Stack IA Complète de l'Auteur de Newsletter — Kit Substack / Beehiiv

What's in this pack (six layers: idea / research / draft / edit / image / send)

This is the stack a serious solo newsletter writer assembles once the weekly send stops being a hobby and starts being a job. Ten picks chosen so that every layer of the pipeline has one good default, no decisions to make on Tuesday morning.

• Idea layer (2 picks): Folo turns the entire web's RSS into one inbox you skim once a day. Fabric's 100+ prompt patterns gives you a hook-writer, a summariser, an angle-finder, an analogy-maker — all one shell command away.
• Research layer (2 picks): STORM does the deep-research multi-source synthesis when you commit to a topic. Jina Reader is the lighter daily tool — paste any URL, get clean markdown an LLM can actually read.
• Draft layer (2 picks): Content Marketer agent turns the brief into an outline plus draft. Zettlr is the markdown editor you actually finish in (footnotes, citations, distraction-free).
• Edit layer (1 pick): Vale catches the prose tells — passive voice, weasel words, your personal banned-word list. The thing that stops your newsletter sounding like every other AI-assisted newsletter.
• Image layer (1 pick): Together AI Image Generation produces the cover/hero image in 60 seconds at indie pricing. One per send, no Canva tab.
• Send layer (2 picks): Ghost is the self-hosted publishing + newsletter platform that gives you the subscriber list and the public web archive in one place. Resend Audiences is the API-first option if you want to drive sends from your own scripts or agents.

Install in this order (research → draft → edit → image → schedule → grow)

1. Folo — set this up first. The compounding asset of a newsletter writer is taste, and taste is built by reading. Folo's AI-powered RSS reader pulls every blog, Substack, and news site into one queue; you skim, you star, you flag-for-newsletter. Five minutes a day, three months in, your idea backlog is bottomless.
2. Fabric — install second. 100+ prompt patterns means you stop writing prompts at 9pm on a Monday. extract_wisdom, write_hook, find_logical_fallacies, create_summary — every common writer move is a one-liner. Wires cleanly into a shell pipeline with the other picks.
3. Jina Reader — install third. The day-to-day research tool. Paste any URL → clean markdown an LLM can read without HTML noise. Combine with Folo: read flag → Jina Reader extract → drop into your notes for the draft.
4. STORM (by Stanford) — install fourth. When a topic deserves a real essay rather than a quick take, STORM does the multi-source research and produces a structured draft you then rewrite in your voice. Don't ship STORM output verbatim; use it to short-circuit the "open 30 tabs" phase.
5. Claude Code Agent: Content Marketer — install fifth. The agent that turns brief + research into outline + first draft, with explicit hook variants. Pair with a style.md so the draft already speaks in your cadence.
6. Zettlr — install sixth. Where you actually finish. Markdown-first, footnotes, citation manager hooks, distraction-free mode. Substack and Beehiiv both import markdown cleanly; Ghost takes markdown natively.
7. Vale — install seventh. Prose linter you can extend with your own banned-word list. Catch "dive into," "in this issue we'll explore," "the world of X" — the AI tells that make readers unsubscribe.
8. Together AI Image Generation — install eighth. Cover image, generated in 60 seconds. One per send. Stop opening Canva.
9. Ghost — install ninth. The platform layer. CMS + newsletter + member payments in one self-hostable Node app. The bet: own the list, own the archive, never get rugged.
10. Resend Audiences — install tenth. The API-first send infrastructure for when you want an agent to schedule sends, segment by cohort, or run drip sequences without a UI in the loop.

How they fit together (ASCII pipeline)

  ┌─ Folo (RSS inbox) ─┐    ┌─ Fabric (prompt patterns) ─┐
  │ daily 5-min skim   │    │ hook / angle / summary       │
  └─────────┬──────────┘    └──────────────┬───────────────┘
            │                              │
            ▼                              ▼
        starred items ──────► idea queue (one doc)
                                    │
                                    ▼
                ┌────────── topic chosen ──────────┐
                │                                  │
                ▼                                  ▼
         Jina Reader                      STORM (deep research)
         (single URL → md)                (multi-source synth)
                │                                  │
                └────────────┬─────────────────────┘
                             ▼
          Claude Code Agent: Content Marketer
          (brief + research → outline + draft + hook variants)
                             │
                             ▼
                       Zettlr (write)
                             │
                             ▼
                       Vale (prose lint)
                             │
                             ▼
           Together AI Image Gen (one cover)
                             │
                             ▼
             ┌─── Ghost ───┐    or    ┌── Resend Audiences ──┐
             │ post + send  │          │ API-driven send      │
             │ + archive    │          │ + segments + drips   │
             └──────┬───────┘          └──────────┬───────────┘
                    ▼                              ▼
                subscribers read · open · share · forward
                    │
                    ▼
             back into Folo as growth signal (referrer, mentions)

The critical joins: Folo + Fabric (raw signal becomes idea), Jina Reader + STORM (research at two depths), and Vale before send (catch the AI tells before they ship).

Tradeoffs you'll hit

• AI draft consistency vs personal voice — A Content Marketer agent gives you a draft every time. The draft is competent. It is not you. The lever that closes the gap is a maintained style.md (banned phrases, your sentence rhythm, the three things you always argue) fed to the agent as system context. Update it monthly. If you skip this, your last six issues sound identical to six other AI-assisted newsletters and your open rate slides.
• Substack vs Beehiiv vs Ghost vs ConvertKit — Substack: fastest start, they own discovery and Notes, but the list is hostage and the revenue cut is real. Beehiiv: better analytics, recommendations network, no revenue cut on subs but ad-tier monetization is the pitch. Ghost: you own everything but you run the server (or pay Ghost Pro $9-25/mo). ConvertKit (now Kit): strongest creator-commerce tooling but weaker reader-facing web archive. Rule: if you'd cry losing the list, Ghost; if you want growth-by-platform, Substack or Beehiiv; if you sell digital products to the list, Kit.
• Paid newsletter vs free — Paid only makes sense once you have a free audience that already trusts you. Most newsletters that launch paid on day one die. Pattern: 18+ months free, 5K+ engaged subscribers, a clear premium-content thesis (deep research, private community, course access). Then paid. Switching free → paid is easier than switching paid → free.
• Self-host (Ghost / Listmonk) vs hosted (Substack / Beehiiv) — Self-host costs ~$5/mo + your time when something breaks. Hosted costs 0-10% of revenue plus deal-change risk. For a one-person newsletter still finding its voice, hosted is fine. For a newsletter that's part of your business, self-host the moment you can afford the ops.
• Stock photos vs AI-generated cover images — AI cover images are good enough at 60 seconds. They start to look generic by the 20th issue. Mitigate by maintaining a visual style guide (palette, composition rule, recurring motif) and feed it to Together AI as part of every prompt. Otherwise rotate in real photography or illustration every 4th issue to break the AI-template feel.

Common pitfalls

• The AI-newsletter voice. Symptoms: every issue opens with "Have you ever wondered," closes with "Until next week, friends," and uses "dive into," "the world of X," "in this issue we'll explore." Fix: put those phrases in your Vale config as errors. Force yourself to rewrite the opening line of every draft by hand, without the agent. Cadence asymmetry is what makes readers feel a human is on the other end.
• No cohort segmentation. All 3K subscribers get every send. The 200 hyper-engaged ones get the same email as the 200 who haven't opened in 6 months. You burn the engaged group and the dormant group never reactivates. Even a two-cohort split (engaged-last-30-days vs everyone) lets you write a re-engagement send without spamming people who don't need one.
• Substack lock-in. You spend three years building 10K subscribers on Substack. You export the list (Substack does let you). You import to Ghost. Open rate drops because you've lost the Substack-network discovery and your domain reputation is fresh. The lock-in isn't the export button; it's the distribution. Plan an exit while you're on Substack: cross-publish to a personal domain via custom domain, build a SEO archive elsewhere, accustom the list to your real URL.
• Cover images that scream AI. Smooth gradient, vaguely human face with one extra finger, painterly lighting on a person at a laptop. Readers spot it now. Fix: pick a specific visual identity (e.g., flat editorial illustration, 1990s magazine collage, single typographic poster) and constrain Together AI to that grammar. Stop using the default "hero image of X" prompt.
• Shipping STORM output verbatim. STORM produces a structured, citation-rich draft. It also produces encyclopedia voice: balanced, hedged, neutral. Newsletter readers signed up for your hot take, not Wikipedia. Use STORM as research compression; rewrite the entire prose in your voice before sending.