Newsletter Autopilot
A finished issue draft every send day, researched all week, every claim cited.
It researches and drafts. You edit the draft and send it from your own email tool; it never sends, schedules or publishes.
The eval-harness fight is really a fight about who owns the definition of “works”. Three takes worth your time, one worth your money.
Browser agents left beta quietly. The latency benchmark explains why nobody announced loudly.
A 4-person team automated their whole CI triage. The case study includes the prompts.
The newsletter dies because the research eats the week. This agent sweeps X and the web for your topics every night, keeps only what is genuinely new, and on send day assembles a complete issue draft in your voice: sections, links, subject line options. You edit it and hit send. Sending stays yours.
One paste into your agent.
Your first issue draft, minutes later.
Install superagnt's Newsletter Autopilot skill for me. It's the newsletter-autopilot plugin from the github.com/superagnt/plugins marketplace, plus the superagnt MCP server (on Claude it's a connector I add in my browser; elsewhere I sign in through my browser). The setup guide for each kind of agent is at https://www.superagnt.com/install?skill=newsletter-autopilot» installs the newsletter-autopilot skill (read every file first, or the raw SKILL.md) and enables data:x, data:web, database
One night's research and the draft written from it, Nov 3, 2026
An SEO newsletter's week on a superagnt test workspace, 18:46 to 18:47 UTC: three X searches and three web news searches, then the four pages worth a sentence read in full. 65 finds came back, 5 were kept. No past issue was given, so the draft is in a plain default voice.
“Run the Newsletter Autopilot skill for my SEO newsletter and draft this week's search news section.”
- 01X: this week's top posts on AI Overviews and SEO3.5 s wall time→ call
data_x_discover{ "action": "search", "query": "\"ai overviews\" seo", "section": "top", "min_likes": 20, "start_date": "2026-10-27", "limit": 20, "response_format": "raw" }← result{ "results": [ { "tweet_id": "2117602399947026591", "creation_date": "Tue Nov 03 12:58:49 +0000 2026", "text": "Most people still think SEO is just about ranking on Google.\n\nThat’s outdated. 👇\n\nModern SEO is evolving into 5 powerful systems:…", "favorite_count": 45, "reply_count": 6, "retweet": false, "user": { "username": "ai_with_orbit" } }, { "tweet_id": "2117297378739052728", "creation_date": "Mon Nov 02 16:46:47 +0000 2026", "text": "Google revealed how sites *must* be named if you want to show in traditional and AI search.\n\nThis comes straight from Google's John Mueller.…", "favorite_count": 84, "reply_count": 26, "retweet": false, "user": { "username": "alexgroberman" } } ] }» Trimmed: 19 more posts of 21, the next-page token, and the fields this step doesn't read (views, media, the author's profile)
- 02Web: this week's news on Search Console1.8 s wall time→ call
data_web_search{ "query": "Google Search Console", "sources": [ "news" ], "tbs": "qdr:w", "limit": 8 }← result{ "success": true, "data": { "news": [ { "title": "What Search Console Can Still Tell You When Indexing Data Lags", "url": "https://www.searchenginejournal.com/what-search-console-can-still-tell-you-when-indexing-data-lags/595016/", "date": "4 days ago" }, { "title": "Gary Illyes Implies Click Data Not Coming To Google Search Console AI Reporting", "url": "https://www.seroundtable.com/click-data-search-console-ai-reporting-42437.html", "date": "1 day ago" } ] } }» Trimmed: 6 more results of 8, their snippets and images, and the payload's other fields
- 03Read the keeper that earns the lead item1.3 s wall time→ call
data_web_scrape{ "url": "https://www.seroundtable.com/google-ai-overviews-mode-ctr-42421.html", "formats": [ "markdown" ], "onlyMainContent": true }← result{ "success": true, "data": { "markdown": "…Everyone has been asking, what is the click-through rate you get from Google Search AI features such as AI Overviews and AI Mode. Well, a new small study says the click-through rate is about 0.9%. The study also showed that when you opt out of showing in Google's Search generative AI, it also leads to a drop in clicks in the main web results.…" } }» Trimmed: the page's navigation before this paragraph, the rest of the article (33,770 characters in all), and the payload's other fields
The five keepers went into issue_material with a short summary each, which is also the skill's repeat ledger: a URL already there is skipped on the next night. Claude (Opus 5.5) in Claude Code wrote the draft below from those five, by the skill's send-day step.
| Search | Source | Found |
|---|---|---|
"core update" google | X | 0 (not charged) |
"ai overviews" seo | X | 21 posts |
"search console" | X | 20 posts |
Google search update | Web news | 8 articles |
Google AI Overviews | Web news | 8 articles |
Google Search Console | Web news | 8 articles |
The night's searches. The X searches read the top posts since Oct 27 with 20 or more likes; the web searches read the past week's news.
The draft section, with two subject lines:
- Subject: AI Overviews: about 0.9% click-through, in one site's test
- Subject: That Search Console click spike may be a bug
AI Overviews and AI Mode turned about 0.9% of their impressions into clicks, in one site's test. Cyrus Shepard used Google's opt-out controls to pull one site out of AI Overviews and AI Mode, then compared before and after: AI impressions came to an estimated "~0.9% CTR", and clicks from the regular results fell after the opt-out too. One site and a small sample, and Barry Schwartz says he hasn't seen anyone else publish data like it (Search Engine Roundtable, Oct 30).
Don't wait on clicks in Search Console's AI report. Gary Illyes says the team wants to add "data that makes sense" and has nothing to announce, which reads as no click or query data soon (Search Engine Roundtable, Nov 2).
A click spike this week? Check it before you report it. Joy Hawkins and Glenn Gabe both posted sudden jumps in Search Console clicks. On some properties the API shows the spike and the web report doesn't, and Search Engine Land has asked Google whether it's a bug (Search Engine Land).
Your PageSpeed score can rise with nothing changed. PageSpeed Insights now runs Lighthouse 13.6.0 (released Oct 27), which changes how it simulates First and Largest Contentful Paint. Barry Pollard says scores generally go up, most on JavaScript-heavy sites (Search Engine Journal).
A night of this research costs 3.44 credits
| Call | Count | Credits each | Credits |
|---|---|---|---|
| X search | 2 charged, 1 empty and free | 1 | 2.00 |
| Web news search | 3 | 0.288 | 0.864 |
| Page read | 4 | 0.144 | 0.576 |
At this night's mix, seven nights come to about 24 credits before the send-day draft. That is an estimate from one night, not a measured week: a busy news week reads more pages, and pages read with heavier settings cost more.
»Charges read from the workspace's usage log for the run, Nov 3, 2026, 18:46 to 18:47 UTC.
The hand-off to your email tool
- The draft is markdown with every link inline: the open, each section written out, and two or three subject lines.
- Move it into whichever tool you send from, Beehiiv, Substack, Kit, Mailchimp or Gmail. The skill never connects to your email tool.
- Mark what you used. The rows it drew on are stamped with the issue, so next week starts from what's new.
Three steps to your first issue draft.
- 1
Name the topics and the send day
Tell it what you cover, when you send, and hand it one past issue so it learns your voice and section shape. Saved once, read every run.
agnt_db_insert
the letter, learnedexampleai agentsgtm toolingsends fridayvoice + section shape read from one past issue
- 2
It researches all week
Every night it sweeps X and the web per topic, skips what it has already filed, and scrapes the pieces worth keeping. By send day the raw material is sitting in your table, sourced.
data_x_search_get_search_search · data_web_search · data_web_scrape · agnt_db_upsert
researched all weekexample03:00 · thu sweep» kept · benchmark thread, summarized
» kept · pricing-change post, scraped
» skipped · 38 already in the pile
- 3
Send day: the draft is waiting
A complete issue draft in your section shape: the takes, the links, every claim cited, plus subject line options. You edit it, then you hit send.
agnt_db_select · agnt_db_upsert
send day · the draft is waitingexampleIssue 47: the week agents started buying things
3 sections · 11 claims, 11 citations · 2 subject options
20 min of editing · you hit send
»After the first good run, the skill offers the always-on version: a small hosted runner on a schedule, so the work keeps happening with your laptop closed. You confirm the cadence before anything is scheduled.
Read every file before you install it.
SKILL.md tells your agent to learn your letter from one past issue, keep 1 to 3 X searches and 1 or 2 web searches per topic, file only finds that would earn a sentence, and on send day assemble the whole issue with every claim linked. Its last rule: it drafts, and it never sends email, posts or touches your email tool. Its first two headings are setup for the agent.
Plain markdown your agent reads while it runs. SKILL.md is the procedure. This is the plugin exactly as it installs from github.com/superagnt/plugins: 4 files, 7.7 KB.
Common questions
01Will subscribers know it is AI-assisted?
Only if it reads that way. The skill learns your voice from one past issue, summarizes in your words instead of copying, and links every claim. You edit before you send, and whether you mention AI help is your call.
02How much time does it actually save?
We haven't timed the editing yet, so we don't claim a number. What we measured: one night's research, six searches and four pages read, took 20 seconds of tool time and 3.44 credits. You still read and edit the draft.
03What email service providers work best?
Any of them. The skill never touches your email tool: it hands you the draft as markdown with every link in place, and you move it into the one you send from, whether that's Beehiiv, Substack, Kit, Mailchimp or Gmail.
04Can I use this for a paid newsletter?
Yes. It drafts and you own the issue. It quotes at most one short line per source and links every one. The skill is free; the research runs on superagnt credits, and plans are $49, $149 or $399 a month.
05How does the AI ensure content quality?
With three rules it follows every run: no cite, no claim; a find earns its row only when it would earn a sentence (new information, a real number, a take worth reacting to); and anything already filed is skipped. You are the last check before send.
06Why does my AI newsletter not sound human?
Usually because it writes from a prompt with no voice to copy. This skill asks for one past issue first and records how you open, how many sections you run and how you frame a take, then drafts in that shape.
Already connected? Your agent can list every packaged skill: agnt_tools_search + agnt_guidance_search