How to write an AI-assisted newsletter people actually open

A weekly research-to-send workflow with a 50%+ open rate. Three tools, three hours, zero recycled prompts.

The "I made an AI newsletter" graveyard is enormous. Most fail for the same reason: they sound like every other AI-generated newsletter, because they were written from the same generic ChatGPT prompts. The ones that work treat AI as a research and drafting accelerator, not a ghost writer.

This is the weekly workflow I use to ship a newsletter with a 50%+ open rate and a real reply-to-author rate. It's a Sunday-afternoon ritual.

The three-tool stack

You'll also need a sending platform, Beehiiv, Substack, ConvertKit, or Buttondown all work. The platform is the least important choice; pick whichever feels lightest.

• Perplexity for research and sourcing.
• Claude for drafting and rewriting in your voice.
• Midjourney for the header image that makes the issue feel curated.

The thing nobody tells you about AI newsletters

The reason a "ChatGPT, write a newsletter about AI" issue gets a 12% open rate next week is that the AI has no taste. Taste is what makes a newsletter feel like a person. Taste is which 6 things to include from the 60 things that happened this week. Taste is the one detail you remember when you tell a friend about it.

Your job, every week, is to bring the taste. The AI does the heavy lifting on everything else.

Step 1: Pick five things, ruthlessly (45 minutes)

Open Perplexity Pro. Run 4-5 search queries for the past 7 days in your topic. Examples for an AI newsletter:

Perplexity gives you sourced summaries with citations. Skim everything. Then close it and write down, from memory, the five things that stuck with you. Memory is taste. The five things you remember without scrolling are the five things your readers will remember.

For each of those five, prompt Perplexity again with the specific topic and "give me 3 primary sources, include the original announcement and any teardown analyses." Save those as bookmarks.

• "Most significant AI model releases this week"
• "AI funding rounds last 7 days"
• "AI policy news this week"
• "Notable AI failures or controversies past 7 days"
• "Best AI demo or experiment shared this week"

Step 2: Set up a Claude Project for your voice (one-time, 15 minutes)

This is the move that separates AI-newsletter slop from one that sounds like you. In Claude, create a Project called "[Your Newsletter] Drafting". Add as Project files:

Now every conversation in this Project starts from your voice, not the generic Claude voice. The difference is night and day after issue 3.

• Your last 5-10 newsletter issues (paste each into a .md file)
• A "voice guide" document: 5 sentences describing your tone (e.g., "Direct, slightly sardonic, never breathless, prefers concrete numbers to adjectives")
• A "format guide": describe your section structure, how long your intros are, how you do callouts, etc.

Step 3: Draft each section as a separate Claude conversation (60 minutes)

For each of your five stories, open a new chat in your Project and prompt:

Draft the [story name] section for this week's issue. Sources: [paste 3 Perplexity links and 1-2 quoted snippets]. The angle I want is [your one-line take]. Match the voice from previous issues. 120-180 words. Include the most concrete number you can find.

Read the draft. Don't accept it. Edit one sentence, even just to add your specific "I" phrasing, so it stops sounding like Claude and starts sounding like you.

The editing-pass rule that works: rewrite the first sentence and the last sentence of every section by hand. The middle can be 90% AI. The bookends are where readers feel the human.

Step 4: Write the intro yourself (15 minutes)

Don't let AI write your intro. The intro is where readers decide whether to keep going. It needs to be:

Examples that work: "I tried to switch to a new code editor last week and gave up after 3 hours. The new ChatGPT pricing this week reminded me why switching costs are so brutal, and why everyone's mad about it." That intro takes 15 minutes to write right and is the highest-leverage paragraph in the whole issue.

• 2-3 sentences max
• Tied to something happening in your life or the wider world this week
• Slightly off-topic, then snapping back

Step 5: Generate the header image in Midjourney (10 minutes)

Open Midjourney. Use one --sref code that you reuse every week, this is what makes your newsletter visually identifiable in an inbox. (Generate 5-10 candidates the first week, pick your favorite, save its style reference.)

Prompt: [short visual metaphor for the issue's main story] --ar 3:1 --style raw --sref <your-code>

3:1 aspect ratio crops cleanly into Beehiiv / Substack / ConvertKit headers without the platform clipping it. Generate, pick, upload.

Step 6: Final pass and send (30 minutes)

Read the entire issue out loud. You will catch:

Subject lines are the second-highest-leverage part of the issue, after the intro. Use 4-7 words, ideally with a specific number or name. "OpenAI just shipped a model that codes like a senior engineer" beats "AI weekly: model releases and other news" by 30 percentage points on open rate.

• Sentences that sound like Claude (rewrite them)
• Repeated words across sections (Claude loves "moreover", search and destroy)
• A boring subject line (rewrite it three times, A/B test if your platform supports it)

What this looks like in practice

A weekly 3-hour Sunday ritual. Not nothing, but maintainable for years.

Mistakes that kill an AI newsletter

• Generic AI voice. No Project, no voice files, no by-hand edits. Reads like every other newsletter, gets unsubscribed at the same rate.
• Too many stories. Eight things you mostly didn't read about beats five things you actually have a take on, every time. Five is the cap.
• No image consistency. Different visual style every week and your newsletter never becomes recognizable in an inbox.
• Unedited intros. AI-written intros are the fastest way to feel inhuman. The intro is non-negotiably you.

Where to go next

After 10 issues, you'll know if newsletter writing is your thing. If it is, the upgrades are:

The compounding effect is real. By issue 50 you have a research workflow that's 4x faster than when you started, and a body of work that proves you have taste. Both are worth more than the newsletter itself.

• A swipe file. Save subject lines that worked (yours and competitors'). Reference before every send.
• Subscriber surveys. Every 25 issues, ask 3 questions. The answers will reshape your taste in productive ways.
• A second writer. AI plus you plus one human guest contributor every 4 weeks adds variety without breaking the rhythm.