my favorite use for a coding agent is korean rom-coms
what happens when you build a spoiler-bounded multimodal context protocol for korean romantic subtext

About 6 months ago, I started watching Kdramas and got confused by a bowl of ramen, and since then, I’ve collected dozens of scene notes, written a custom media-player script, built a taste file explaining my type, and given Codex enough context about my romantic life that some of its conclusions will never leave my laptop.
So how the hell did watching television turn into a software project?
10 seconds of eye contact
Korean shows leave much longer gaps between lines than the Western shows I grew up with, and 6 months in, I still haven’t adjusted.
2 characters lock eyes and hold it for 10 full seconds, and when somebody finally looks away, the other notices and neither of them breaks the awkward silence. Most Western rom-coms I grew up with would’ve cut away 7 seconds ago, but Korean television gives me enough time to see who broke eye contact.
I understood something in those 10 seconds because of everything that had happened earlier, but I couldn’t point to a single line of dialogue that had told me. The subtitles gave me the words and left out what those words meant between these 2 characters.
As you go through this article, you’ll see this workflow marker show up.
- Something Codex could already do.
- The new ability I just added.
I’m going to do my best to explain how my Codex workflow gets more capable as we go, with the older abilities in regular type and whatever I just added in bold.
the subtitles are technically correct
I don’t understand Korean, so I needed English subtitles, and every so often, the subtitle looked completely ordinary while everybody on screen suddenly froze or started yelling.
About 3 or 4 weeks into my first Kdrama, I got annoyed enough to open Codex, and a long-running chat had already helped me build software, so why not use one for Korean romantic subtext?
At first, I told Codex what I was watching and roughly how far I’d gotten, and anything after that point was off-limits. I’d dictate whatever I thought I’d just seen and go back to the show as soon as Codex filled in the missing context.
I asked why a perfectly normal offer of ramen had made everybody lose their minds.
- You
- Codex
Then the pillow fight happened.
- You
- Codex
I gave every show its own chat, and early on, Codex had my rough speech-to-text description plus whatever it already knew or could find online, and that was usually enough to explain the social cue I’d missed.
Some answers took 30 seconds, other times I looked up 30 minutes later and realized I was still reading about Korean honorifics because 2 characters had dropped them in the middle of a pillow fight.
Giving Codex the context by hand
I described the scene and set the spoiler cutoff for every question.
- Answer a spoiler-safe question about a scene.
sometimes there isn’t even a question
Sometimes, I wasn’t looking for an answer, I just wanted to use Codex like a journal, so I got into the habit of starting every show chat with the same rule.
- You
- You
- Codex
6 episodes into the same show’s chat
Those little entries piled up fast, and by episode 9 of Something About 1 Percent, I’d spent 6 episodes dictating reactions into the same chat, including plenty I’d forgotten. I asked about 3 forms of address at once, and Codex already knew who the couple was, how their contract had started, what they’d argued about, and which promises they were pretending not to care about.
- You
- Codex
If I’d started a fresh chat, Codex could’ve defined the words and stopped there.
I’m terrible at remembering Korean names, but Codex had enough context to figure out who I meant anyway.
- You
- Codex
Using the show chat's history
One chat per show let later questions use earlier scenes and reactions.
- Answer a spoiler-safe question about a scene.
- Use earlier context from the same show.
Codex got the right actress from an awful description that would’ve been useless in a search box, so I got bolder and asked about every little Korean television detail I’d usually shrug at and forget.
the sandwich is sponsored, the car is wearing a fake mustache
Subway is impossible to ignore once you start watching Korean TV, and it looks strange enough to Western audiences that John Oliver spent a segment explaining it. The logo keeps appearing behind completely serious conversations while actors hold the sandwiches towards the camera, and nobody on screen acts like any of this is strange.
- You
- Codex
- You
- Codex
For comparison, Community took Subway’s money and gave a character the legal name “Subway,” and the completely absurd sponsorship literally was the joke.
After 30 or 40 episodes of seeing Hyundai and Genesis cars everywhere, I assumed Korean television would be thrilled to leave the badges alone. Then the logos started disappearing, some were covered and others had a completely different emblem in the same spot. Then an unmistakable Hyundai Ioniq 5 appeared with a plain black oval where the H should be, like somebody had given the car a fake mustache even though it was still obviously an Ioniq 5.
I couldn’t name the black oval, but I could show it to Codex, so why not drop the frame into that show’s chat and let it do most of the explaining?
- You
- Codex
I knew I’d look up why the badge was covered again, and I’d saved the explanation only in the show chat.
if i paid for the tokens, i’m keeping them
If I’d already spent an hour getting an explanation, why leave it in a chat I might never open again? Codex already let me turn repeated instructions into a skill I could call later and keep a useful chat as a document I could point to, so I saved the Kdrama chats too, text and media.
I moved the text into Obsidian, where I already kept my homelab notes as Markdown. Codex could read any server configuration or networking decision I’d written down there, so I threw the half-dictated Kdrama reactions into the same vault.
- You
- Codex
Saving the chat as Markdown
I moved the questions and reactions from each show chat into Obsidian as Markdown.
- Answer a spoiler-safe question about a scene.
- Use earlier context from the same show.
- Read the saved show note in a new chat.
Once the chats were in Obsidian, I found the explanation again, but “somewhere around the middle of episode 6” still left me scrubbing through an hour of video, so why not save the exact moment too?
please don’t tell me what happens next
A timestamp would get me back to the scene, and Codex still needed to know that anything after it was off-limits, so I typed the show, episode, rough timestamp, and spoiler cutoff into every question by hand.
- You
Searching for a character was risky because cast pages and thumbnails from later episodes could tell me what happened next, and Google only needs half a name to identify somebody’s biological mother.
Asking about the Ioniq taught me to send the frame first, and IINA, the media player on my Mac, already knew which file it was playing and the exact timestamp. The filename gave me the show, season, episode, and title, and I could pull a small subtitle window around that timestamp. The rest of the episode never entered the chat.
The frame, filename, timestamp, and subtitle window were already there, so why was I still warning Codex about spoilers by hand?
- You
- Codex
Codex wrote a small script that I loaded into IINA and bound to the s key. Now I press s and get the current frame plus a JSON block on my clipboard with the file name, timestamp, bookmark, and snapshot path.
The script handled the keypress, but Codex still needed the same instructions every time, so I turned those instructions into 2 skills I could reuse. The Kdrama skill handles the show note and passes the capture to scene-save, and scene-save reads the JSON and pulls the subtitles around that timestamp for the note.
<scene_capture>
{
"source_type": "sonarr",
"parsed_metadata": {
"show": "The Story of Park's Marriage Contract",
"season": 1,
"episode": 8,
"episode_title": "On a Rainy, Windy Day",
"filename": "The Story of Park's Marriage Contract - S01E08.mkv"
},
"video_path": "/TV/Parks-Contract/S01E08.mkv",
"timestamp_sec": 3226,
"timestamp_display": "53:46",
"bookmark_url": "iina-bookmark://open?file=S01E08&start=53:46",
"snapshot_path": "/captures/parks-contract-s01e08-53m46s.png"
}
</scene_capture>I pasted the JSON into the show chat and added what I thought was happening and why I cared.
- You
- Codex
Codex turned the JSON and my half-dictated reaction into an Obsidian note I could reopen weeks later, and I added the cast links afterward.
---
show: "The Story of Park's Marriage Contract"
season: 1
episode: 8
timestamp_display: "53:46"
screenshot: "../assets/s01e08-hide-your-pain-like-an-idiot-3226.jpg"
characters:
- "[[cast/Kang Tae-ha]]"
- "[[cast/Park Yeon-woo]]"
---
## The Scene
After [[cast/Kang Tae-ha|Tae-ha]] admits that he likes her,
[[cast/Park Yeon-woo|Yeon-woo]] asks whether he plans to
hide his pain and keep quiet about it like an idiot.
## The Hit
Good. He needed the roasting.

[▶ Play in IINA](iina-bookmark://open?file=S01E08&start=53:46)i know the face, who the hell is that?
The timestamp got me back to the scene, but I still had to match each Korean name to a face by hand, and Plex and Jellyfin already pulled the episode details and character images from TMDB and TheTVDB, so why wasn’t I putting those images beside the names in Obsidian?
- You
When I draw a blank on a Korean name, I hover over it in Obsidian and see the character in costume. If Codex picked the wrong person, I fix the link.
2 weeks later, who is anybody?
With the running note, cast pages, saved scenes, and exact spoiler cutoff in one folder, I wondered how much a fresh chat could reconstruct. When I came back 2 weeks later, I pointed one at the show folder.
- You
- Codex
The Play in IINA link took me back to the exact timestamp, and my note reminded me why I’d stopped there.
Then I’d find some old theory about a hand movement and, of course, want another episode.
There is literally no person on Earth I’d make reconstruct 2 weeks of my theories and then get excited about a weird face halfway through episode 9, and Codex can follow all of it, so I don’t have to wonder whether I’m being annoying.
Building a reusable show directory
The scene note kept the exact moment and linked each name to a face. The latest note gave a new chat the spoiler cutoff.
- Answer a spoiler-safe question about a scene.
- Use earlier context from the same show.
- Read the saved show note in a new chat.
- Read the latest scene note and rebuild the show context in a new chat.
If one folder gave a fresh chat enough context for one show, what could Codex find across a few dozen of them?
ask me cold and i have almost nothing
I was collecting more little observations than I could remember, so I wrote a short verdict at the top of each show note. Months later, I could read that one line and remember exactly how I’d felt about the show. Somewhere in all of that, there had to be a pattern, unless I’d been inventing one the whole time.
Ask me cold what I like in a Kdrama romance and I have almost nothing, but Codex could compare all the scraps, and I only trusted it when it showed me the scenes.
- You
- Codex
Comparing notes across shows
Codex compared saved notes across shows and found reactions I kept having.
I didn’t want to think about Codex’s last sentence yet, and figuring out why Young Lady and Gentleman annoyed me felt much easier.
i only knew i wasn’t enjoying it
Sometimes I don’t know why a show isn’t working for me, I only know it isn’t, and with Young Lady and Gentleman, I couldn’t get any more specific, so I told Codex exactly that and pointed it at the saved notes.
- You
- Codex
- You
- Codex
I didn’t like how accurately Codex had read my reactions across the shows, and my personal notes were going to be worse.
My personal directory has years of journal entries about relationships and some very bad times, including the hard decisions I was trying to work through, and Codex already reads those files when I ask it to help me think through something in my life. I knew it was going to tell me something I didn’t want to hear, but screw it, I opened another chat.
- You
- Codex
I was embarrassed by how accurate it was, so I read some of it twice, closed the laptop for a minute, reopened it, and got annoyed that Codex could quote my own notes back at me.
And no, I am absolutely not sharing any of those here.
i have rules for this now
Codex had found the same preferences across several shows and in my personal notes, and if Codex was going to say something that personal, I wanted the rules in a file I could argue with later. A rating left out the reasons and exceptions I cared about, especially when I might save 20 screenshots from a show and drop it anyway.
- You
- Codex
- You
- Codex
Building a taste file
I asked Codex to test its first guesses against more shows and keep the rules soft.
Once those rules were written down, I didn’t want to spend another 16 hours on the wrong show.
okay, but will i like it?
For each new show, I opened a chat and had the Kdrama skill load the taste file, then asked about the premise and relationship setup. I left the ending off-limits and tried Jealousy Incarnate first.
- You
- Codex
Codex said I might like some of the early chemistry, so I ignored everything else it had said and made it through 6 episodes.
After that, I asked about The Law Cafe, and Codex was much more confident.
- You
- Codex
Sure enough, by the time I finished watching The Law Cafe, I’d saved 30+ scenes and greatly enjoyed the conclusion, so Codex absolutely nailed this one.
fine, codex was right. what did it miss?
Codex had predicted I’d like Love Your Enemy too, and I finished it.
- You
- Codex
Correcting recommendations
I brought finished shows and drops back to Codex for the next taste-file review.
- Answer a spoiler-safe question about a scene.
- Use earlier context from the same show.
- Read the saved show note in a new chat.
- Read the latest scene note and rebuild the show context in a new chat.
- Compare saved notes across shows.
- Save each preference with its evidence and exceptions.
- Review finished shows and drops when I ask.
I still decide when the taste file changes.
Eventually I want the Kdrama skill checking it whenever I save a scene or leave a verdict, but I haven’t told it to do that yet. I have no idea how many scenes should change a rule built across several shows.
Most of the time, I don’t know what I’m saving a scene for, and when something makes me stop, I press s and tell Codex why, then go back to the episode.
I might not touch the note for months, then another show makes the old reaction look familiar and I finally know what to ask.
I built a spoiler-bounded multimodal context protocol because everybody except me knew the ramen was about sex.