English
Subscribe

Editor's notes

KakaoTalk's 37-Character Cutoff, +2 I Measured

KakaoTalk cards cut titles at 37 characters and drop the description. Plus 156 runs testing a "65% token savings" claim and a scorecard for 11 reels: three things I measured instead of guessing.

KakaoTalk's 37-Character Cutoff, +2 I Measured

The theme of this Friday's log is "measure, don't guess." I first guessed that KakaoTalk cards were cut by width, then corrected it to character count. I ran a "65% token savings" claim 156 times. And when I re-read the reel gap I had written down as 2.2x on November 2, it was 1.8x. In all three cases, the story changed once I measured. I am writing down what worked and what I got wrong.

1. KakaoTalk cuts titles at 37 characters

On the morning of Tuesday, November 3, I published ANG⁺ Briefing #3 and pasted the link into KakaoTalk, Korea's dominant messenger. The preview card showed only a truncated title, and the line describing the article was missing entirely. That morning I tested cards myself to find the rule, and the same day I audited all three sites I run.

What looked wrong

  • At the time, #3's Korean title was a long sentence listing all six items. The site appends " · ainewground" to the share title (og:title), so the title KakaoTalk received was 137 characters long.
  • The card showed the start of the title followed by "…", and the summary (og:description) line did not appear. By contrast, the site's market-map index (19 characters) and gptssam.com/new/ (31 characters) showed both the title and two lines of description.

My first guess was wrong

  • At first I assumed cards were cut by width — two lines, about 40 Korean characters — and wrote a rule to keep Korean titles under 26 characters (10:23am).
  • More test cards showed it was not about width. Three truncated cards (#3's old title, a 78-character gptssam guide, and the 38-character gptssam home page) were all cut at exactly 37 characters. Even the 38-character home page was cut, while cards with 31, 30, and 19 characters showed the title plus two lines of description. About 40 minutes later (11:05am), I corrected the rule to "37 characters by count."
  • Whether the og tags sat within the first 4KB of the page, or whether a description tag existed, was not the cause. I could not find a length rule in Kakao's official documentation or its developer forum (DevTalk), so this rule is based on observation and may change if KakaoTalk changes its behavior.

What testing taught me

  • KakaoTalk caches the card for each URL. After fixing a title, pasting the same URL again brings back the old card. When testing, I added ?t=<number> to the end of the URL so it looked like a new address.
  • Images are cached separately, too. So I added a version value derived from the file's contents to share-image URLs, so the URL changes whenever the image does (gptssam and hiddensideof).
  • It is all or nothing. Within 37 characters, you get up to two lines of title and two lines of description. Go over, and the description disappears completely. A single extra character costs you half the card.

Auditing 502 pages on three sites

  • The same day, I built a link-preview audit tool. It opens 502 pages across ainewground, gptssam.com, and hiddensideof.com, checks their og tags and images, and flags pages that KakaoTalk's 37-character limit would cut. Truncated titles, missing tags or images, and response errors on Korean pages are red; truncated titles, missing descriptions, and image ratios on other-language pages are orange.
  • The first run found 176 red pages. I shortened the titles of ANG⁺ #1–#3 and the market-map introductions and maps, and for hiddensideof and gptssam I created a table that keeps a shorter share title separately. By that morning, red was down to 33 pages. A rerun in the afternoon showed 32. The remaining pages are on my to-do list, in order.
  • Korean titles are now 23 characters or fewer, spaces included, so they still fit within 37 after the site adds its 14 characters. Lists of items go in the summary, with the key point in the first 45 characters. This post's Korean card title is 36 characters.
Try it yourself: in KakaoTalk's "Chat with myself," paste the URL you want to check, with any number added to the end, such as ?t=1234. You will see the current card, not an old cached one.

2. I measured "65% token savings" 156 times

Before installing the caveman skill, which passed 110,000 GitHub stars on a "65% token savings" claim, I measured it myself. On October 31, using Claude Code 2.1.314 and claude-opus-5-5, I ran the same prompts under four conditions: default, "Answer concisely." added in Korean before the question, caveman, and ultracave. Ten questions and three work tasks ran three times per condition, for 156 runs in total.

Results

  • Median output tokens on the 10 questions: default 1,100, "Answer concisely." 335, caveman 571, ultracave 499.5.
  • Caveman used 48.1% fewer tokens than the default. But it used 70.4% more than "Answer concisely.", which was shorter on all 10 questions.
  • Median output tokens on the 3 work tasks: default 863, "Answer concisely." 677, caveman 895, ultracave 798. With only nine runs per condition and overlapping ranges, I am not concluding that caveman increases tokens on work tasks.
  • Accuracy: a person checked all 156 answers, and all were correct. On these prompts, shorter did not mean wrong.

Two things that went sideways

  • The tool changed mid-measurement. When I opened a recording session during the first run, Claude Code auto-updated from 2.1.306 to 2.1.314. I stopped at run 36 and threw those numbers out. Then I pinned the binary, turned off auto-updates, and ran all 156 again from the start.
  • The grader got seven wrong. Reading the seven answers it had marked wrong, I found all of them correct. In six, bold markers (**) split the words so the text did not match; one was a correct sentence caught by a forbidden pattern. I only added a step to strip emphasis markers, and did not change the questions or criteria after measuring.
This does not mean "65% is a lie." The author's README also names "be concise" as the baseline. But on our 10 Korean questions, two words beat the skill. Before installing it, compare it with one line asking for a concise answer.

3. Scorecard for 11 reels: TikTok at 1.8x Instagram

This morning (November 6), I re-read the view counts for 11 reels from the Instagram and TikTok @gptssam profile pages. To keep my own views out of the numbers, I did not open any videos.

  • Total views: TikTok 3,126, Instagram 1,706, so TikTok is about 1.8x. On November 2, with seven reels, it was 2,646 and 1,198, or 2.2x.
  • The top reel differs. On TikTok, "Moving 3 Years of AI Memory" (IG003) has 929 views, while the same video has 151 on Instagram. On Instagram, the top is "The Formula for the Top 4 AIs" (IG002) at 437.
  • Instagram is flat. Apart from that one reel at 437, the other 10 all sit between 106 and 153. TikTok spreads from 48 to 929.
  • On the latest four (IG008–011) alone, Instagram leads: 503 to TikTok's 403. But these went up only days ago, so it is too early to judge.
  • Followers: 1 on Instagram, 11 on TikTok. In the November 2 table, the four reels with comment-keyword auto-DMs set up (IG004–007) all showed zero.
The sample is too small to say which platform is the right one. What I can say is that the same video can differ by about 6x between platforms. Next, before looking at more totals, I need to open Instagram's detailed metrics: skip rate and average watch time.

What's next

If any of these numbers change, I will write them up again. If you see a KakaoTalk card behave differently, or get a different measurement result, let us know at hello@ainewground.com.

Ad · from our publisher: free lectures released this week

These are the five lectures released this week. All are free and in Korean.

  • Is "65% Token Savings" Real? · about 9 min — Shows how to run the measurement in item 2 fairly. You can download the meter kit and test your own prompts, which suits Claude Code users who want to check a token-saving tool before installing it.
  • The Home Internet and Phone Contracts You Didn't Know Had Expired · about 14 min — Calculates the money leaking when a 25% contract discount has lapsed or an old bundle still applies, with a five-field bill checker and a 36-month total-cost comparison for budget-carrier bundles. For anyone who has not opened a phone bill in a while.
  • Reading a Building Register with AI Before Signing a Lease · about 13 min — Checks the risks that can get a jeonse deposit guarantee refused even with a clean title record, using five fields of the building register and a calculator based on HUG's official formula. For anyone about to sign a jeonse contract in Korea.
  • Building Sign-Up, Subscription, and Cancellation Screens Without Dark Patterns · about 14 min — Measures 16 screens from nine Korean and overseas services and gives a 12-item pre-launch scorecard based on Korea's E-Commerce Act as amended effective February 14, 2025. For anyone checking a cancellation flow before launch.
  • The Day My Channel Vanished Overnight · about 14 min — The real record of a Shorts channel that passed 1,000 subscribers and 1.03 million total views within 21 days of launch, then was permanently removed under the "circumvention policy" with zero warnings. If you run an AI video channel, start with the policy checklist and the AI-label comparison table for five platforms.