Automate Your Weekly Newsletter Without Losing Your Voice
A realistic pipeline for using AI to handle the summarizing grind while you keep full ownership of opinion, order, and the words that make readers open next week's issue.
👁 115 reads
Most newsletter writers don’t quit because they run out of ideas. They quit because every week they face the same two-hour slog: open twelve tabs, write twelve summaries, stitch them into prose that sounds like a tired version of themselves. AI can take that slog off your plate without touching the things that actually make your newsletter worth reading.
The pipeline below splits the work honestly. AI does the mechanical part - condensing sources into draft summaries. You do the editorial part - choosing what matters, deciding the order, writing the intro, and putting your name on a subject line. It takes about thirty minutes once the habit is in place.
Collect all week into one note, not a dozen bookmarks
The pipeline starts on Monday, not on the day you send. Keep a single running note - in Notion, Obsidian, Apple Notes, anywhere you’ll actually use - and paste every link into it as you encounter it during the week. Add one raw sentence of your own reaction when you paste: ‘this contradicts what I said in issue 12,’ or ‘readers keep asking about this,’ or even just ‘interesting, unsure why yet.’
That sentence is not for the AI. It is for you, Friday afternoon, when you’ve forgotten why you saved something. It also becomes the seed of your actual opinion later.
By the end of the week you might have eight to twenty links. Not all of them will make it in. The act of culling - deciding which four or six actually belong together this week - is the first editorial decision that no model should make for you. Kill anything that doesn’t fit the issue’s implied thread, even if the piece was good.
Give the AI one job: summarize what’s already there
Once you’ve selected your links, the summarizing prompt is the only place AI enters the production chain. The instruction below is one you can copy and adapt:
Summarization instruction (copy and modify)
You are summarizing source material for a newsletter. For each link I provide, write a summary of exactly three sentences. Sentence one: what the piece says. Sentence two: the most useful or surprising specific detail from it. Sentence three: a question or implication it raises - do not answer it, just name it. Use plain language. Do not editorialize or express enthusiasm. Do not add anything that isn’t in the source. If you cannot access a URL, tell me plainly rather than guessing.
Paste this prompt, then paste your selected links - or, if the AI can’t browse, paste the key excerpts you’ve copied from each source. Keep the AI anchored to what you actually gave it. The moment it starts ‘adding context’ from its training data, you’ve lost the factual tether. The last line of the instruction (‘tell me plainly rather than guessing’) is doing real work.
What you get back is a stack of neutral summaries. They should feel slightly flat. That’s correct. The flatness is the blank canvas.
Write the intro yourself, every single time
The intro is the one non-negotiable manual task. Two to four paragraphs, written from scratch, no AI draft. This is where you say what you think about the week - what connects these pieces, why you’re annoyed or optimistic or confused, what changed your mind. Readers subscribe to a point of view. The intro is where it lives.
If you genuinely have nothing to say about the week’s material, that’s a signal the curation was weak, not a reason to skip the intro. Go back to your raw reaction notes and find the thread.
Your voice guide: five rules you write for yourself
A voice guide isn’t a style sheet for the AI to follow - it’s a checklist you run through before hitting send. Five rules is enough. More than that and you won’t use it. Below is a sample block you can rewrite in your own terms:
Sample voice guide (rewrite these in your own words)
- Sentences stay under twenty words unless a longer one is doing something on purpose.
- No hedging phrases like ‘it seems,’ ‘perhaps,’ or ‘one might argue’ - take the position or don’t take it.
- Every piece gets one specific detail, not a generality. Numbers, names, or direct quotes from the source.
- The last line of any section earns its place - if it could be cut without losing meaning, cut it.
- Read it aloud before sending. If you stumble, the sentence is wrong.
After you’ve edited the AI’s summaries and written your intro, run this checklist once. It takes four minutes. It catches the places where editing fatigue made you accept a sentence that isn’t yours.
Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →Order, the subject line, and the final edit
Sequence is an opinion. The first item in your newsletter tells readers what you think is most important this week. The last item is what you want them walking away thinking about. Middle items support or complicate the framing. If you let the order be arbitrary, the issue feels arbitrary.
The subject line is the last thing you write, not the first. Write it after the issue is assembled because only then do you know what the issue is actually about. Keep it specific enough that a reader who opened five past issues would recognize it as yours.
Before sending, read the whole issue once more as a reader, not an editor. The question isn’t ‘is this correct?’ - it’s ‘would I keep reading if I hadn’t written this?’ If a section feels like filler, it is. Cut it and nobody will notice the gap.
The pipeline works because it protects the right things. Summarizing is slow, error-prone, and thankless when done manually - hand it to the AI with a tight brief. Opinion, sequence, and the subject line are fast when you’re not also exhausted from summarizing - keep them. That division, held consistently, is what lets you publish reliably without the newsletter starting to sound like everyone else’s newsletter.
More on this from QUANTSLANT
guides·How to Keep a Character Consistent Across AI-Generated ImagesPrompt drift, mutating faces, and shifting wardrobes are solvable problems - if you build a repeatable workflow before you start generating.