if you treat the draft as finished. The polish pass is short, but it is the difference between a newsletter that sounds like you and one that merely avoids sounding like a robot.
- Add one thing only you know. A detail from Tuesday, a reaction someone had, something you noticed. This is the highest-value edit and it takes thirty seconds.
- Break one sentence in half. AI drafts drift toward even rhythm. One deliberate short sentence resets it.
- Check every fact and number. Anything the model filled in that was not in your notes is a guess.
- Rewrite the opening line. It is the sentence most likely to sound generic and the one most people read.
- Read the sign-off aloud. If you would not say it, change it.
✕ Straight from the draft
We are excited to share some updates on the project and look forward to keeping you posted as things continue to progress over the coming weeks.
✓ After thirty seconds of polish
The pilot finishes Friday. I will send numbers next week — early signs are better than I expected, which is not something I say often.
If a draft needs more than this, do not start over. Work out which input caused the gap and fix that one element, using the diagnostic approach in our guide to refining AI output without starting from scratch.
Here is my edited draft. Do not rewrite it. Instead, list only: 1. Any sentence that still sounds like generic corporate writing, quoted exactly 2. Any claim, number, or name that does not appear in my original notes 3. Any paragraph longer than 3 sentences Return a numbered list and nothing else. No compliments, no summary. [Paste your edited draft]
Keeping Client Data Out of Prompts
This is where most workplace newsletter workflows quietly go wrong. Past issues are exactly the kind of document that contains headcount, client names, unreleased dates, and internal metrics — and you are about to upload several of them.
Published issues, your own writing, anonymised examples.
Structure notes, style rules, banned-word lists.
Client and colleague names, internal metrics, headcount figures.
Anything about a deal, product, or date not yet public.
Salary data, personal information about employees, anything under NDA.
Redacting is faster than it sounds: swap real names for role labels once, and your samples stay reusable forever. “Client A” teaches the AI your sentence rhythm just as well as the real name does.
Two things worth knowing before you store anything permanently:
- The instructions field is not a vault. OpenAI's own documentation notes that content from your use of custom instructions may be used to improve model performance unless you opt out. Treat it as you would a shared drive.
- Consumer accounts and business accounts behave differently. The meaningful line is not free versus paid — it is a personal account versus a company-provisioned business or enterprise account, where no-training is the contractual default. If your employer has a licence, use it.
If you are working under a company AI policy, our guide on whether ChatGPT is safe to use at work covers the settings and the conversations to have with IT before you upload anything.
When Voice Matching Still Fails
Three failures account for nearly everything that goes wrong, and each has a specific cause rather than a vague “the AI is bad at this.”
| What you see | Why it happens | The fix |
|---|---|---|
| Banned words still appear | List is too short or too vague | Name exact words, not categories |
| Voice fades halfway down | Long generation, early instructions decay | Move samples lower, cap length |
| Reads like an essay | No structural limit was set | Add a hard 3-sentence paragraph cap |
| Facts you never mentioned | Model filled a gap silently | Add the do-not-invent rule |
| Every issue sounds identical | Same samples, same structure | Rotate samples every few months |
Drift on longer pieces is the most common one. The fix is mechanical rather than clever.
The second half of this draft stopped matching my voice. Re-read the samples in <samples>. Rewrite ONLY the section starting at "[paste the first line that goes wrong]" to the end, matching the sentence length and vocabulary of the samples. Keep every fact identical. Change nothing before that line. <samples> [Paste past issue 1] </samples>
Honest limits, because nobody else will tell you:
- AI cannot decide what is worth saying. It will give equal weight to a major announcement and a minor schedule change. Editorial judgement stays yours, permanently.
- It cannot make a boring week interesting. If you have nothing to say, a voice-matched draft about nothing still says nothing.
- It will not catch a wrong number. Verification is a human step every single time, and the one people skip when they are in a hurry.
- Two samples will not survive a genuine change in your voice. If you have shifted roles or audiences, rebuild the file rather than patching it.
- Nobody should be reading a newsletter you did not think about. The method saves you drafting time, not thinking time.
Will subscribers know? Not from the prose, if you follow this properly. They will notice if the content is thin, if the details are vague, or if the enthusiasm sounds borrowed — and none of those are prompting problems. They are a sign the week genuinely had nothing in it, which is worth knowing.
Set the voice file up once and the weekly cost drops to about ten minutes: four to dump your notes, one to generate, five to polish. That is what it looks like to write email newsletter using AI as a system rather than a lucky prompt. That is a small enough number that you will actually keep sending it, which was always the harder part. If you want prompt structures for the rest of your week, the free AI work templates library is the place to start.
Build your voice file this week
- Pick two past issues you were genuinely happy with Run the voice-audit prompt to turn them into 8–12 written rules Write your banned list — start with the words you already hate Store the instruction and the banned list in your tool's instructions field Draft one issue with the master prompt, then polish it for five minutes Add every word that slips through to the banned list