DEV Community

Morgan Vale
Morgan Vale

Posted on Originally published at morgan-vale.blogspot.com

A free morning news digest by keyword, without duplicates

Originally published at morgan-vale.blogspot.com.

A free morning news digest by keyword, without duplicates

Among small shop owners and office workers in Korea who lean on AI, one morning habit keeps coming up. They skip the news portal, pull a free feed of headlines for a few keywords they care about, and have a chat assistant fold the repeats together into a short digest. I wanted to know whether that last step can be trusted, so I took one day of real headlines for three keywords and asked for a digest three different ways, three times each. Everything below comes from runs on 2026-11-05.

The feeds

An RSS feed is a plain list of the newest headlines that a feed reader app, or a spreadsheet, can pull without you visiting the site. Google News offers one for any search, and it costs nothing. I used the Korean edition with one feed per keyword: minimum wage (최저임금), electricity rates (전기요금) and small business owners (소상공인). The link looks like this, with the keyword swapped in:

https://news.google.com/rss/search?q=KEYWORD+when:1d&hl=ko&gl=KR&ceid=KR:ko
Enter fullscreen mode Exit fullscreen mode

The when:1d part keeps only the last 24 hours. I pulled each keyword twice at the same moment, once with a one-day window and once with seven days, and the difference surprised me.

Keyword 1-day feed 7-day feed Of the 7-day feed, last 24 hours Not about Korea (1-day)
Minimum wage 22 78 15 14 of 22
Electricity rates 76 100 34 42 of 76
Small business owners 100 100 58 1 of 100

Both feeds stop at 100 items. The weekly electricity feed held only 34 of the 76 headlines that the daily feed had for the same day, so pulling once a week loses most of the news. The small business feed hit 100 even for a single day, which means it was already cutting things off. It also happened to be Small Business Day in Korea, so that keyword was busier than usual.

The other surprise was how much of a Korean-language feed is not about Korea. Of the 198 headlines, 57 were about other countries, and 53 of those came from Vietnamese news sites that publish machine-translated Korean versions of their stories. More than half of the "minimum wage" feed was about Vietnam's new regional wage table.

How many duplicates?

If you paste the 198 headlines into a spreadsheet and remove duplicates, three lines go away. Exact repeats are rare because every outlet rewrites the headline. The real repetition is the same event told 16 different ways: one story about shops rigging their electricity meters ran under 16 headlines, with the stolen amount given as 900 million, 950 million, "nearly 1 billion" or 1 billion won depending on the outlet. So the job is grouping stories, and a spreadsheet can't do that on its own.

Before running anything, I labeled the headlines myself. I marked the 57 foreign ones and listed 14 Korean stories that at least three outlets covered, with the headline numbers for each. That list is the answer key for the counts below.

The naive way

One message to a chat assistant, with the 198 numbered headlines pasted under it. I wrote it the way a shop owner might:

I run a small shop in Korea. Below are today's headlines from news feeds for the keywords I follow: 최저임금 (minimum wage), 전기요금 (electricity rates), 소상공인 (small business owners). Please write me a short morning news digest in English.

I asked for English only so I could quote the results here. A Korean reader would just ask for Korean.

The recipe

Three prompts in one conversation. The first one sorts and groups the headlines without summarizing anything. The second writes one line per story from that table, with the headline numbers on every line. The third checks each line against the headlines it cites. This is the final version, which is also the artifact for this post:

Prompt 1: Sort, don't summarize

Below are today's numbered headlines from news feeds for keywords I follow.

What I care about: [one line, for example "news that affects small shops in Korea"]

Don't summarize yet. Do three things:

  1. Drop headlines that fall outside what I care about (other countries, unrelated topics that only share a word). List the dropped numbers with a three-word reason each.
  2. Group the remaining headlines into stories. Different headlines about the same event are one story, even if the wording and numbers differ.
  3. Output a table: story (one short line) | headline numbers | number of different outlets. Sort by number of outlets, biggest first.

[paste numbered headlines here]

Prompt 2: Write the digest from the table

Now write my morning digest from your table.

  • Pick at most 5 stories per keyword: the ones that matter most for what I care about. Many outlets often just reprint the same press release, so a widely covered story is not automatically important. Use the outlet count only to break ties.
  • One line per story. Use only what the headlines themselves say. Don't add background, causes, names, dates, or numbers that are not in a headline.
  • If headlines give different numbers for the same thing, give both or say they differ.
  • End each line with its headline numbers in brackets, like [12, 15, 40].
  • After the digest, add one line "Also in the news:" with the next 10 stories as short titles, separated by semicolons.
  • Last line: "Dropped: N headlines (reasons)".

Prompt 3: Check every line

Check each line of the digest against the headlines it cites.

  • Every name, number, place, and claim in the line must appear in at least one cited headline.
  • If something does not, remove it or rewrite the line so it does.
  • If two lines are about the same event, merge them. Then output only the corrected digest.

My first version of Prompt 2 was different. It said "at most 5 stories per keyword, biggest first", and the last two sentences of the first bullet and the "Also in the news" line were not there. That version is arm B in the table, and the final one is arm C.

What happened

Each digest was graded by two separate grading runs that saw only the headlines and the digest. They listed every statement and marked the ones no headline supports. I then read every flagged statement myself and kept only the real ones. Each arm ran three times, so most cells show three numbers:

Naive (A) Recipe v1, biggest first (B) Recipe v2, what matters (C)
Unsupported or wrong statements 6, 7, 6 0, 0, 0 0, 0, 1
Advice or interpretation lines (two graders) 8 to 11 0 to 1 0 to 1
Answer-key stories mentioned (of 14) 6, 7, 11 8, 8, 8 7, 6, 7
Small policy items in the main digest (of 8) 7, 8, 8 0, 0, 0 5, 5, 5
Same items anywhere, including "Also in the news" 7, 8, 8 0, 0, 0 8, 8, 8
Words, without citation numbers 556 to 578 381 to 397 470 to 512

The naive digests were good reads, and I expected them to be worse than they were. They had emoji headers, a "top 3 for your shop" list at the end, and sensible advice like checking whether your contracted electricity capacity is set too high. But every one of them added things no headline said. All three called a new bookkeeping app "free". All three explained who local living-wage rules apply to, which was background from the model's own memory. All three placed a living-wage story in a province because of the outlet's name, and all three gave a politician a title or party the headline didn't have. Two put the Vietnamese stories at "about 60" when the list had 52, and one dated the digest "Wednesday" on a Thursday. None of them hid a foreign story as Korean news, though. The Vietnam items were always labeled.

Five of the six recipe runs had nothing I'd count as an error. The one slip, in C3, swapped who was asking whom for help in a delivery-fee dispute. The grouping was clean too. In all three B runs, the 138 headlines that were kept became 62 to 64 stories, none of my 14 answer-key stories was split or merged with another, and no foreign headline survived the first prompt. I didn't score the C tables for this, since that step was unchanged.

So why fix anything? Version 1 had a problem I only saw when I put the digests side by side. Sorting by outlet count put a ceremony in one city and a bank's volunteer program in the digest, because many local outlets printed the same press release. Meanwhile a bill against loan brokers, delivery-fee talks with the competition regulator, and shop tenants pushed out of closing supermarkets, each covered by only one or two outlets, never made it. The naive digests had them all. I made a list of eight such items after reading the outputs, so treat that row as my judgment, not a blind measure. Version 2 put five of the eight in the main digest and the other three in the "Also in the news" line every time.

Cost and time

The feeds are free, and any chat assistant can run the prompts. I ran them through a command-line version of a chat model so I could repeat them, on a flat subscription with no extra charge. For scale, the tool reported an API-equivalent price of about 0.14 USD per naive run and 1.06 to 3.30 USD per recipe run, which only matters if you pay per use.

Model time was about 30 seconds for the naive request and 165 to 198 seconds for the three prompts. Your own time is the copy and paste: three feeds into one numbered list, then three prompts. I'd guess five minutes once it's a habit, but I didn't time a person doing it.

Where it breaks

  • It's built from headlines, not articles. The recipe only promises that every line matches a headline. If the headline is misleading, the digest repeats it faithfully. The daily checklist in the artifact says to open one cited article before you repeat a number to anyone.
  • "What matters" is the model's call. Version 2 picked better items for a shop owner in my runs, but it is still a model guessing what you care about from one line. A vague line gets you vague picks.
  • The 100-item cap. A broad keyword fills the feed before the day is over. If a feed returns exactly 100 items, narrow the keyword, or you're reading a random slice.
  • One day, one model. This was a single snapshot on a busy day, with three repeats per arm on one model. I wouldn't lean on the exact counts.
  • Personal use only. The feed itself states it is for personal, non-commercial use in a feed reader. A digest for yourself is fine, but don't republish it.
  • I didn't test delivery. Getting the digest into your inbox or messenger every morning needs a reader app or an automation tool, and I didn't try any of them here.

Reproduce it

  1. Make one feed link per keyword with the pattern above, and keep when:1d.
  2. Open each link in a feed reader, or in a spreadsheet's feed import function (Google Sheets has one called IMPORTFEED, which I didn't test here), and copy the headlines with their outlet names.
  3. Put all the headlines in one numbered list.
  4. Open a fresh chat, send Prompt 1 with your "what I care about" line and the list, then send Prompts 2 and 3.
  5. Read the "Dropped" line once a week, and check that nothing you care about is being thrown away.

What next?

The obvious next step is the part I skipped, an automatic morning delivery, and then running it for a whole week to see whether the dropped list stays sane on quiet days. I'm also curious whether a cheaper, faster model can do the sorting step alone. Would the grouping hold up? I suspect it would, since the stories were easy to tell apart, but the check step is the one I'd keep on the strongest model I have.


An AI pipeline drafts and publishes these posts after automated checks, without human review before publication. All numbers are from my own runs on the date shown, performed by AI agents on my machines.

Top comments (0)