Personal Newsletter With AI Assistance

A personal newsletter with AI assistance should still sound like a specific person who read specific things and changed their mind in public. AI is there to cluster notes, challenge a weak argument, and catch a missing definition — not to become the author of record. When the model is the author, readers churn for a reason: they subscribed to you.

This post covers a production workflow for a one-person newsletter: capture, claims, draft, edit, send, archive. It places Cadensend accurately. Cadensend is an MIT-licensed, self-hosted email curriculum engine with no hosted signup (GitHub). It plans sourced lesson series and, in the MVP, delivers to your verified address. It is not Ghost, not ConvertKit, not Substack, and not a hosted newsletter product. Use it to build the course you wish your newsletter were. Use an ESP or CMS with consent features to reach other people.

Key Takeaways:

• Subscribe-to-a-person requires a claims table and a voice pass the model is not allowed to take.
• Cadensend is for private, sourced curricula you host; list tools are for opt-in audiences.
• Batch research into extracts; do not paste Perplexity essays into issue one.
• Schedule writing holds; AI will not create time.
• Never put subscriber PII or off-the-record mail into a consumer model to “find themes.”

Table of Contents:

• What a Personal Newsletter Is For
• The Weekly Machine
• Where Cadensend Fits
• Comparison: Newsletter Stack Options
• AI Roles With Hard Bans
• Voice, Trust, and Legal Floor
• From Private Curriculum to Public Issue
• Worked Example: One Week of a One-Person List
• Frequently Asked Questions
• Conclusion

What a Personal Newsletter Is For

Short answer: A cadence of owned judgments, not a recap engine.

Formats that work:

• “I tried X in production; here is the tradeoff I would repeat.”
• “I was wrong last month; here is the correction with the source.”
• “One procedure you can run this afternoon.”

Formats that die:

• Ten links with model blurbs (the “linkblog of a bot”).
• A recap of HN that a thousand other newsletters also recapped, 20 minutes later.
• A “thought leadership” essay with no artifact.

If you actually want a course, admit it and sequence it. That is closer to Cadensend than to a weekly vibe dump. If you want a wiki, stay in Notion: Notion AI vs dedicated learning tools. If you want launch-and-teach for a product, side project marketing emails that teach — still separate the ESP.

HinterBuild’s own blog is a public canonical, not a newsletter CMS. Different surface, same honesty rule: about.

The Weekly Machine

Short answer: Capture all week; decide once; draft once; send once.

Capture. URLs, quotes ≤25 words, one-line opinions. Perplexity research without copying: citations, not pastes. Batch process a reading list with AI on Friday morning if the pile exploded.

Decide. One thesis. If you have three, you have next week. Claims table: ship / needs source / kill.

Draft. You. Model may turn the table into an outline. AI for writing technical blog posts if the issue is technical canonical-length; keep newsletter issues shorter unless the promise is “deep.”

Edit. Grammarly vs Claude for editing. Voice pass last, out loud. Write better emails with AI for subject/preheader — still your promise, not “you won’t believe.”

Send. ESP with unsubscribe. Human click. Idempotent provider calls so a retry is safe: idempotency in distributed systems.

Time. Calendar AI scheduling for deep work. A newsletter without a hold becomes a guilty Google Doc.

Interview-adjacent writers. If the newsletter is also how you clarify stories for hiring, keep drills separate: AI for interview prep. Do not send your STAR stories as if they were reported journalism.

Where Cadensend Fits

Short answer: Cadensend is the compiler for a sourced series you receive as the first reader.

Use Cadensend when:

• You can name a learning goal with exclusions and an audience level (even if the audience is future-you).
• Sources will be files and URLs you ingest, not “whatever the model remembers.”
• You want coverage validation, locked issues, structural citations, timezone-aware exact-once delivery.

Do not use Cadensend when:

• You need to mail 4,000 subscribers tomorrow.
• You need a web magazine with members (Ghost).
• You need automations, tagging, commerce (ConvertKit-class).
• You need the model to browse autonomously. It will not; that is intentional.

MVP: single verified recipient. Roadmap later mentions newsletter-creator features (opt-in, suppression, archives) only after those controls are real. Until then, the honest architecture is:

1. Cadensend → your inbox (quality bar).
2. You copy or rewrite into Ghost/ESP (audience bar).

Self-host, MIT, no Cadensend SaaS signup. Page: Cadensend. Cluster: products. Engineering help: backend API engineering, contact.

Gmail is the mailbox, not the authoring environment of record. AI in Gmail without leaking data.

Comparison: Newsletter Stack Options

If a roundup article lists Cadensend next to ConvertKit as “open source ConvertKit,” it is wrong. Correct it with this table.

AI Roles With Hard Bans

Allowed

• Cluster this week’s claim cards into possible theses (you pick).
• List questions a skeptical subscriber would ask (you answer from evidence).
• Flag jargon on first use.
• Compress a long canonical post into a newsletter-sized issue without new facts.
• Quiz you on last month’s issue so you remember what you taught.

Banned

• Invent a “study.”
• Invent a subscriber quote.
• Soften a retraction.
• Impute motives to named people.
• “Make it more viral.”
• Read the mailbox to “see what people care about” unless you have a DPA and a pipeline: PII.

Hallucinations in a personal newsletter destroy the only asset you have. LLM hallucination causes and fixes. Untrusted PDFs in the research drop: prompt injection.

Voice, Trust, and Legal Floor

Voice is specific nouns from your week: the queue name, the failing test, the person (with permission). Models prefer abstract nouns. Delete them.

Trust is corrections. When you blow a claim, the next issue leads with it. AI will try to bury the lede to save tone. Do not allow that.

Legal floor: unsubscribe, identity of sender, don’t buy lists, don’t scrape. Educational tone does not grant extra rights. HinterBuild is not your counsel.

Deliverability is an engineering problem (alignment, bounce handling, complaint loops) — ESP territory, not Cadensend MVP territory.

From Private Curriculum to Public Issue

Many good newsletters are public slices of a private syllabus.

1. Run a 6-issue Cadensend series on a topic you actually need (e.g. citation rendering, calendar holds, Gmail redaction).
2. Note where you argued with the draft. That argument is the newsletter.
3. Publish one issue a week to the list, rewritten for strangers (more context, fewer in-jokes unless in-jokes are the product).
4. Link the long form on hinterbuild.com or your own site when it deserves to rank.

That path respects both products. Cadensend taught you. The ESP spoke to them.

Keep the calendar honest or the private series dies and the public one becomes AI sludge. Calendar AI scheduling for deep work.

Worked Example: One Week of a One-Person List

Short answer: Private lesson Tuesday, public issue Thursday, no model in the byline.

Sunday capture. Twelve URLs, three quotes, two opinions (“defrag tools shattered my Wednesday,” “citation IDs beat model footnotes”). Perplexity used only to find the AWS page you had lost: Perplexity research without copying. Batch the rest Monday: batch process a reading list with AI. Notion is the shoebox: Notion AI vs dedicated learning tools.

Monday decide. Thesis: “Your newsletter should not be the first time you see your own argument.” Claims: (1) you shipped a duplicate email during a retry (your logs); (2) keyed sends fix it (idempotency in distributed systems); (3) Cadensend is not your ESP. Kill: “studies show newsletters are back.”

Tuesday Cadensend. You already have a three-issue private series on exact-once delivery, sourced from your own design notes and public specs. Issue 2 arrives at 07:30. You read it in a DEEP hold. You argue with one paragraph. That argument is Thursday’s newsletter. Cadensend — MIT, self-host, verified recipient = you, GitHub. Not Ghost. Not ConvertKit.

Wednesday write. Human draft, 700 words. Claude: list unsourced claims. It finds “everyone hates Substack.” Deleted. Grammarly: two hyphens. Grammarly vs Claude for editing. Subject: write better emails with AI, then you rewrite it because the model went clickbait. No subscriber CSV in the prompt. AI in Gmail without leaking data. No incident thread. If you later want a canonical, AI for writing technical blog posts. If you are also interviewing, do not paste this as a STAR: AI for interview prep. If this were a launch sequence, it would be side project marketing emails that teach with an ESP.

Thursday send. ConvertKit (or Ghost members, or any list tool with unsub). Human send. Retry-safe. Calendar already held 09:00–11:00: calendar AI scheduling for deep work.

Friday. One correction if a reader caught a sloppy sentence. Lead with it next week. The model will suggest burying it.

Stack reminder. Cadensend compiled. ESP delivered to others. HinterBuild’s products page keeps that split. Need queues? Backend API engineering. Contact. About. Hallucination discipline: LLM hallucination causes and fixes.

If Thursday’s draft still sounds like a model, cut the first paragraph and the last, and keep only the proof: the duplicate send, the key, the timezone you actually used. Readers subscribed for that object, not for a smoother tone. You can run one more pass with the prompt “do not add facts; list sentences a close friend would not recognize as mine.” Delete those sentences. Then send. Skipping a week is allowed; shipping a generic recap is how the list dies.

Frequently Asked Questions

Can I run my public newsletter on Cadensend today?

No. Cadensend’s MVP delivers to your verified address. There is no hosted Cadensend signup and no list product yet. Use an ESP or a CMS with membership for other people.

Is Cadensend like Ghost or ConvertKit?

No. Ghost is a publishing CMS. ConvertKit is a creator marketing platform. Cadensend is a self-hosted email curriculum engine that turns a learning goal into sourced, scheduled issues.

How much AI is too much?

If a close colleague cannot tell you wrote it, too much. Keep the proof paragraph 100% yours.

Should I summarize the whole internet each week?

No. Pick one judgment. Point to primaries. Batch the rest of the pile into extracts for you, not for the list.

What about paid newsletters?

Same ethics, higher legal and support load. Cadensend Update 2 on the public roadmap mentions billing in the context of future newsletter-creator features — not a promise that you should take payments through Cadensend today.

Can HinterBuild build me a custom newsletter backend?

Possibly, as backend API engineering. That is not the same as adding hosted Cadensend. Contact.

How do I pick a cadence I will keep?

Start fortnightly if weekly dies. The archive of six honest issues beats a graveyard of twelve AI recaps. Tie the send to an existing hold, not to a burst of motivation after a viral tweet. If you cannot name next week’s thesis on Sunday, skip the send rather than generating one.

What belongs in an archive vs an email?

Archive (Ghost, your blog, git): the version you will defend in a year. Email: the weekly judgment plus a link. Do not let the ESP be the only copy. Do not let Cadensend be the public archive in MVP — there is no public list, and the product is not a CMS.

Can I use subscriber replies as training data for a model?

Not by default. Replies are correspondence. You need a policy, a redaction step, and a reason. “The model writes better if it reads my fans” is not a reason. If you quote a reader, ask.

Conclusion

• A personal newsletter is judgments on a cadence, assisted — not authored — by models.
• Cadensend is the private curriculum compiler (MIT, self-host, not ConvertKit/Ghost).
• ESPs and CMSs remain how you reach opt-in readers.
• Freeze claims, protect subscriber data, write on a calendar hold.
• Promote lived-through lessons, not chat exports.

Related: write better emails with AI, side project marketing emails that teach, LLM hallucination causes and fixes, idempotency in distributed systems. Products, about, Abdul Sami on LinkedIn.