Build a scheduled LLM newsletter pipeline that crawls sources, scores items and drafts each issue
An open-source TypeScript kit that crawls source sites, tags and scores articles with an LLM, then drafts a newsletter issue; the author reports about $0.20 to $1 in API cost per issue.
Evidence: The source shows its work: steps, screenshots, code or data you can inspect.
The business problem
A one-person newsletter that curates news from many sources needs hours of reading, sorting and writing per issue. The author wanted an issue to assemble itself for pennies, with a human only checking the result.
What was tried
The kit runs as a pipeline you host yourself. First it fetches list pages from your chosen sources, skips URLs it has stored, then fetches and parses each detail page and saves the article. Next an LLM tags each article, analyses images in parallel and gives each item an importance score from 1 to 10. A publication check decides whether there is enough to send; if so the model writes a title and Markdown body, which is rendered to HTML with inlined CSS and saved. Optional steps send a preview email, extract deadlines or events from sources, and rewrite a title that repeats a recent one. Every stage is a swappable provider, so you can plug in any model supported by the AI SDK, and you supply your own scraping, storage, scheduling and subscriber delivery. The author extracted it from a live niche newsletter and suggests rule-based parsers rather than LLM scraping for production.
What was reported (positive)
The README says LLM API cost is roughly $0.2 to $1 per issue with optimised model choices, and reports a 15% click-through rate and near-zero maintenance for the author's own newsletter. These are self-reported and tied to that one service, not to the kit in general.
Limitations
Cost and click-through figures are self-reported from one small newsletter, with no models, article counts or token volumes given, so they cannot be reproduced from the page. Needs a recent Node.js release and your own scraper, scheduler, database and email sending, so setup is heavier than no-code tools. Deduplication is by exact URL only, retries can repeat LLM calls, retry settings do not cap total spend, and source-grounding checks are prompt-based, so a human still has to verify facts before sending. The author promotes their own newsletter service.
What you need
Comfort with TypeScript and Node.js, an LLM API key, sources you are allowed to crawl, and your own storage, scheduling and email delivery.
The quick start uses an OpenAI model through the AI SDK; the repo's playground also supports Anthropic, Google and TogetherAI.
Sources
- GitHub, kimhongyeon (README) ↗ Code repository, publication date unknown
Source published: unknown. Last reviewed here: October 17, 2026. Spot a mistake? Tell us.
Tools in this workflow
- LLM Newsletter Kit: Pipeline framework for crawling, scoring and drafting issuesTry LLM Newsletter Kit
- AI SDK: Model interface the kit callsTry AI SDK
Related workflows
Pick an AI model and post-editing setup per content type for Chinese localization, using a blind 774-output benchmark
A localization benchmark found AI workflows beat human translators on four of six content types, but humans led on informational and SEO text, and model choice mattered more than post-editing.
Draft a training course outline with Claude Projects: your own voice documents, 16 design questions, then edit every step
A learning and development professional loads her voice, expertise and story documents into a Claude Project, gives it her outline structure and about 16 design questions, and edits the draft outline and workbook text it returns.
Build a small SaaS with Lovable, Supabase and Stripe using a living spec file and a habit of checking the database yourself
A founder builds with Lovable, keeps one living spec file the model reads before every feature, ships one feature at a time, and audits the database, logs and billing after each one because the tool reports success without checking.
Build a course with AI only where it helps: human outcome and outline, AI for examples, analogies, copy and voice
A course creator with over twenty years of experience describes keeping course design human and using AI for examples, analogies, marketing copy and voice matching.