Draft your weekly newsletter from the links you saved, in your own voice
Save links during the week. At the end of the week an agent opens each link, confirms it could read the page, summarises it, groups items into themes and drafts the issue in your voice. You check each summary against its link, edit, and send it yourself. The main reported risk is a confident summary of a page the agent never read, so the agent must say which links failed.
You get a draft issue with a short, sourced summary of each link you saved, grouped by theme, plus a list of links the agent could not read.
- Difficulty
- Beginner
- Time
- 30-60 minutes a week once set up
- Sources
- 6 · 6 independent
- Updated
- 2026-10-13
The prompt
Fill the blanks here, or leave them for your agent to ask. Pick the tab for your agent. The hand-off button gives any agent this page as markdown.
Draft this week's issue of my newsletter, [newsletter name], for [audience]. Links I saved this week: [paste links, or point to the file, bookmark folder or read-later list]. Three past issues, so you can match my voice: [paste or link]. Format: [intro of 2-4 sentences], then 3-5 themed sections, then [a short "also worth a look" list]. Step 1. Open every link. For each one, record: title, author, publish date, and whether you could read the full text (yes, partly, no). If a page is paywalled, blocked, or you only saw the URL, mark it "no". Do not summarise any page marked "no". STOP and show me this list before you write. Step 2. After I reply, for each readable link write a summary of 2-3 sentences with one concrete detail from the page. Keep numbers exactly as the page states them. If you quote, copy the words exactly and keep the quote under 20 words. Step 3. Group the items into themes. Drop items that do not fit the audience and tell me which ones you dropped and why. Step 4. Write the issue in my voice. Use short sentences. Do not add opinions I have not given you. Mark any place you want my own take with [MY TAKE: add your view here]. Rules: - Every item links to its source. - Never invent a title, author, book or statistic. If you are unsure, leave it out and tell me. - Do not send, schedule or publish anything.
Human checkpoints: the agent stops and asks you
- Review the list of links the agent could not read before it writes.
- Check every summary, quote and number against its source.
- Send or schedule the issue yourself in your email tool.
A newsletter goes to many people at once and cannot be unsent. A wrong summary or a made-up title reaches every reader. You read the draft against the sources and press send yourself.
What you need
- A list of saved linksdata
A read-later app, a bookmarks folder, a Notion table or a plain text file all work.
- Two or three past issuesdata
Writers who gave the model their own past writing report better voice.
- An agent that can open web pagesapp
One newsletter writer says Claude Desktop could not search until he added search and fetch tools.
- Your newsletter platformaccount
You paste or import the draft and send it yourself.
Steps
Save links in one place all week
Use one list so the agent finds every item. Add a short note to a link if you already know why it matters.
Check: The list has this week's links and nothing older.
Let the agent read and report
Paste the prompt. The agent opens each link and reports which pages it could read. Replace or drop links it could not read, or paste their text in yourself.
Check: Every link is marked yes, partly or no.
Get the themed draft
The agent summarises readable links, groups them and writes the issue in your format.
Check: Each item has a link, and the draft follows your usual structure.
Check every item against its source
Open each link and compare. Look hard at titles, author names, numbers and quotes. Large studies of AI assistants by the BBC and the EBU found altered quotes and sourcing errors in a large share of news answers.
Check: Every summary matches its page, and every quote is exact.
Add your own take
Fill each marked spot with your view. This is the part readers subscribe for.
Check: No placeholder text is left.
Send it yourself
Paste the draft into your newsletter tool, send a test to yourself, read it on your phone, then send or schedule it.
Check: The test email looks right and every link works.
What goes wrong
The agent summarises a page it never read, from the URL or from memory.
Fix: Make the agent report read status for each link before writing, and never summarise a "no". Test once with a broken link to confirm it admits the failure.
Read the lesson →The draft includes a book, title or fact that does not exist.
Fix: Allow only items from your list. In 2025 a newspaper printed an AI-made summer reading list with books that do not exist, as 404 Media reported.
Read the lesson →Paywalled or HTML-less pages get only a link, or a thin summary.
Fix: Paste the text of those pages yourself, or move them to the short "also worth a look" list.
The prompt drifts. Small changes give different output each week.
Fix: Keep the prompt in one versioned file. One writer says prompt changes behaved unpredictably until he added version control.
The voice is generic.
Fix: Include past issues as samples and leave marked spaces for your own opinion.
Evidence
Two or more public accounts from different people or sites describe the same job. We summarise each source in our own words. Open the original before you rely on it.
- Blog2025-07-11· ClaudeHow I created an AI-based newsletter workflow using Claude Desktop and MCP servers
The author uses Claude Desktop with search and fetch tools to find and read pages for his weekly issue. He reports about 36 prompt versions, and says the process takes about 45 minutes. Claude could not search until he added search tools, and prompt changes were unpredictable until he versioned them.
Luis Poveda · povedaaqui.substack.com - Blog2024-12-05· Custom agentBuilding an automated research newsletter
The author has Claude screen arXiv titles and abstracts against his interests and summarise the papers that pass, sent twice a week. Papers without an HTML version cannot be read in full and appear only as links. He still reads the interesting papers himself.
Wesley Pasfield · wesleypasfield.substack.com - Blog2026-09-29· Custom agentHow to automate your newsletter with an AI agent
The author has Claude draft issues in his voice from a topic queue, with a human preview and confirm step before sending. He says about 90% of first drafts are approved unchanged and the agent cannot fix a weak topic. The page promotes his own services.
Alejandro Rioja · alejandrorioja.com - Blog2025-07-17· Claude CodeHow I built an AI system that finds ideas I want to write about (comments)
The author uses Claude Code over a local copy of his newsletter archive to find ideas. Commenters say they moved much of their newsletter work into a Claude project and that the output is far better when it is grounded in their own past writing.
Wyndo · aimaker.substack.com - Blog2023-03-10· ChatGPTChatGPT couldn't access the internet, even though it looked like it could
Willison gave ChatGPT links to pages that did not exist, and it wrote a confident summary from the words in the URL. He later noted the post is out of date because ChatGPT gained browsing, but the test he suggests still applies.
Simon Willison · simonwillison.net - News2025-05-20Chicago Sun-Times prints AI-generated summer reading list with books that don't exist
404 Media reports a printed reading list, partly made with AI, that credited real authors with books that do not exist. The person who made it said they were embarrassed.
Jason Koebler · 404media.co
- EBU and BBC News Integrity in AI Assistants study (press release)checked 2026-10-13
- Research Live on the BBC study of AI news answerschecked 2026-10-13
- Schedule recurring tasks in Claude Coworkchecked 2026-10-13
Report how it went
Every report makes this page better for the next person and the next agent. Reports are reviewed before they count. Agents can report with the report_run tool.
Loading reports…