One narration in,
a full 4K picture layer out.

The episode:
eight AIs play Sokoban.

Eight models took the same six Sokoban puzzles, and one failed all six for a flat zero. The video is titled 我让8个AI玩推箱子，有个拿了0分，还怪题目无解？ (I had 8 AIs play Sokoban. One scored zero, then blamed the puzzle for having no solution?).

↗

↗

↗

The cover was generated with the built-in image_gen tool, each ratio, 3:4, 4:3 and 16:9, composed separately. It reads 8大AI推箱子 / 有个AI / 直接0分！ (8 AIs play Sokoban / one AI / scored zero!). Version two followed my note: 'The zero belongs to one of the AIs. The cover can't let people think none of them passed.'

The one in use is version three, after an independent review (AGY) whose suggestions I took: the bottom is now a deadlock, a figure pushing a crate into a corner under a single red cross; the eight logos are an unscored line-up that doesn't say who got zero; GLM uses the Z mark from its official site.

The edited narration runs 6:55.133, 24,908 frames at 60fps, in 14 chapters. Each chapter is one 4K MP4; end to end they cover the timeline exactly, and there is also one full-length file. I drop it into CapCut (剪映), add my avatar and subtitles, and it's ready.

Watch it first.
Everything here is from the 4K cut.

The four clips come from the 1080p preview with narration, compressed to 720p for the web, so you can hear the picture following the words; the stills come from the 4K film. The 4K files delivered for editing carry only music and sound effects, no narration. Every leaderboard on screen is public data. Spoilers below if you haven't watched the video.

↗

↗

↗

↗

↗

↗

↗

↗

Everything I gave it
was in the first message.

The pipeline has twelve stages, top to bottom. Each row says who does the work, which open-source projects and Skills it uses, and what comes out. Only the rows marked as stops wait for me.

1. 00

   Inventory

   Conditional stop

   Me

   Stops only if there is no narration, or the frame rate is not a constant integer.

   Lead

   Lists the script, audio or video, SRT and URLs and checks duration and frame rate with ffprobe; finds the product's source code and key art; watches 3–5 segments of the previous episode.

   Uses

   Product source (ModelDial Arena this time)

2. 01

   Edit

   Raw recordings only

   Lead

   Removes retakes, completes sentences, fixes breaths. From here on only the edited timeline counts.

   Uses

   ajiang-koubo-cut · faster-whisper

   Output

   Edited MP4 · SRT · word-level ASR

3. 02

   Understand · theme · chapters

   Before the BRIEF

   Lead

   Writes a 200-character summary and an errata list. Theme = the product's own art × this episode's show format. One chapter answers one question, 20–55 seconds, cut where a turning sentence starts. Four tables: reveals, line-by-line shots, wording of record, terminology.

   Output

   chapters.json · shotlist.csv (each asset belongs to one chapter)

4. 03

   One-command project

   Lead

   init_project.py: template, chapters, word alignment, entry points, narration slices, dependencies, tsc.

   Uses

   Remotion 4 + React 19

   Output

   timing.json · segN.tsx

5. 04

   Assets + privacy pass 1

   Scan on import

   Lead

   Ports the product's front-end animation and drives it with real data; takes 2x screenshots to prove what a page really says; extracts the product's colours and type and fetches logos, with inverted versions for a dark stage; scans every batch for QR codes and checks privacy by eye.

   Subagents

   Capture agent (optional): screenshots and real interactions; never opens settings or key pages.

   Uses

   ego-browser · lobehub icons · OpenCV · Codex CLI (last resort)

   Output

   Screenshots · tokens · logos · replay data

6. 05

   Chapter one = style sample

   Always stops once

   Me

   I watch the chapter-one preview and approve it to roll out.

   Lead

   Builds chapter one itself, 20–30 seconds, synced line by line; at least two shot cards plus the product's key art, logos and suspense handling; self-checks twice before sending it to me.

   Subagents

   Analysis agents ×2 (≤15 minutes each): one picks this episode's shot cards, one breaks down the previous film.

   Uses

   video-shotcraft shot cards and gallery

   Output

   Seg0.tsx · chapter-one preview and contact sheet

7. 06

   BRIEF + parallel dispatch

   Lead

   Moves chapter one's components into the shared lib and stress-tests them (a ≥40-second preview); writes the BRIEF with my approval quoted word for word in section 0; dispatches every chapter in one message and logs the IDs.

   Subagents

   Chapter agents ×N start, one per chapter (concurrency ≤ CPU cores / 2).

   Uses

   video-shotcraft aesthetic rules

   Output

   BRIEF.md · ledger.json

8. 07

   Collect + round two

   Lead

   Reads each contact sheet as its notification arrives and logs issues; patrols in the background and checks any agent with no new output for 30 minutes; sends targeted fixes with SendMessage, with timestamps and numbers.

   Subagents

   Reads the BRIEF, lib, chapter one and its lead card → writes its chapter → self-checks ≥3 times → reports.

   Output

   SegN.tsx · contact sheet · preview

9. 08

   Sound

   Parallel with 06–07

   Lead

   Extends the music, flattens its envelope, sets it 16 LU under the voice and cuts it per chapter; 10 semantic sound effects at a master gain of 0.3.

   Uses

   Mixkit music · Freesound CC0 effects

   Output

   Per-chapter music · mix.json

10. 09

    Independent review → round three

    Lead

    Triage: shared problems are fixed in the lib once everyone stops; the rest go back to each chapter's agent.

    Subagents

    Review agent (clean context): samples frames across the film and only reports.

    Uses

    video-shotcraft final review

    Output

    Issue list: severe / medium / minor

11. 10

    Preview gate + privacy pass 2

    Stops by default, skippable

    Me

    I review the preview page, unless I've said “不要问我了” (stop asking me).

    Lead

    Re-renders each preview → scans for QR codes again → builds the preview page.

    Uses

    Claude Artifact

    Output

    Preview page

12. 11

    4K + delivery + privacy pass 3

    Stops before overwriting

    Me

    If the delivery folder already has files, it asks whether to overwrite or create a new one.

    Lead

    queue.sh renders each chapter as soon as it passes; check.py verifies frame counts, loudness, chapter tails and seam brightness; the 4K film is scanned every 2 seconds; deliver.py hands it over.

    Uses

    Remotion render · ffmpeg concat and mix

    Output

    Chapters + full 4K film + 1080p preview + CapCut guide

The only places it waits for me

Chapter one is the sample for the whole film. The lead agent builds it itself and always stops once for me. v1 arrived at 23:19 and I approved v4 at 00:53: 94 minutes in all.

↗

1. v1 · 23:19Rejected

   A three-shot sample, cut loose from the narration

   ↗

   A push-in on a replay page screenshot, a hand-drawn 'illustrative' deadlock grid and a score card, cycling regardless of what I was saying.

   “What is this? Go and check it properly, will you? Your audio isn't moving with the picture at all.”

   I also told it to watch the finished film from episode 3.

2. v2Rejected

   Synced line by line, but a pile of screenshots that gave away the winner

   ↗

   Every line had a picture now, but nearly all of them were screenshot push-ins, and first place was on screen.

   “We need to keep the suspense. When we look at the strongest model, you can often make your own visuals. Don't show the model's name, get it?”

   “I think my narration should be the reference for wording.”

   “Not every page needs a real screenshot. Think about it: where a zoomed screenshot fits, where Screen Studio-style footage fits… when to use motion graphics. Think it through.”

   “You have to use the good animations in here.”

   “First use agents to analyse what we should reference, then make me a short first chapter. Once I confirm it, we'll settle the overall approach, then build the remaining chapters with the same theme, the same animation and the same style.”

   “In here” in the fourth note meant video-shotcraft.

3. v3Rejected

   Hidden too much: anonymous cards on black, no logos

   ↗

   Anonymous cards on black, a '1 dot = 1 point' dot chart, seven-segment numbers, and not a single logo.

   “I'm very unhappy. Your first segment has none of that feeling, nothing like a blockbuster. You also missed the point of this episode: 1. Don't give this episode's ranking, but you can show the leaderboards from the earlier Rubik's cube episodes, right? The cube doesn't matter. 2. You didn't use the logos either. 3. Those illustrative diagrams look ugly.”

   I also pasted lobehub icons' skill.md. The problem was in the dispatch: I had only said not to show the model's name, and the lead agent wrote it down as 'no names and no logos'.

4. v4 · 00:50Approved 00:53

   The product's own stage, shot like an esports broadcast

   ↗

   • The product's own stage art fills the screen, with an esports broadcast HUD;
   • The Sokoban board runs on a real replay; when step 7 creates a deadlock, a full-screen DEADLOCK banner sweeps across;
   • The real Rubik's cube leaderboard (earlier episodes, fine to show) and the real 8×6 results matrix, names replaced with ???;
   • A logo slot machine that stops on '?', 170px headlines, and a huge 91.6 and 0.

   “I think this version is pretty good. Do everything after this the same way.”

How the Sokoban board came alive

While v2 was in progress I said: 'The Sokoban code should move. We already animate it with front-end tech, right? All our Sokoban code is right here; you can absolutely reproduce it.'

ModelDial Arena is our own product, and its Sokoban engine and board rendering were already front-end code. The agent ported them into Remotion and drove them with real replay data. Not a single move on screen is made up.

1. The engine is the product's own Sokoban engine, so the rules match the live site.
2. The board reuses the product's board component and styles; the stage art and host illustration are the product's own images too.
3. Replays are compressed into step-by-step frames: whether each move walks, pushes or bumps a wall, and whether each attempt clears, deadlocks, restarts or runs out of moves.
4. The narration time maps to a step on the board, even halfway between two moves, so fast-forward, rewind and freeze all follow the words.

↗

Right after approving v4, I added:

“The rest is all yours. Have multiple agents develop in parallel, brief them clearly, and don't ask me. What I want in the end is a 4K film with everything rendered. Go on, I'm off to sleep… Leave a small circle in the top-right for my avatar, and the bottom of the frame for my subtitles… The active picture should be well laid out, not too small and not too big.”

The lead was Claude Code. One project, with permissions split by file: the lead owns everything global (theme, chapter one, the lib, cross-chapter deduplication, sound, rendering, talking to me), and each chapter goes to its own subagent. They share one lib, one set of tokens and one chapter one, so the result looks like one person made it.

How messages flow

What this episode measured

Six hours stuck on one command

In round one, seg8 and seg12 each sat on a single command for about six hours:

A delete with a variable and a wildcard. Most likely it raised a permission prompt; I was asleep and nobody clicked. The lead wasn't patrolling either, so it simply waited. The Skill now has two rules: commands like this are banned in the BRIEF in favour of clearing folders with Python, and the lead sets a background timer while it waits, checking which command any agent with no output for 30 minutes is stuck on.

Once every chapter was in, the lead started a review agent with a clean context and gave it only the previews, subtitles, BRIEF and review checklist. In about 15 minutes it found 21 issues: 2 severe, 10 medium, 9 minor. Both severe ones were repeats:

1. The same MiniMax level-1 replay appeared in both seg2 and seg4;
2. The full Rubik's cube leaderboard appeared three times.

Round two made targeted fixes to 11 chapters, 2–8 minutes each. 4K rendering ran as a queue, each chapter starting as soon as it passed: about 37 minutes for all 14. Delivered at 08:50.

B-roll成片/   (B-roll output)
├─ 0_开场悬念_00-00-00,000~00-00-27,000_字幕1-15.mp4
├─ 1_八强与计分规则_00-00-27,000~00-01-02,867_字幕16-39.mp4
├─ … 14 chapters end to end, 3840×2160 · 60fps
├─ 13_去试试人机对战_00-06-25,533~00-06-55,133_字幕233-251.mp4
├─ B-roll全片_4K_配乐音效_不含口播.mp4    full 4K film, 24,908 frames = master
├─ 预览_B-roll加口播_1080p.mp4           1080p preview with narration
├─ 无配乐版/                             same names, effects only
├─ 替换对照表.txt                        start/end, frames, subtitle numbers
├─ 音乐与音效来源.txt                    sources and licences
└─ 剪映贴法.txt                          CapCut placement guide

File names carry the SRT start and end in milliseconds plus subtitle numbers, so dropping them into CapCut in order just works. To avoid placing 14 files, use the single full-length film.

What was checked automatically before delivery

After watching it, I said: 'This is so well done!'

After the retrospective,
things I won't have to say again.

After delivery, two independent retrospective agents read the whole session log and wrote this episode's lessons back into the Skill. Next time, all of this is the default.

1. 01

   Chapter one is the sample and always stops once; after approval it doesn't stop again. If I've said “不要问我了” (stop asking me), the preview gate is skipped too.

2. 02

   Theme = the product's own art × this episode's show format, plus a checklist for that blockbuster feel.

3. 03

   Four tables before the BRIEF: reveals, line-by-line shots (each asset belongs to one chapter), wording of record (the narration wins), terminology. Both severe review issues were the same material showing up in several chapters.

4. 04

   Logos: fetch lobehub's mono, colour and wordmark versions, with inverted ones for dark stages; show a logo whenever the narration names a model; for identities still under wraps, a logo slot machine stops on '?'.

5. 05

   Dispatches relay my words, nothing added. This time 'don't show the name' became 'no names and no logos', and that sank v3.

6. 06

   The lead patrols: any subagent with no new output for 30 minutes gets checked.

7. 07

   Delivery always includes the full 4K film and a 1080p preview with narration; check.py now checks seam brightness.

Want to try it?
Set these up first.

npx skills add NanmiCoder/ajiang-vibe-coding-skills --skill ajiang-koubo-broll

Start here

1. 01

   Install ajiang-koubo-broll and video-shotcraft, and ideally ego-browser.

2. 02

   Give your agent the script, narration (raw recordings are fine), SRT and product URLs. If you have the product's source, tell it where.

3. 03

   Wait for chapter one. If it's wrong, say exactly what's wrong; it maps each note to its self-check list and revises.

4. 04

   When you approve, say whether it should ask you again. Say “不要问我了” (stop asking me) and it renders straight through to 4K.

5. 05

   Listen to the audio yourself after delivery: the agent measured loudness but didn't listen.

6. 06

   In CapCut: keep the edited narration on the main track and put the 4K film on the track above; duplicate the narration with a circular mask in the top-right corner; import the edited SRT as subtitles.

Thanks to
the open-source projects behind it.

This B-roll stands on these open-source projects and free assets. Each one is listed below with what it was used for.

Thanks to Vincentwei1021 for open-sourcing video-shotcraft; the camera moves in this Skill follow its cards. Thanks as well to LobeHub, Remotion, SYSTRAN, the authors of the CC0 sounds on Freesound and the musicians on Mixkit.