A 2 a.m. Schedule, +2 Things That Failed Quietly
The 2 a.m. schedule never fired, the KakaoTalk card and its link pointed to different posts, and the audio read numbers wrong. Three things that broke without an error message, and how I fixed them.
A 2 a.m. Schedule, +2 Things That Failed Quietly
about 4 min
The theme of this Friday's log is "automation that fails quietly." This week I added three things that run while no one is watching: a schedule that drafts the briefing before dawn, a link that announces the latest post in our KakaoTalk open chat every morning, and audio that lets you listen to an article in about four minutes. Each one went wrong once, and none of them produced a single error message. Every time, the problem was noticed first by a person who opened the result. Here is what broke, how I fixed it, and what I still have not confirmed.
1. The 2 a.m. schedule never fired
On Friday, November 6, I set up the publishing routine. At 2 a.m. on Tuesdays and Thursdays, a Claude Code session drafts the ANG⁺ Briefing; on Fridays, it drafts this editor's log. I get a review copy on my phone before 3 a.m., and only posts I answer with "deploy" go live.
What happened
- At 2 a.m. on Tuesday, November 10, nothing happened. No draft, no notification, no error.
- When I opened the app's log, the cause was not the schedule but the session holding it. The desktop app shuts down sessions that have been idle for a while, and the schedule lived only inside that session. With the session off, there was nothing to ring at 2 a.m.
- The session came back when I sent a question at 7 a.m. I started writing #5 after that and published it the same morning.
What I tested and chose
- That morning I had an app-level scheduled task, which lives outside the session, send a message to the session. The message went out at 9:32, and the sleeping session started again at 9:33:06. The wake-up method itself worked.
- But when I tried to make it a weekly recurring task, the app's automatic safety check blocked it. My best guess is that the turn had been started by a message from the test session, not by me.
- On November 11, I decided: "no schedule; I will ask on the days I need it." When a person asks, it runs for certain.
What surfaced again this morning
- This morning (Friday, November 13), there was no editor's log. With the schedule removed, that was expected. But it showed up somewhere else first: the "latest post" card that goes up in the open chat at 8 a.m. every day carried Thursday's #6 again.
- "Ask on the days I need it" meant it only runs if I remember the day. So today I set the schedule again. This time it is not inside the session: an app-level task wakes the session at 1:55 a.m. every Tuesday, Thursday, and Friday. The first run is Tuesday, November 17.
To check this yourself: a schedule being "set" and a schedule having "fired" are different things. Instead of checking whether its name is in the list, check whether the run history has a line at that time.
2. The KakaoTalk card and its link went to different places
Every morning at 8 a.m., one link goes up in the open chat: gptssam.com/latest. It is a short address I made on November 2 that redirects to the newest ainewground post, lecture, or guide. The address stays the same, but the idea is for the card to change every day.
What went wrong
- The KakaoTalk card (title, description, image) is set when KakaoTalk first reads the address. Where the link goes, however, is calculated by the server at the moment someone taps it. The two happened at different times.
- For example, the Wednesday 8 a.m. card showed Tuesday's news, but if a lecture went up Wednesday afternoon, anyone who tapped the card after that landed on the lecture. The card said one thing and opened another.
- At first I guessed that "a lecture posted the same day also overtakes the news." Rereading the code, that guess was wrong: on the same date, the ainewground post comes first. The real mismatch happened the next day.
What I fixed
- On November 10, I changed the destination from "the newest at the moment of the tap" to "the newest item published before the most recent 8 a.m." Now the 8 a.m. card and wherever you land all day long are the same.
- Posts with a publish time (published at 7 a.m.) make that day's 8 a.m. alert; lectures and guides that carry only a date become eligible from 8 a.m. the following day.
- I tested nine scenarios at different times, all passed, and then deployed. The response now also records which 8 a.m. it used as the cutoff.
What remains
- The rule has a cost: on days with nothing new, yesterday's post goes out again. #6 going up again this morning is the rule working as designed. The system ran correctly, but for the reader it is "that post from yesterday."
- On the morning of November 12, I sent the #6 link in KakaoTalk and no card appeared. The site's og tags had exactly the same structure as #5's. When I pressed "look up meta information" in Kakao's sharing debugger, the image, title, and description had all been read correctly, and after that the card showed normally too. I changed nothing on the site. I still do not know why it did not show the first time.
To check this yourself: if a card does not show, do not fix the site first. Look up the address in Kakao's sharing debugger and see what KakaoTalk actually read.
3. The listening audio read numbers wrong
Starting with #5 on November 10, every article gets listening audio. It is not a reading of the article but a podcast-style summary of about four minutes, with one host talking through the key points. The voices are made with Qwen3-TTS running on my own computer, in Korean, English, Japanese, and Chinese.
How I check it
- I transcribe the generated audio with Whisper and match the numbers and dates in the transcript against the script. Every chunk goes through the small model first, and suspicious chunks get a second pass with the larger medium model.
- I regenerate only chunks that both models transcribed wrong. If only one model got it wrong, I treat it as a transcription mistake rather than a voice mistake and leave it.
Where it went wrong
- #5, Japanese: "from November 5 to 9" came out as "from November 4 to here" in both models. I rewrote the dates in kana and regenerated.
- In #6, I regenerated 7 of 32 chunks (8 chunks × 4 languages).
- The Korean for "four dollars" came out as $1 in the small model and $3 in the medium model. Spelling it out as "four dollars for input, twenty dollars for output" fixed it.
- The English "four and twenty dollars" came out as
$420. That was the script's fault, not the voice's: I had written "$4 and $20" in a way a listener could also misread. - In Japanese, ordinal counters kept going wrong. "The fourth" (四つ目) came out as "the eighth," a kana version came out as "the sixth," and only 「四番目」 got it right. "A thousand at once" came out as "2,000," so I rewrote it in kana.
- So I added one rule: in Japanese scripts, write ordinals as 「〜番目」 or in kana, and prices in Arabic numerals.
What I still do not know
- Transcription only confirms the numbers. Where "Anthropic" and "Gemini" came out slightly misspelled in the Korean transcripts, I have not yet sorted out whether the voice or the transcriber was wrong. Listening to all four languages end to end with human ears is still to be done.
To check this yourself: do not only listen to TTS audio. Pull the numbers out of the transcript and set them next to the numbers in your script. Your ear may hear "four dollars" correctly, but that is no guarantee the machine, or every listener, will.
What's next
The next checkpoint is Tuesday, November 17, at 1:55 a.m. I will check the run history to see whether the schedule actually fired, and write it up in the next log. If you see a KakaoTalk card behave differently, or hear something odd in the audio, let us know at hello@ainewground.com.
Ad · from our publisher: lectures released this week
These are the four lectures released since the last log. The first three are free; the last is paid. All are in Korean.
- Build Your Own Content Archive Site with AI, No Coding · about 14 min · free — Build the 1200×630 KakaoTalk preview card and the
/latestredirect from item 2 yourself. For anyone who wants to gather scattered posts into one site. - Make Your Lecture Audio with Local TTS and Check It by Transcript · 13 min 24 s · free — Follows the checking method from item 3 step by step: match Qwen3-TTS audio against a Whisper transcript to find where numbers and negations changed. For anyone making lectures or audio with AI voices.
- Six Things I Actually Handed to Pickle · intro + 7 lessons · free — Real iPhone screen recordings of what the AI assistant Pickle did and where it got stuck: an overnight report, an 8 a.m. scheduled run, a real phone call, Government24 certificates, and three ways of logging in.
- Do It, GPTssam! — The Same 10 Jobs for Pickle, Dots, and Grok Bot · 55 lessons, 1 h 13 min 30 s · paid — A measured scorecard from handing ten jobs to three AI assistant apps on the same day with the same sentences, plus a way to judge your own tasks. The first three lessons are a free preview; early-bird price ₩99,000 (regular ₩132,000, through November 30), with access for 12 months from the payment date.