Writing a newsletter with AI: Building a Curation Engine

There is a specific kind of friction that happens when you try to turn a pile of interesting links into a cohesive newsletter. It is like trying to organize a junk drawer by hand. You pick up an object, remember why you kept it, and then realize it does not fit with the other five things you just laid out on the table. In design, we talk about affordances, the qualities of an object that tell you how to use it. A well-crafted newsletter should have a high degree of legibility. It should tell the reader exactly why these specific ideas are grouped together.

Most people think writing a newsletter with AI means asking a chatbot to write five paragraphs about AI trends. That approach creates a high volume of noise but very little value. The result is often a flat, repetitive artifact that lacks the texture of a human perspective. If you want to maintain subscriber retention, you need to treat AI as a research assistant, not a ghostwriter. You want to automate the curation engine while reserving the final editorial voice for yourself.

What you will have at the end

By following this guide, you will build a system that automatically pulls content from your favorite sources, synthesizes the core arguments using an LLM, and prepares a draft in your workspace. You will move from a manual copy-paste workflow to a high-flow state where your only job is to add the 'so what' to each entry.

This system solves the problem of subscriber fatigue. Data shows that newsletters written entirely by AI often see a 12 percent drop in open rates over a six-month period. In contrast, hybrid newsletters, where AI handles the structural synthesis and humans handle the voice, maintain an average open rate of 38 percent, which is consistent with premium manual publications. You will also have a protocol to verify links and prevent hallucinations, ensuring your credibility remains intact.

Prerequisites

Before we start, ensure you have access to the following tools:

1. An automation platform: Make.com or Zapier to connect your data sources.
2. An LLM API: An API key for Claude 3.5 Sonnet or GPT-4o. Claude is generally preferred for editorial tasks because of its ability to follow complex style guides. You can find more on this in our comparison of ChatGPT vs Claude for marketing copy.
3. A destination workspace: Notion or Google Docs. Notion AI is particularly useful here for internal refinements.
4. RSS Feeds: A list of 5 to 10 high-quality RSS feeds from industry blogs or news sites.

Step 1: Connect live RSS feeds to your database

The first seam in our workflow is the connection between the web and your database. Manually checking sites is a flow state killer. We want these items to flow into a central 'holding tank' without any manual effort.

1. Create a new table in Notion or Airtable with columns for 'Title', 'URL', 'Source', 'Raw Text', and 'AI Summary'.
2. In Make.com, create a new scenario. Add the 'RSS' module and set it to 'Watch items'.
3. Add your first RSS feed URL. For example, use a primary source like the Official Anthropic News Feed.
4. Add a 'Text Parser' module to scrape the full content of the URL. This prevents the LLM from hallucinating based on just the title.
5. Connect this to your Notion module. Map the RSS fields to your table columns.

This creates a live feed of research. The cost of this automation is negligible compared to traditional freelance research. While a freelancer might charge $50 to $100 just to curate a list of links, your API and automation costs will likely hover around $0.15 per issue.

Step 2: Design a verification and synthesis prompt

Now we must address the biggest risk of writing a newsletter with AI: hallucinations. You cannot trust an LLM to accurately summarize a technical paper without a specific protocol. We will use a few-shot prompting technique, which involves giving the AI multiple examples of how you want it to think. This is similar to showing a junior designer a mood board before they start a project.

In your automation tool, add a module for your chosen LLM. Use the following system prompt structure to ensure structural synthesis:

Role: Senior Research Assistant
Task: Summarize the provided article for a professional newsletter.

Constraint 1: You must only use information provided in the 'Raw Text'. 
Constraint 2: If the text mentions a statistic, include the exact number.
Constraint 3: Verify the URL is active and matches the content.
Constraint 4: Use a neutral, analytical tone. Do not use marketing buzzwords.

Output Format:
- One sentence summarizing the core argument.
- Three bullet points of supporting evidence.
- One 'counter-point' or limitation mentioned in the text.

This prompt creates a consistent mental model for the AI. It focuses on the artifacts of the text, the facts and the figures, rather than trying to mimic your voice yet. You can find more tips on refining this kind of output in our guide on AI for UI copywriting.

Step 3: Manage the editorial seam

This is where most people fail. They take the AI output and hit 'send'. This creates a visible seam between the machine-generated summaries and the human reader. To fix this, you must apply your own editorial layer.

Open your draft in Notion or your preferred editor. Use Grammarly to check for clarity, but do not let it strip away your personality. Your job is to look at the AI-generated summary and ask: 'Why does my specific audience care about this today?'

One small thing people ignore is the impact of AI-generated text patterns on email deliverability. Modern spam filters, especially those used by major providers like Gmail and Outlook, are increasingly sensitive to 'low burstiness' text. This is text that has a very consistent sentence length and predictable word choices. To avoid the spam folder, vary your sentence lengths manually. Use a mix of short, punchy statements and longer, more complex observations. This increases the 'human' signature of your email.

Troubleshooting

If you find the AI is hallucinating facts, the issue is usually the 'temperature' setting in your API call. For curation, set the temperature to 0.2 or lower. This makes the model more deterministic and less likely to wander into creative fiction.

If your open rates are dropping, check your subject lines. AI tends to suggest titles like 'The Future of X: A Comprehensive Guide'. These are boring. They lack the affordance of curiosity. Write your own subject lines based on a specific, small observation you made while reading the curated links.

If you encounter technical errors in the automation, it is often due to a broken RSS feed or a site blocking your scraper. You can use tools like Claude Code to write a small custom script that handles these edge cases more gracefully than a standard no-code tool.

Next steps

Once your curation engine is running, you can start experimenting with visual artifacts. Use Stable Diffusion to generate unique header images for each section. This adds a layer of craft that stock photos cannot match.

To test your new system, run this simple experiment:

1. Pick one topic you usually cover manually.
2. Let the AI engine curate and summarize three articles on that topic.
3. Spend exactly 10 minutes adding your own 'Editorial Why' to those summaries.
4. Send it to a small segment of your list and compare the click-through rate to your previous issues.

You will likely find that the clarity of the summaries, combined with your unique perspective, creates a much stronger flow state for your readers. For more advanced workflows on using AI for complex tasks, see our staff engineer's guide to code refactoring.