How I Manage 47 Newsletters a Day With One AI Prompt

I get 47 tech newsletters a day. I read 3 of them. The other 44 are read by an AI that scores each one against a 4-line prompt and only surfaces what is relevant to me. This is not a productivity hack. This is a Saturday afternoon project that took 200 lines of Python and a cron entry. The whole thing runs on the M720q in my living room.. Prompts Are Code Now: My Claude Opus 4.8 Playbook

The 47-newsletter trap

It started during the pandemic. Like everyone else, I was stuck at home and anxious about falling behind. The reflex was familiar: subscribe to everything. I told myself I was “staying current.” What I was actually doing was performing learning without absorbing anything.. Post 2: Setting Up Hermes Agent — Your AI Project Manager

By 2024, my morning routine was 45 minutes of triage. Open Substack. Open LinkedIn digest. Open the eight AI newsletters I had somehow accumulated. Open three dev newsletters from companies whose product I do not use. By the time I got to actual work, I had read 200 headlines and remembered three. The remaining 197 were a tax on my attention that compounded through the day.

The worst part was not the time. The worst part was the false sense of progress. I felt like I was learning because I was reading. I was not. I was consuming headlines at the rate my eyes could move and retaining at the rate my morning brain could focus on a topic I had not asked to read about.

The bottleneck in my learning was not access to information. It was the discipline to filter 200 headlines a day. I cannot do that. I have tried for three years. I will keep failing at it. The only way I found to stay current on more topics, with less effort, than I did when I was reading 47 by hand, was to let an AI filter them on criteria I wrote.

The prompt

The whole system pivots on a 4-line prompt. I run it on every article before I see the headline. Here it is, copy-paste ready:

Score this article from 0 to 10 against these criteria:
(1) is it relevant to a self-hosted homelab builder in Indonesia,
(2) does it ship working code or commands I can use this week,
(3) does it change how I would build or operate a system,
(4) is it from a source I trust?
Return JSON with score and 1-line reason.

That is the entire filter. There is no machine learning, no embedding, no vector database. Just an LLM call with a tight question. The question is not “is this article interesting.” The question is “is this article useful to me, this week, on the things I am shipping.”

The first criterion is my context. The second is my time horizon. The third is my leverage. The fourth is my trust network. Articles that score 7 or higher on those four questions get into my morning summary. The rest are dropped before they cost me attention.

You will need to rewrite the criteria for yourself. The structure is what matters: 4 specific questions that, together, define what “useful to me this week” means. Mine is calibrated to building homelab infrastructure and shipping AI agents. Yours might be calibrated to running a marketing team, or studying for an exam, or learning a new framework. The point is that the criteria are specific, and that an LLM can apply a specific test 200 times in a row, and you cannot.

The pipeline

The pipeline is 200 lines of Python. I will walk you through the shape of it.

The crawl runs at 23:00 WIB every night. A cron entry on my M720q fires a Python script. The script maintains a list of 9 sources — dev.to, The New Stack, Towards Data Science, MarkTechPost, InfoQ, and a few Indonesian tech sites. For each source, it calls a Playwright-backed crawler to fetch the latest articles, then parses the HTML to extract titles, URLs, and the first 500 words of body text. The crawl takes about 90 seconds for all 9 sources.

The LLM scoring runs on each article. For every article, the script calls the LLM with the criteria prompt above. The call returns a JSON object with a score from 0 to 10 and a one-line reason. Articles with a score of 7 or higher go into a list. Articles below 7 are dropped. The whole scoring pass takes about 3 minutes for 50 articles, which is the average nightly volume across all 9 sources.

The summary is built and sent to Telegram. The script takes the top 3 to 5 articles by score and writes a 2-sentence TL;DR for each, again via the LLM. The summary is formatted as a Telegram message and sent to a bot I registered for this purpose. The message lands at 7am WIB. I read it in 90 seconds. I open one article. I close Telegram. I start work.

The LLM API call is the only thing that costs money. At my volume — 50 articles scored plus 5 TL;DRs generated per night — the bill is under Rp 25,000 a month. The crawl, the cron, the script, the bot, the homelab box, the static IP, the electricity — all of it is either free or already paid for by the rest of the things the M720q runs.

Why this works for me, and might not for you

The reason this works is that the AI is not doing anything magical. It is applying a specific test I wrote, 200 times in a row, before my eyes see the article. The test captures 4 specific things I care about. Articles that pass all 4 are 80 percent relevant to the things I am actually building. That compares to the 3 percent I retained from skimming 47 newsletters by hand. The 27x improvement is not because the AI is smarter than me. It is because the AI is more consistent than me, on a question I could not answer at 7am with coffee still brewing.

It works because the criteria are mine. They encode my context, my time horizon, my leverage, and my trust network. They are not transferable. You have to write your own, against the things you are actually working on.

It works because the AI is allowed to say no. About 88 percent of articles scored each night fail the 7-point threshold. The AI drops them, and I never see them. The whole point is that the AI is willing to be ruthless on my behalf.

It works because the cost is low enough that I run it every night without thinking. The LLM API calls cost under Rp 25,000 a month. The crawl, cron, bot, and homelab box are already paid for. If the system cost Rp 500,000 a month, I would not run it. The version I run costs Rp 25,000 a month and 30 minutes of maintenance per quarter.

What to actually build this weekend

If you want to try this, you do not need a homelab. You do not need a fancy pipeline. You need three things: a list of 4 to 6 sources you actually want to learn from, a script that crawls them, and a 4-line prompt that scores relevance. The whole thing is a Saturday afternoon project.

Step 1: write the criteria prompt. Be specific. Mine is the one above. Yours might be “score from 0 to 10: (1) is it about a problem I am solving this month, (2) is it something I would not have found by myself, (3) is it from a source I trust?” The prompt is yours. The point is that an LLM can apply a specific test 200 times a day, and you cannot.

Step 2: set up the crawl. A cron job, a simple Python script, and a write to a JSON file or a database. If you have never written a cron job, this is the project to start with. The basic version is 30 lines of code. The homelab version with proper error handling and a Telegram bot is 200 lines. Both work.

Step 3: pipe it to Telegram. The point of using Telegram, not email, is that the summary shows up at the same place I already am in the morning. Email is where newsletters go to die. Telegram is where I am before I open my laptop.

That is the system. 47 newsletters in, 3 summaries out, 90 seconds of my attention. I learn more, with less effort, than I did when I was performing the act of reading 47 newsletters by hand and remembering nothing.

The bottleneck in learning was never access to information. It was the discipline to ignore most of it. An AI that ignores for you, on criteria you wrote, gets close to solving it. The whole thing fits in a single file. You can build it this weekend.

For the homelab + automated daily research setup I described, see my $0/month blog stack breakdown. The crawls run on the same M720q. The pattern I describe here is part of a larger approach I wrote about in my post on shipping AI agents before they are perfect. And if you want the technical deep-dive on what 32 AI agents hitting a homelab server at once looks like, the homelab AI workload guide covers the breaking point that pushed me to the architecture I use now.