DEV LOG · GLASS BOX ENTRY · team jf1shh · Unreal Engine 5.9 · 2026 Epic AI MegaJam

How Jump Cut was made, in the open.

Solo entry from team jf1shh, made in Unreal Engine 5.9 for the 2026 Epic AI MegaJam. The theme was "Director's Cut". This is my entry for the Glass Box prize, so the page covers how the game was built as well as what it is. I worked with a team of AI agents (Claude, through Claude Code). They wrote almost all of the code and ran the tests. I made the calls, played the game, and submit it by hand.

13,710,362,734
tokens processed since the window opened
90,892,528
tokens the agents wrote (output)
182
game saves in the window (Diversion)
1,013
numbered hub-relay blocks (clock-stamped)
209
hub seat number now (hand-overs, no human in the loop)

Every figure on this page comes from a script that reads the real logs, saves and hand-over notes. Numbers measured at 2026-11-15 18:41:05 PST. Where nothing was measured, the page says so.

SCENE 01who did what

The crew

In a build test, the lead makes a numbered build of the game. The handheld packages it into a zip, starts it, plays a script, and checks the log for lines like "results screen shown". If everything passes, the build is called CLEAN. If not, the list of problems goes back to the lead.

The hold look during a MUST CUT, from the final capture
Holding SPACE on a MUST CUT: the look while the cut is held.
SCENE 02how much the agents read and wrote

Tokens: the running meter

Tokens are the units an AI model reads and writes. This is every token the jam sessions on this machine and on the test handheld logged since the window opened at 9 PM PT on Nov 11.

90,892,528output: what the agents wrote
318,791,152fresh input + cache writes
13,300,679,054cache reads: context re-read, 97% of the total

Running total across the jam. Flags mark moments from the hub relay. Hover for the value. The running-total and per-hour charts leave out the test handheld, whose hourly split was not measured; the day, model and role charts and the headline include it.

Per hour: new work against re-reading

By day

By model

By role

Role is a rule, not a label from Claude: hubs, leads (the Unreal project session) and helpers (every sub-agent and other session). Cache reads are the same context read again at a lower rate, so the all-token view is mostly cache; the output view shows what was actually written.

The plan-usage crunch

The percent of the plan the coordinator agent logged in its own block headers; the line only connects the figures it quoted, so gaps mean none was quoted. The drop is the reset I clicked; the usage rules that followed are in the Misses scene.

Same numbers as a table
dayinputoutputcache readcache writeall
2026-11-113,234932,768188,899,7333,101,769192,937,504
2026-11-1259,35914,754,3692,835,930,32344,948,2112,895,692,262
2026-11-1346,23925,879,0363,680,204,96677,007,2593,783,137,500
2026-11-1476,36143,812,6995,891,146,562171,545,6196,106,581,241
2026-11-1511,0815,513,656704,497,47021,992,020732,014,227
modelinputoutputcache readcache writeall
claude-fable-5-122,544944,30353,392,7234,305,25158,664,821
claude-fable-5-211,67810,757,0931,557,045,03124,966,7491,592,780,551
claude-haiku-5-52,0531,884,45062,753,2829,546,82674,186,611
claude-opus-52,3461,815,789242,862,4274,771,406249,451,968
claude-opus-5-5100,90943,229,5747,900,031,078136,340,7138,079,702,274
claude-sonnet-5-556,74432,261,3193,484,594,513138,663,9333,655,576,509
roleinputoutputcache readcache writeall
helpers138,38661,655,1438,404,208,206242,221,4178,708,223,152
hubs25,01015,422,9582,874,101,36037,943,0832,927,492,411
leads7,1902,566,994430,818,0147,809,126441,201,324
test handheld25,68811,247,4331,591,551,47430,621,2521,633,445,847
SCENE 03no human in the loop

The hub relay: a baton, now at hub 209

One Claude session at a time sits in the hub seat. When its memory fills, it writes a hand-over note and the next session starts and takes over. The hub seat is now number 209; this page was ordered by hub 209. Hand-over folders written since the window opened: 91; a hand-over every 50.5 minutes (median). Hubs handed off at a median of 296K tokens of context (20 logged).

Each dot is a hand-over folder, placed at its first file time. Ringed dots are also takeovers recorded in the hub-orders file.

The relay log has 1013 numbered, clock-stamped blocks (block 69 to 1080); its surviving record begins on Nov 13 at 06:39, after the file-loss notice. The hub also wrote 119 project checkpoints to the shared notes vault since Nov 11 (per day: 13, 13, 34, 39, 20).

SCENE 04numbered saves on Diversion

The build ladder: every save, passes and fails

182 saves in the window (save 3 to 184). Each square is one save, in order. Green: a PASS/GREEN word on it. Red: a FAIL/RED word. Amber: both. Grey: saved, but no gate word was found for it (not the same as untested). Outlined: a ship or final save.

pass 43 fail 7 both 1 no gate word 133

The latest save with a logged test line reports 224 tests passed and 0 failed. Relay blocks with a gate result: 100 with a pass word, 20 with a fail word.

SCENE 05adversarial, new-player and makes-sense gates

Reviews: the agents checking each other

Fresh agents with no notes read pictures of each build and said what a new player would do. Other agents tried to break each patch before it was saved. Counts are relay blocks whose header names the review kind (a block is not a person, and one block can cover several reviewers).

SCENE 06from the draft, in my words and the agents' words

Day by day

Day 0: Wed Nov 11

tokens 192.9Moutput 932.8Ksaves 19relay blocks not measured

Before the window: Epic revealed the theme, "Director's Cut", on the 11 AM PT kickoff stream. The contest window did not open until 9 PM PT. In between, the agents wrote idea notes against the theme and I picked one. At 1:31 PM I said "Yes, lock it" and Jump Cut became the design doc. It was text only: no project, no level, no code, no art. The agents then wrote a prep pack from it (shot lists, HUD sketches, a metric sheet per scene, a data table) and made 17 labelled reference pictures. Every one of those pictures says "PRE-JAM REFERENCE, NOT A GAME ASSET" across the bottom and none is in the game. All of this, plus the tools and scripts built in the weeks before, is listed in the pre-jam declaration.

At 9 PM the window opened. The Unreal project was created at 9:03 PM. The first C++ files landed at 9:04: the project set-up, a plain blockout level (gray shapes), a press-to-start screen and a sky. The jam plan said work until 1 AM, then sleep until 8:30.

Who did what: The agents had written the set-up scripts and checklists before the jam; those tools are in the pre-jam declaration. Every game file, level, Blueprint, mesh, sound and line of gameplay code was made after 9 PM. The lead built the blockout. I watched and answered questions.

The decision I made: "Yes, lock it." Three concepts were scored, then a review tried to break them. Jump Cut scored 23 of 35 on a realistic build, against 20 and 19 for the others.

The tooling and workflows had early hiccups, which made me nervous. I had slept little the night before and was tired but excited to build.

Thu Nov 12 (day 1)

tokens 2.90Boutput 14.8Msaves 85relay blocks not measured

What got built: most of Scene 1. A car drives down a road. Hold to look, release to cut. Driving a stunt clean scores up to +3. Cutting scores nothing and spends one piece of film. A crash gives one free retake. Each result gets a stamp: WIDE, CLOSE, PERFECT, CUT. A wrap screen shows points and stars. Music and sound effects came later that day, from a synth script (more below). I also added a title screen and a quit key.

Who did what: The lead wrote the code. The handheld ran builds B5 to B9 and sent back reports. Helpers wrote the look guide and the plan for Scenes 2 and 3. I read each report and said go or fix.

What broke:

The decision I made: No more posting to itch until the very end. Nothing would go up, not even the private test page, until one upload on Sunday that I do by hand. (On Sunday I uploaded both zips to the private draft; see Nov 15.)

The PERFECT stamp, from the final capture
The PERFECT stamp after a clean cut.

B10 test (build 65): this build adds a "SECOND UNIT presents" card at launch, hints that change with keyboard or gamepad, and no open ports. B10 did not come back clean, so the page video moved to a later build.

Later that day, test build 11 fixed the window (full size on the handheld) and turned off one heavy rendering feature (Nanite) on weak graphics chips, with the same look. Test build 12 fixes the CLOSE stamp and adds stunt sounds, a loudness limiter and the first Scene 1 look (real props, a dashed-lane road, a studio skyline). On the handheld, this build removed the first-time cut hitches but still missed the frame-rate target, so a performance pass came next.

The day entry as filed (2026-11-12)

(Written on Thursday Nov 12; later entries say what changed.)

Correction 2026-11-13: the music was replaced on Nov 13 with tracks generated with ElevenLabs Music (via fal.ai). The lines below describe the earlier code-made music.

Jump Cut, by jf1shh (solo entry). Theme "Director's Cut". Day 1 of the jam. Times Pacific.

What shipped today

  • A playable Scene 1 with a score, sound and a wrap screen. Hold to look, release to cut. Driving a stunt clean scores up to +3, a cut spends one piece of film, a crash gives a free retake, and THAT'S A WRAP shows points and stars. The music and five action sounds come from a synth script: no samples, no AI audio.
  • Test build 9 passed, but the game was listening on a network port, left over from the engine's debug trace tool, so a judge could get a firewall pop-up first thing. Build 64 compiled that tool out.
  • Test build 10 (game build 65) closed the port and failed the window check. No listening ports this time. But on the handheld's sharp screen the window opened at 960 by 536 and was stretched to twice its size, so it looked blurry, and it ran at 42 to 48 frames a second. The build agent found the cause: high-DPI support was off, and our own window code forced a small window. That fix is in build 71.
  • Test build 11 (game build 72) passed, except one stamp. The window now opens at the handheld's full 1920 by 1200, and the SM6 cook cut the primitives from 103,732 to 18,780. Nanite is now turned off on weak integrated GPUs, with the same look. At that point the CLOSE near-miss stamp never fired (fixed in test build 12 below). Captures for the page are being recorded at a steady 30 frames a second until the frame rate is fixed.
  • Test build 12 (game build 92) was in test when I wrote this. It fixes the CLOSE stamp, so near misses come in three sizes (WIDE, CLOSE, PERFECT). It also adds a sound and a warning cue for each stunt, a loudness limiter so nothing clips, and the first Scene 1 look: real props, a dashed-lane road and a studio skyline behind the set. One prop came out of the generator with a white car baked onto its crane arm, so we took it out of the scene. On the handheld, this build removed the first-time cut hitches but still missed the frame-rate target, so a performance pass came next.
  • The audio kit was made, but not in a build yet. 14 sound effects and a four-part music track (70s stunt-show funk), all made in code with no samples; 8 short director voice lines from AI text-to-speech; and a mix plan that ducks the music under the voice. It was ready for the build agent to import.
  • The page kit for Sunday. The cover and banner (the cream car I picked), the page text, the entry form answers, the credits, and a one-command pipeline that turns the handheld's captures into the 45-60 second video, stills and a GIF. Nothing was to go to itch until Sunday (on Sunday I uploaded both zips to the private draft).

The performance find

The same Scene 1 that draws 1,793 primitives and no shadow passes in the editor drew 103,732 primitives and 18 shadow draws on the handheld. The Windows build was cooked for the older shader model only (SM5), so Nanite was off and every mesh fell back to a heavy version. The next test build cooks for SM6 with SM5 as a fallback, plus a lighter preset for the handheld's integrated GPU. A desktop check of build 71 (72 draw calls, 2.96 ms of GPU time a frame, nearly half of it upscaling) pointed at the handheld's integrated GPU as the limit.

Decisions since the last entry

  • Goal: 35 out of 35 (me, 09:16). Every judging criterion at 5. If something has to give, Fun and Gameplay stay at 5 no matter what. My own words: "we want AAA quality, we want FUN, we want addictive and immersive gameplay, and we want to WIN this competition."
  • Nothing goes to itch until the end (me, 07:47). One upload on Sunday, by hand. No early safety copy.
  • The cover is the cream car (me, morning), to match the car in the game.
  • The hero car is a Meshy model (my pick, "Car 1 (Meshy)"): a 3D car made from a concept image, cream with red stripes. The import kit has three detail levels (about 15,000 triangles up close), separate wheels so they spin, and warm-white headlights. It replaces the blockout box once it is in a test build; it is not in one yet.
Model render, not an in-game shot: the Meshy hero car from the side, cream with red stripes
Model render, not an in-game shot: the Meshy hero car from the side, cream with red stripes
Model render (Blender), not a game screenshot. In-game shots come from the next test build.
  • The dev log stays in my words where it says how it felt: the coordinator agent asks me one question at a time and the agents only fix grammar.

The game director review

A Claude Fable "game director" review scored build 65 at 17 out of 35: Ambition 3, Theme 3, Fun 2, Gameplay 3, Visuals 2, Level 2, Audio 2. Its main point: the cut button is the move, and at the wrap the Director's Cut of your own run should play back what you kept and what you deleted. It ranked ten changes by points per hour: sound for the car and the crash, a grade and best take at the wrap, the crash as a beat instead of a reset, the wrap replay with one deleted scene, and the look pass. A second Fable pass turned that into a plan for 35: plan for 35, expect 30-31, floor 28, with Fun and Gameplay buildable by Friday evening.

What's next

Fix the CLOSE stamp and hold the handheld near 60 frames a second. Import the Meshy hero car and the audio kit. Then the fun block: the crash as a beat, the grade and best take, and the wrap replay. The final video and screenshots come from a later build that has all of that, not from today's.

One thing that surprised us

Our own tests never showed the CLOSE stamp. At first that looked like the tests' fault: they only dodged in the PERFECT band (0.42-0.59 s) or the WIDE band (2.56 s and up). The handheld's next run swept the middle, 0.75 to 2.2 seconds, and CLOSE still never fired. So it is a real bug in the game, and it is on the build agent's list.

AI disclosure

Claude agents (Opus for the coordinator agent and the build agent, Sonnet for helpers, Fable for the reviews) wrote the code, ran the tests and drafted this text. FLUX.1 [schnell] on fal made the concept image the cover is built from; Meshy made the props. Google Gemini text-to-speech, through fal, made draft voice lines for the director (I chose an AI voice at 9:23 AM). No local AI models ran. The music and sound effects are made in code, with no samples. Full list on the game page.

Build/version

Latest tested build when I wrote this: 72 (test build 11): clean except the CLOSE stamp; 39 -> 57 frames a second on the handheld with Nanite off for weak integrated GPUs. Before it: build 65 (test build 10), not clean (window size). At that point the hero car and the audio kit were not in a build yet.

Fri Nov 13 (day 2)

tokens 3.78Boutput 25.9Msaves 11relay blocks 354

What got built: Thursday night and Friday went in numbered builds. B13 to B13c added a curved road, an opening ramp, a Director's Cut take with a clapper slate, wide shots that earn no film, and a cut in on-screen text (I said it was "way too busy with a lot of text"). Friday I added a hold cue: while you hold Space, the screen says what letting go does and what holding on does. Then a stick fix for controllers, a heavier lane change, a setting so any controller can steer, and a log of every key press for the friends who play tonight. The music became two 70s-style funk tracks. After I heard them I said "Use these". In the evening: clearer cut words, a rule that cuts for you if you are still holding at the end, and a water tower that leans while its false front falls.

What broke, and how it was found:

What the agents got wrong, and how it was caught:

The handheld test runs: B13b passed its break-it run (16 launches, no crash, no soft-lock). The friends' build passed all main rows, four screenshots at the real screen size and 34 break-it cases over 41 launches, with no game crash. One case failed: a fast lane tap can be lost for a frame. I chose "Send as is". I played it on the handheld's own controller and said "done stick feels great lock it in": 38 stick lane changes, none ignored. The stick is now locked.

Decisions I made: I kept the AI music, with the credit line on the page. The terms are unclear on one point, and I sent a written question to the platform's support; no answer has come yet. If the answer is no, the game goes back to the code-made loops. I also said "players even me I am making this game need cues and hints in order to understand". The three friends got the build that evening.

Page facts that were wrong: before I sent the page, a helper listed what was really inside the packaged Windows build and compared it with the page text. The page said no AI made the music; the music tracks in the build are AI-generated (ElevenLabs Music, through fal). It also said the Rodin water tower was left out of the game; it is in the build. A free plywood texture (CC0, Poly Haven) was in the build but not on the sourced-assets list. All three were fixed. The page is now checked against the package before it goes up.

The HUD icons: I asked a Claude helper for a set of simple icons as vector code, looked at three versions, and told the hub "go with your v3 pick". Eleven icons ship in the build; no image model drew them.

The on-screen cue rows, from the final capture
The cue rows (lane cues) shown before a shot.
The day entry as filed (2026-11-13)

(Written on Friday Nov 13; later entries say what changed.)

Jump Cut, by jf1shh (solo entry). Theme "Director's Cut". Day 2 of the jam. Times Pacific.

What got built

  • Late Thursday night, three builds went to the handheld. B13 (pacing, a ramp jump, a call sheet and countdown), B13b (a curved Scene 1 road, a re-timed Scene 2, an opening ramp, rumble, and a cut in on-screen text after I said it was "way too busy with a lot of text") and B13c (a Director's Cut take with a clapper slate, wide shots that earn no film, a free "mercy" retake reel, ramp walls, a lighter hero car, and the Meshy water tower). B13b passed the break-it run: 16 launches, no crash, no soft-lock. A test of a cheaper anti-aliasing mode failed narrowly (car seams and far edges crawled), so the game keeps its current one.
  • A hold cue. While you hold Space, the screen now says what releasing does and what holding on does. On a stunt you can drive, releasing is a cut worth +0 and holding on is DRIVE IT. On a must-cut stunt the row goes amber near the end of the window. The agents also found that a must-cut needs a fresh tap after the hold runs out, so the words became "TAP [SPACE] AGAIN TO CUT". Seven small text patches went in, each applied alone with its own tests.
  • New music. The agents generated 70s-style funk tracks with ElevenLabs Music (through fal.ai). I heard two and said "Use these" and "those are cool". They replaced the code-made loops in the game.
  • Candidate builds. From mid-morning the work went in numbered candidates: cand1 (cue words plus music), cand1f (plus an input fix), cand5 to cand8, then cand9b for friends. I also moved the title logo to the top left, asked for it smaller, then asked for the line under it on two rows; I saw a picture of each step.
  • Controller fixes. A setting so any connected controller can steer. A stick fix so a slow push no longer hops lanes many times. I played cand9b on the handheld's own controller and said: "done stick feels great lock it in". The log showed 38 stick lane changes, none ignored. The stick is now locked: nobody changes it without my word.
  • A log for the friends. cand9b records every key press and choice, and on quit it opens a mail draft to me with a short questionnaire. It also got a heavier lane change (0.45 s), which the numbers say is fair.
  • Final-build work started for Saturday: hazard tape on the road, air tricks, bigger crashes, a river, an outdoor desert look, a second camera crew, and 70s-style icons. My rule: polish beats new features.
  • In the evening (save points 110 to 116 on Diversion): clearer cut words, a rule that cuts for you if you are still holding at the end, a lane glow that no longer lights early, and a bolder shot title. The water tower leans and its false front falls.

What broke

  • The Legion refused to take input (about 09:13). The game window would not come to the front and "last input" read zero while the cursor never moved. The stick-drift theory was measured and ruled out. The built-in pad streams packets with no change; which device resets the idle timer was never identified. Keys sent as window messages still reached the game, so tests went on, but they are not a human keyboard.
  • Performance failed its floor: 23.0 frames a second (1% low) in a Development build against a 26.1 floor.
  • The packaged game would not compile (about 18:50). That day's outdoor-light code used an editor-only call in four places. The compile check had only built the editor. It now builds the game too.
  • The save gap (20:15). I asked whether the game was being saved to Diversion and GitHub. The answer was no: the last Diversion save was 05:27 with 116 files unsaved. I said "Yes, push it". A save went in at 20:30 and the build agent's steps now say to save after every passed test run.
  • A stuck build agent. A build agent sat 30 minutes in a hand-off that was not happening, because it believed the line was 300K when it had been moved to 400K. The coordinator agent measured it and messaged it.

What the agents got wrong, and how it was caught

  • A caveat was dropped. The build agent reported a pad-hint defect from a log and wrote that the code cause was not yet read. The coordinator agent told me within a minute that it was confirmed. The build agent read the code and corrected itself six minutes later. The coordinator agent's own note: a caveat someone attaches to their own finding is not yours to drop.
  • "Corners while holding" was killed. A patch would have drawn gold corners during the hold. A coordinator agent read the clock code: the cut window ends first, so the corners would have taught players to release on gold, into the +0 action.
  • The star cap was misnamed all night. Documents called it a retake cap. The code counts crashes. A free crash is not counted; mercy crashes are.
  • Colour predictions were wrong. For the hold cue the coordinator agent wrote predictions before the run. Two colour predictions failed and the build was judged by the code.
  • Break-it reviews tried to break each patch. Break-it reviews of cand9's pieces found two real bugs: the stick fix would have stopped lane changes on worn controllers, and the play log forced a disk write on every key press, which could stutter. Both were fixed and attacked again. Later in the evening, reviews caught a hint that the clean-up had hidden, a trick prompt that would have shown dots, and speed lines that drew over the lanes.
  • Three viewers who had been told nothing watched the first 12 seconds. All three understood hold, cut and drive. None could say how scoring works, and none understood the film counter. Later, three fresh agents read pictures from the build I played and all failed at the same three places: the must-cut cue, a crash with no stated reason, and a mid-air "[E]" prompt nobody had taught. That build did not pass the new-player gate.
  • Offscreen captures show on-screen messages about ten times too briefly. Real players are fine. Trailer footage has to be captured at real speed.
  • The test agent on the Legion refused to launch a build (its safety check) and the coordinator agent could not talk it past that. I typed "yes, launch" myself.

What the Legion runs found

Each test run was a Claude session on the handheld. On cand8, launch and sound passed, any controller could steer, and holding the D-pad moved one lane. The stick problem was found there. cand9b was cleared after all main rows, four screenshots at the real screen size and 34 break-it cases over 41 launches, with no crash and no soft-lock. One case failed (a fast lane tap can be lost for one frame) and I chose "Send as is". On prev1 (my preview), controller rows, title and help screen passed, and lane taps in bends landed 19 of 21.

What I decided

  • "no reason to wait for 9am": the proof freeze lifted at 07:57.
  • "push the friend playtest until tonight": friends were busy until the evening.
  • Every coordinator agent runs Fable 5.2 at max and hands each task to the lowest model that fits. The coordinator hand-off was raised to 500K.
  • Bots with no prior knowledge at three intelligence levels should try to learn the game.
  • "install a sound recorder on the legion": so agents can hear a build. A recorder using Windows' own audio API recorded 115 seconds beside the game. Music ran its full length. Nobody had yet listened to it.
  • Permission for the coordinator agent to close and open test agents on the Legion.
  • Keep the AI music and credit it. The terms are unclear on one point; a question to the platform's support is drafted.
  • Send cand9b as is. All three friends got it that evening. "Players even me I am making this game need cues and hints."
  • Push to GitHub and Diversion for me from now on, and give the Legion its OKs.
  • When the game still has the cut held at the end: "Yes, cut for me".
  • "I want you to hit 30/35", and: "things have to make logical sense....if a car impacts something on the ground it probably shouldnt flip, if its not in the air".

Where it stood at midnight

Save 116 was a working tree, not a player build. The Legion still had prev1. The score estimate was about 21 of 35, and 16 of 35 from a cold red-team reading of pictures of save 116.

AI disclosure

Claude agents (Fable for the coordinator agents and hardest reviews, Opus for the build agent and break-it reviews, Sonnet and Haiku for helpers) wrote the code, ran the tests and drafted this text. The music is AI-made: ElevenLabs Music through fal.ai. Director voice lines are Gemini text-to-speech through fal. No local models ran.

Sat Nov 14 (day 3)

tokens 6.11Boutput 43.8Msaves 61relay blocks 491

From midnight to about 11 AM, the saves ran from 118 to 161. I said "i want you to hit 30/35", and "good night keep working and dont stop no idleness".

What got built: save 118 is about telling the player what to do. Every stunt you can steer past or cut now shows both choices with their buttons for the whole window: the lane action, and under it "or hold to cut". The old cause was that the lane cue turned off once the car was in the safe lane. Save 119 polished the tower. Then, in order: crashes that follow the car's momentum, a dry causeway where the ford was, ground and road looks made in the engine (the build's one sourced texture is Poly Haven plywood, CC0, on the ramp wedges and the set walls you crash into), a rebuilt wrap screen with less text, a new sound mix, a cut grade, a movie replay of your kept cuts, and crash hit-stop and camera kick. The Director's Cut now means one thing: your own cuts, driven. The handheld's own safety check passed on the build I played (47 launches, close-crash 15 of 15, break-it 16 with 0 fail).

My overnight rule, in the words I typed: "things have to make logical sense....if a car impacts something on the ground it probably shouldnt flip, if its not in the air". So a car that hits something on the road stops hard and slides, and only a car in the air tumbles. I chose scoring option B, answered "Yes, grade the cut", and said "No, keep C" to a no-crash B. After playing, I said "textures and materials wil definitely need more polish but great work over night the game is looking much better!" and "that's a wrap end screen still has too much text though".

What broke:

What the agents got wrong, and how it was caught: after each batch of patches, three fresh agents read pictures of the build, with no notes, and answer per shot "what do I do, and which button". An agent that tries to break the patch reads it first. Reading 2 (pictures of save 120) scored 15 of 35 from a cold red-team; before it was 16. Reading 3 failed shot 7 for 2 of 3 readers. Reading 4 showed that "STAY CENTER > RIGHT" reads as do nothing. Reading 6 failed shot 7 for 3 of 3 and shot 3 for 2 of 3. "EARLY DODGE +0" showed up when a player steered safely as told. A cue text said a must-cut was free, and it was not. A looker comparing stills with a patch on and off caught a car floating on a teal lake and ripple rings drawn on dry asphalt. Break-it reviews caught a touch tap on the menu that switched glyphs to keyboard and a latch that kept a bright "STEER RIGHT" after a fast press. Two saves went in over a looker's "STOP", on purpose, because the fault was in the base code and not the patch. One hub counted its own memory 200K too low and was caught by the context guard.

The first friend's log, from the Friday build: 3.5 minutes, both scenes finished, 32 crashes. Early moves crash him. I answered "Clearer cue". Lanes and hazard timing stay locked, and a wait-then-now cue is ordered.

Still open then, at about 11 AM on Saturday (what happened to these is below): the horizon fix and one shortened HUD line were to go in the next batch. The oncoming truck had to pass review or be out. A reading on the newest build, a switch list and the README checks were owed. The wait-then-now cue was not built. A Windows Shipping build had not been run: the Linux compile was running, and Windows was to come after a new handheld package. The plan then was a twin cut at 13:00, a handheld pass from 14:15 to 15:00, and the freeze at 17:00 (the freeze later moved; see below).

From 11 AM to about 12:15 the saves ran 163 to 165, and I stopped adding features. Save 163 fixed the HUD badge at 16:9; small overlaps remain at 4:3 and at 16:10 with the keyboard. Save 164 is the "Clearer cue" I asked for: WAIT, then a bar, then NOW. A looker still said BAD, for four small drawing faults. Save 165 put "DIRECTOR'S CUT" and "N OF M STUNTS CUT" on the wrap card, cut the hold down to one cue, and gave the oncoming truck lamps and a 1.0 s brake. 161 tests pass, 0 fail.

I also added a rule: anything that moves needs a real-life physics reason, and a "makes sense" review runs as a gate. Its first run was still going at 11:26. The handheld's safe gate passed on the test build (33 launches, 0 fail); scene 1 ran 13.81 ms median, scene 2 14.26 ms. I dropped a HUD clamp follow-up because redoing it changed nothing. Left for after the jam: the water tower falls too fast, props jump at a crash, and the front wheels do not steer.

The last four fixes (effects follow slow-motion, the car's pitch eases on the ramp, the light-rig tripod gets casters and a winch, and a polish for the cue faults) were in test then. The final save was due 12:50 and the Shipping cut planned for 13:00, after one Windows check on the final tree.

The final save, 166, went in at 12:34, 16 minutes before its target. Effects now follow slow-motion, the car's pitch eases on the ramp, the drums get a kicker plank, and the cue text is cleaned up. The light-rig tripod's tow line, winch and casters ship switched off: its look said STOP because a line ran across the lanes. A review of the whole diff said ship with fixes and found no code blocker; one fix, a Windows compile of the final tree, is still open. That compile finishes about 12:58, the Shipping cut is expected about 13:05, the handheld pass comes after it, and the freeze is at 5 PM.

The Windows check of save 166 was green at 12:47 and I gave the cut word at 12:48, 12 minutes before the planned 13:00. The Development twin package finished at 12:54 (764.8 MB). The first package named Shipping was really a Development build, because a script variable was lost; comparing file hashes caught it at 13:00. The real Shipping package of save 166 is proven: 562.1 MB, a different exe hash from the Development twin, and the build log says Shipping config. A hand start of it was still owed then. On pictures of 166, first-time readers got 7 of 7 on keyboard, all pictures and gamepad (an earlier gamepad 1 of 7 was a picture-kit artifact). The judge-style red team gave about 19.5 of 35, up from 18, and the main loss is too many words and rules on screen. The physics-sense pass still judges 25 of 61 moving things wrong. A text-only candidate, a four-line HOW TO PLAY card and DIRECTOR'S CUT on the wrap card, was waiting; if it was not ready by 13:40, 166 would ship as it was.

After the cut I allowed one more save, text only, because the judge-style review said the top loss was too many words on screen. Save 167 has a four-line HOW TO PLAY card (the old one was 105 to 135 words) and shows DIRECTOR'S CUT once, on the wrap card, as a key row. Save 166 stays packaged as the fallback, and 167 was to ship only if it read no worse. Three fresh readers per side, card only, scored 9 of 15 on both. Crash went from 0 of 3 to 2 of 3 and Director's Cut from 0 to 1 of 3, but the cut question fell from 3 of 3 to 0 of 3, because the cut cost is no longer on the card. I accepted that for a much shorter card. Both 167 packages are proven (Development twin 764.9 MB, Shipping 562.1 MB with a different exe hash and Shipping config in the build log). The handheld check of 167 was still running then, with a target of 14:20; if it was red or late, 166 would ship.

The 5 PM freeze did not happen. At 14:39 I answered "Yes, polish overnight" to a card that moved the game freeze to Sunday 11:00 PT. The rule became: the last build that passed its handheld gate ships. That evening save 168 fixed the HUD boxes (rounded plates, text that fits, no old text under new); I played it and it played fine. Save 169 failed its picture check at 16:10 and 4:3, where the WAIT and NOW tags vanished at stunt 7, so it was never handed out. Save 170 (cue arrows and a credits fit) went in at 17:38 and passed the handheld gate, and save 171 (the menu logo and the new icon) passed too. Save 172 went in at 19:35 and its full gate passed at about 22:09, including a DIRECTOR'S CUT reload of a cut stunt. At 19:48 I ruled that the submission zip is the Shipping build.

The day entry as filed (2026-11-14)

(Written on Saturday Nov 14; later entries say what changed.)

Jump Cut, by jf1shh (solo entry). Theme "Director's Cut". Day 3 of the jam. Written part-way through the day, from the coordinator agent's checkpoint notes up to about 11:00 PT. Times Pacific. The freeze was planned for 17:00 that day (it later moved to Sunday 11:00 PT).

What got built

Every change below is a save point on Diversion (a working tree, not yet a build handed to players) unless it says otherwise.

  • Cue clarity (saves 117 to 120, 134 to 142, 146 to 153). The main fix of the night. Every optional stunt window now shows two rows: the lane action with its control, and "or hold to cut" with its control. The cause for shots 1, 2 and 6 was that the lane cue turned off once the car was in the safe lane. Later rounds added a step line, a bar that ends at the game's own cut, and "[SPACE] > LET GO" on the shot where that was unclear.
  • Water tower (saves 113, 115, 119). The tower leans and its false front falls. The water, puddle and dust did not appear at first; the cause was a wiring fault (no prop got a valid rig), fixed in save 115. Save 119 polished it.
  • Crashes that follow momentum (saves 124 to 149). My rule: a car that hits something on the ground stops hard and slides; only a car in the air tumbles. Crash fade, crash contact, a ramp wall and a light rig that meet the car.
  • Scoring (saves 144, and a grade after). I picked option B at 05:03: clean dodging can earn +1 (on paper, not yet played), and lanes, stick, grade and hazard timing stay locked. "Yes, grade the cut" unlocked a cut grade. A run with no crashes does not grade B ("No, keep C").
  • Look. Horizon, set dressing and a ford: a dry causeway through a pond (save 141) after the water version failed twice. Ground and road looks made in the engine (the Poly Haven ground and road textures I approved are not in the final build; its one sourced texture is Poly Haven plywood, CC0, on the ramp wedges and the set walls you crash into). A wrap-screen redo with less text.
  • Sound. A new mix, an imported cut whoosh and a hold click.
  • Replay and crash effects (saves 159 and 160). The Director's Cut now has one meaning: your own cuts, driven. A reel plays your kept take. Crash hit-stop, impact burst and camera kick. A credits screen with 13 sourced lines.
  • I played prev2 A (a Development package of save 151) on the handheld. Take 4, 13 points, 5 stars, grade B. On the sound check I answered "Sounded fine".

What broke

  • The wordless box. A prompt plate painted with no words. It is pixel-identical with every switch off, so it is a bug in the base. Saved with the fault named and fixed in a later round.
  • The canal wall material never compiled. A custom shader named a variable line, which is a reserved word. The walls are matte concrete, lighter than expected.
  • Grey horizon cliffs. A material was made from a material that Unreal refuses as a parent, so about two thirds of the cliffs and spires drew the engine's default grey. At that point a fix was cut and reviewed.
  • The ford. Three versions of a wet road failed their looks. The ruling: stop chasing it after two tries. A dry causeway is what shipped.
  • The oncoming truck. Revision 1 slid sideways. Revision 2 still crabbed up to 64 degrees and lurched from 0 to 14 m/s in half a second. Both were rejected. Revision 3 was in review then.
  • The truck-steer pop. The truck body still jumps about 250 pixels in one frame. Three fixes were tried; one was reversed for a failing test and the next one replaced it.
  • Saves 148 and 152 went in over a looker's "STOP". Both times the looker's STOP because the fault was in the base code, not the patch. That rule is written down and used on purpose.
  • A one-frame test line (S1-R LATE) shows up as an old flake: 3 of 3 reruns passed; 0 of 12 logs on recent trees.

What the agents got wrong, and how it was caught

  • Readings. Fresh agents who have never seen the game read pictures of each tree and say, per shot, what to do and which button. Reading 2 (tree 120): the red-team gave 15 of 35 (16 before on tree 116). Reading 3 (tree 133): shot 7 failed for 2 of 3 readers. Reading 4 (tree 137): the cue fix helped shots 5 and 7 but hurt shot 3, because "STAY CENTER > RIGHT" reads as do nothing. Reading 6 (tree 148): shot 7 failed 3 of 3 and shot 3 failed 2 of 3, and "NEXT: RIGHT [D greyed]" read as "press now". Each reading went back to a patch.
  • A cue that was wrong on the safe path. "EARLY DODGE +0" appeared when a player steered safely as told.
  • Looks. A looker checks stills with a patch on and off. These caught a car floating on a teal lake, grey planks, ripple rings drawn on dry asphalt, mesas that read as buried tanks, and a HUD line about 110 px too long.
  • Break-it reviews found a blocker where a touch tap on the menu switched glyphs to keyboard, a latch that kept a bright "STEER RIGHT" after a fast press, and a rule that would have made the clean point pay before the cue showed.
  • A cap that was wrong. A shunt cap of 150 cm failed a test that expected 163. The coordinator agent ruled the cap was wrong, not the test; the cap went to 170. A fix that changes what a test expects is the author's call.
  • A cue text that was untrue. "A MUST CUT IS FREE" was shown, and reading 7 found it untrue. It went into the next HUD round (WP-HUD-15).
  • A coordinator agent miscounted its own memory by about 200K and was caught by the context guard. The next hand-off is planned at 420K, not by feel. Coordinator agents have now changed hands about every hour.
  • A conflict was left unreconciled between two notes about whether prev3 holds one more patch than save 158.

What the Legion runs found

Test agent 25 ran the prev2 A safe gate: 47 launches, all exit 0; close-crash 15 of 15; break-it 16 with 0 fail; soft-lock pass; the hold click present; performance inconclusive. That was weighed as a pass for my sound check only. A later perf check of prev3 against prev2 A: scene 1 median 16.27 ms against 16.09 (budget +0.3), scene 2 14.49 against 14.90. prev3's safe gate passed too (close-crash 15 of 15, 34 launches). The first of the three playtesters' logs from cand9b: 3.5 minutes, both scenes finished, 32 crashes. Early moves crash him. The log also showed two missing fonts.

What I decided

  • "i want you to hit 30/35" and "good night keep working and dont stop no idleness".
  • Scoring option B. "Yes, grade the cut". "No, keep C" on a no-crash B. "Yes, shrink it" for the truck's crash box.
  • "Not yet" on widening shot 7's timing by 0.15 s. Hazard timing stays locked.
  • "Keep them" on the five old test shortcuts. One shortcut, Jump Cut TEST, from now on.
  • On the first playtester's log: "Clearer cue". Lanes and timing stay locked. A wait-then-now cue was ordered.
  • Questions come one at a time on a card.
  • Feedback in my words: "textures and materials wil definitely need more polish but great work over night the game is looking much better!" and "that's a wrap end screen still has too much text though".
  • I asked for a replay of the whole run with its effects, "the action cut reel", and a clearer meaning for the Director's Cut.
  • Six UI/UX links were handed over as reference only; nothing installed.

Still open then (about 11:00 PT)

  • The horizon fix and a shortened HUD line were to go in the next batch. The oncoming-truck revision 3 had to pass review by 11:35 or be out.
  • A reading 8 on the new tree, a switch list, and README checks.
  • The wait-then-now cue was ordered, not built.
  • A Windows Shipping build had not been run. The Linux Shipping compile was running; the Windows one was to come after a new Legion package.
  • The plan: a twin cut at 13:00, a Legion pass 14:15 to 15:00, and the freeze at 17:00.
  • No gameplay video or final screenshots existed yet. My own words for Sunday were not written.
  • The question to the platform about the AI music was drafted, not confirmed sent then.

11:00 to 12:15 PT: saves 163 to 165, then the last four fixes

Fixes began at 11:31, about 40 minutes ahead of the plan. Feature work stopped; only fixes go in from here.

  • Save 163 (about 11:36). The HUD badge no longer overlaps at 16:9. Small overlaps remain at 4:3 and at 16:10 with the keyboard. A follow-up clamp was dropped because its re-cut changed nothing.
  • Save 164 (about 11:53). The cue I asked for ("Clearer cue") is in: WAIT, then a bar, then NOW. The looker said BAD with no STOP; four small drawing faults are left, and a polish patch is queued for them.
  • Save 165 (about 12:14), chain A. The wrap card now says "DIRECTOR'S CUT" and "N OF M STUNTS CUT". The cut hold shows one cue (look PASS). The oncoming truck has lamps on its lenses and brakes over 1.0 s; the brake is smooth, the looker said BAD with no STOP. Tests: 161 pass, 0 fail.

The "makes sense" review is a new gate, from my rule that anything that moves needs a real-life physics reason. Its first run was in progress at 11:26. Its verdict is not in the notes I had.

The Legion gate: the test build (tree 161) passed the safe gate, 33 launches with 0 fail. Perf: scene 1 13.81 ms median and 36.5 fps at the 1% low; scene 2 14.26 ms and 32.8 fps. The test shortcut has pointed at that build since 11:46. The replay reel is not verified yet.

Dropped: the HUD clamp follow-up. Left for after the jam: the water tower falls too fast, props jump at a crash, the front wheels do not steer.

Each patch got a break-it review by a fresh agent before its go. They found real faults, all fixed first: a wrong count on the wrap card, a last-frame timing fault on the cut cue, a sideways skid with a 1.5 s truck brake, and lights dropping under the road on landing. I answered the grade card "Keep as is", so scoring and grades stay locked.

In test then, chain B, the last four fixes (started 12:13):

  • effects follow the world's slow-motion;
  • the car's pitch eases on the ramp;
  • the light-rig tripod gets casters, a tow line and a winch, and the drums get a kicker plank;
  • the polish pass for the four cue faults.

The final save was due 12:50 and the Shipping cut planned for 13:00. Chains A and B were saved without a Windows compile because the Shipping compile holds the handheld, so one Windows check on the final tree is required before the cut. A review of the whole diff is due 12:40.

12:34 PT: final save 166

The final save, 166, went in at 12:34, 16 minutes ahead of its 12:50 target. The build label reads "SHIP (dv.commit.166, 2026-11-14 PT)". Tests: 163 pass, 0 fail; lint 0.

What the last four fixes change for the player:

  • Effects follow the world's slow-motion.
  • The car's pitch eases on the ramp (look PASS).
  • The drums get a kicker plank.
  • The cue polish fixes the four drawing faults from save 164.

One part ships switched off: the light-rig tripod's tow line, winch and casters. Its look said STOP because a line ran across the lanes. The drum kicker stays on.

The other looks: the cue polish, the effects and the drum kicker were each BAD with no STOP. One WAIT pill runs past the panel edge on a gamepad. The slow-down cannot be read from still pictures, and the crash-hit pair was never captured.

The review of the whole diff (saves 156 to 165 plus chain B) came back SHIP-WITH-FIXES, with no code blocker. The build label fix is done. The other fix, a Windows compile of the final tree, is pending.

Next, as planned then: the Windows compile of the final tree ends about 12:58, then the game and editor check on 166. The Shipping cut is expected about 13:05 (planned 13:00). Then the test-machine safe gate (target 15:00), a new-player reading and a physics pass on pictures of 166, and the freeze at 17:00.

13:08 PT: the cut

The Windows check of save 166 came back green at 12:47, and the cut word went out at 12:48, 12 minutes ahead of the 13:00 plan.

Packages:

  • The Development twin package finished at 12:54 and is good: 764.8 MB, with the fixed concrete texture in the cook and the build label line in its BUILD.txt.
  • The first package named Shipping was really a Development build, because a script variable was lost. I found it at 13:00 by comparing file hashes.
  • The real Shipping package of save 166 is now proven: 562.1 MB, its exe hash differs from the Development twin's, and the build log says Shipping config. The board row "Shipping build cut" is ticked. A hand start of it was still owed then.

Reviews of 166:

  • First-time readers: keyboard 7 of 7, all-pictures 7 of 7, gamepad 7 of 7 on real pad-mode pictures. The earlier gamepad score of 1 of 7 came from the picture kit, not the game.
  • The judge-style red team gave about 19.5 of 35, up from 18. The main loss is too many words and rules on screen.
  • The physics-sense pass still judges 25 of 61 moving things wrong. Three have been fixed since the first run.

Paperwork: the judges' README has no open checks (4 are left for the handheld). The pitch is approved, with a Director's Cut sentence. The switch list is at r3. The caveat ledger closed 215 stale items, and a sample check found none wrong. The earlier devlog pull request, 209, merged at 12:47.

One more change is a candidate, not saved: chain C, a four-line HOW TO PLAY card and DIRECTOR'S CUT shown once on the wrap card. Both passed an adversarial review. It starts after the Shipping package and becomes save 167 only on the coordinator agent's word. If it is not ready by 13:40 it is dropped and 166 ships. A third part, showing the step keys early, was dropped after review.

Next, as planned then: the hand start of the Shipping package, the handheld's safe gate (target 15:00), my play with sound on, the video capture, and the freeze at 17:00.

13:34 PT: save 167, the shorter cards

I allowed one more save after the cut, and only a text-only one. The judge-style review of 166 said its top loss was too many words on screen, and the cards were the cheapest place to fix that. Save 166 stays packaged as the fallback, and the rule I set at 13:21 is that 167 ships only if it reads no worse than 166.

Save 167 went in at 13:20. The build label reads "SHIP (dv.commit.167, 2026-11-14 PT)". Before the save: Linux compile OK, tests 163 pass, 0 fail, lint 0. The Windows check of 167 was green at 13:23.

What changed:

  • The HOW TO PLAY card is now four rows. The old one ran 105 to 135 words, and all three readers called it too much.
  • DIRECTOR'S CUT is shown once, on the wrap card, as a key row.

The look at the wrap card said it is the same as 166: a road stripe shows through the translucent panel in both. In 167 it sits under the DIRECTOR'S CUT row like an underline and hides no letter. The wrap card went from 29 words to 36. The look at the how-to card was BAD but not a STOP: at 4:3 only, a row wraps to two lines and stays readable.

The reading, with three fresh readers per side, card only, same questions:

Question166 cards167 cards
Steer3 of 33 of 3
Cut3 of 30 of 3
Crash0 of 32 of 3
Director's Cut0 of 31 of 3
Wrap3 of 33 of 3
Total9 of 159 of 15

It is a tie, 9 of 15 each. Crash and Director's Cut got better. Cut got worse: the cut cost is no longer on the card, and no reader found it. I accepted that to get a much shorter card. Two other caveats: the 4:3 how-to row is cramped, and the title tagline overlaps the clapper prop at 16:10 (166 has that too).

Packages, both proven:

  • The Development twin finished at 13:28 and is good: 764.9 MB.
  • The Shipping package finished at 13:34: 562.1 MB, and its exe hash, name and size all differ from the twin's. The build log shows Shipping config once and Development config zero times. Both BUILD.txt files end with the 167 label. The fixed concrete texture is in both cooks (882 paths each). The Shipping folder has no marker file and no log file; the twin has its playtest log by design. The 166 folders were not touched.

Still running: the handheld check of 167, started at 13:47 with its target at 14:20. It covers a full run on keyboard and pad, both cards, the wrap card, a hand start of the Shipping exe and four README checks. 167 is not cleared until it passes and I have played the build. If it is red or late, 167 is dropped and 166 ships.

Sun Nov 15 (day 4, to 5:59 PM PT)

tokens 732.0Moutput 5.5Msaves 6relay blocks 168

"The last day I didn't sleep much at all, and the usage limit was a real concern, so I made adjustments and many things had to be cut, which I believe is very common in the game development process. But I was also very happy with what I built, and I will continue to work on it in my own free time."

The THAT'S A WRAP card, from the final capture
The wrap card over the dimmed scene.
The day entry as filed (2026-11-15)

(Written on Sunday Nov 15; later entries say what changed.)

Jump Cut, by jf1shh (solo entry). Written at 03:44 PT, from the coordinator agent's checkpoint notes.

What got built

  • HUD without dark boxes (save 182). The HUD now runs at icon level 2 with no dark plates. The Development build of save 180 asserted under a button mash (RenderBatch.NumIndices > 0 on the wrap card). Save 182 adds one draw-only fix: a 1x1 pixel ghost box on the lowest HUD layer. The cause was read from the engine source.
  • Checks. Development 182 survived the mash hammer 3 of 3. The Shipping 182 build was cut at 03:22 and passed its full gate at 03:42 (hashes, mash hammer, one keyboard and one gamepad run to the wrap card with a deliberate miss and a retake, settings page). An adversarial review found no blocker and kept four caveats.
  • Ship candidate. Save 182. Fallbacks in order: 180, 181, 175, 172.

The miss

  • The film-set backlot props were never in any build. The game's own log says built instances=0. Their mesh files were never saved into the project. Textures and material instances were saved, so the old check, which only counted files, passed.
  • The cause: the import script deleted each mesh right after saving it, and still logged OK.
  • The fix is in the import script only. No game code changed. From now on the proof is the game's own log saying it built the props.
  • Save 183 is one try at adding the props. It must be saved by 07:00 or it is dropped. Save 182 stays the ship candidate until then.

Save 183: the props finally place (04:06 PT)

  • The film-set backlot props never drew in any build, not even in save 182. Two causes: the import script deleted each mesh after saving it, and the placer's sand-reach sample pinned the band empty.
  • Save 183 fixes both: the 19 meshes are back, and one small code change to the sand-reach sample. The game's own log now says built instances=246, with no missing types.
  • The Legion test for save 183 was running. No result was in yet (it passed; see 04:57 below). Save 182 stayed the ship candidate until it passed.
  • An art kit of 28 pictures was drawn in code for the page and the video. It is not on the live page yet.

Save 183 passes on the handheld (04:57 PT)

  • Save 183 passed its one Legion test at about 04:55 PT. The Development and Shipping builds both ran clean: hammer x3 each, no crash, no dialog.
  • Keyboard and gamepad hand runs both passed, each with a deliberate miss, a retake and the wrap card. 65 stills at 1920x1200.
  • The game's log on the Legion now places 246 props with no missing types. Save 182 placed none.
  • The cost is speed. Stage 1 ran at 50.4 fps against 50.9 for save 182. Stage 2 ran at 45.9 fps against 49.9. That is about 1 ms a frame (5 %) slower.
  • The picture check said GO. A drum and a camera dolly far out on the sand look tilted about 10 degrees. One far trailer group is too far away to judge.
  • Ship order is now 183, 182, 180, 181, 175, 172. Save 183 is the ship candidate.

Plan for the rest of the day (as of 10:44)

Freeze 11:00, video capture on the ship candidate, page public 16:30, submit 17:11, deadline 17:59 (all PT).

Freeze: save 184 ships (10:44 PT, Sun 2026-11-15)

  • Nothing was in flight. The game was frozen on save 184.
  • Two builds ship. The Windows zip (458 MB) is the entry. The Linux zip (322 MB) is a bonus download. I uploaded both to the private itch draft and said: "Yes, both saved".
  • Linux gate: pass, with one gap. The next-scene button could not be pressed on the shipping build, so scene 2 was reached only on the development build of the same export. The page says "lightly tested".
  • Glass Box page. It is animated and builds clean. The privacy scan found nothing. Nine real pictures are in. Nine draft markers were still open then.
  • Media kit. Eleven AI concept pictures of future ideas, not the game. Nothing was public yet at that point.
  • AI tools used for the concept pictures: Google Gemini (Nano Banana), ChatGPT image creation on the web, and Grok. They are disclosed as concept art only.
  • Itch draft. The description is set. Still to do by hand then: cover, six screenshots, banner, the three bracket notes, genre, tags, video link, the AI boxes, Public, and Submit.
  • Still ahead then (Pacific): video on YouTube before 15:00, page public at 16:30, dev log and pre-jam post 16:25 to 16:50, Submit at 17:11, hard stop at 17:59.
SCENE 07disclosure, tool by tool

Every AI tool, and what it made

Claude

(Anthropic), through Claude Code: wrote the game's C++ and Blueprint setup, drove the Unreal Editor through Epic's Unreal MCP plugin, wrote the scripts that make the 40 code-made sound effects, drew the 11 HUD icons as vector (SVG) code through a Claude Sonnet helper, drafted design specs (the look guide, Scenes 2 and 3, stunt timings) for the entrant's review, ran the tests and the builds, and drafted this page and ...

In the shipped game: See the disclosure

Epic's Unreal MCP plugin

the bridge the agent used to work in the editor.

In the shipped game: Editor only: compiled out of the shipped game

FLUX.1 [schnell] on fal

concept images for the sets, props and hero car, plus mood frames for the look guide, and one image that ships in the game: the city-skyline backdrop texture (edited afterwards by a script, no AI). The cover and banner are not made from these: they are the code-drawn kit art named below (an earlier draft cover used one concept and is no longer shown).

In the shipped game: 1 image ships (a skyline backdrop); the rest is labelled concept art

Meshy

(the jam's Meshy Pro partner credits): turned concept images into 3D models: the hero car you drive, and props for the backlot sets. Five are placed in the game's two scenes: a studio light, a barrel, a water tower, a prop truck and a camera crane.

In the shipped game: Hero car + 6 prop meshes made; 5 props placed in the two scenes

Hyper3D Rodin

(Gen-2.5): made 3D set-dressing props that are in the game: the crew trailer, the water truck and the wind machine (placed in save 184). There is no Rodin water tower. The Rodin props were made from concept images during the jam, like the Meshy props.

In the shipped game: 3 props placed (crew trailer, water truck, wind machine)

Blender 5.2.3

(local tool, no AI): finished the AI meshes (decimate, bake, LODs) and built a few set pieces by script (crate stack, drum pile, scaffold, cable run, fence, telegraph pole, rock, sage); the hero car's wheels are Blender-made. Its output ships.

In the shipped game: No AI: finishes the AI meshes and builds a few set pieces

PBRgen

(Pro web app, app.pbrgen.com): generated PBR material sets (colour, normal, ORM) for the road, sand, props and car. The sets wired by WP-W2-C ship in save 184 as the default look (switch jc.Look.W2Pbr = 1): car (paint, chrome trim, tyre rubber), ground (road asphalt, dirt sand, wet river sand) and props (oil drum metal, plywood walls, steel).

In the shipped game: Material sets for car, ground and props (default look)

GPT image

(through the Codex CLI): tried once for concept images; it produced no images, and nothing from it is used.

In the shipped game: 0: it produced nothing, nothing is used

Cursor

(its built-in image tool; it reports no model name): before the window, made the 17 pre-jam reference pictures for the prep pack. Each is marked "PRE-JAM REFERENCE, NOT A GAME ASSET" across the bottom; none is in the game or on this page.

In the shipped game: 0 in the game: 17 pre-jam reference pictures, each labelled

Google Gemini text-to-speech, through fal

(fal-ai/gemini-tts, model gemini-2.5-pro-tts, voice "Algenib"), voiced the director's 17 short lines in the game, such as "Rolling... ACTION!" and "That's a wrap!", and one trailer line ("You direct the cut.").

In the shipped game: 17 director voice lines

Music generated with ElevenLabs Music

(via fal.ai; fal-ai/elevenlabs/music): the drive loop and the wrap. Claude Code wrote the scripts that make the 40 sound effects in code (numpy and standard Python; no AI model, no samples). The 4 interface sounds are Kenney's Interface Sounds pack (CC0), a sourced asset.

In the shipped game: 2 music tracks

The 11 HUD icons

written as SVG code by Claude (a Sonnet helper) and rasterised; no image model made them.

In the shipped game: 11 icons ship in the build; no image model drew them

The 63 shipped sounds by origin, as stated in the AI disclosure; not re-counted by this page.

What AI made and what it did not (from the draft)

What AI made:

Model render, not an in-game shot: the Meshy hero car from the side, cream with red stripes
Model render, not an in-game shot: the Meshy hero car from the side, cream with red stripes
Model render (Blender), not a game screenshot. In-game shots come from the next test build.
Model render, not an in-game shot: the Meshy hero car from behind, at the game's chase-camera angle
Model render, not an in-game shot: the Meshy hero car from behind, at the game's chase-camera angle
Same car from the game's chase-camera angle, in the game's lighting. Also a Blender model render.
The Water Tower shot with the backlot props, from the final capture
The water tower shot, where the tower must be cut.
Shot 7 on the backlot, from the final capture
Shot 7 on the backlot, with the props on the sand.

What AI did not make:

Concept pictures:

Tried and left out:

A correction: an earlier version of my notes said Rodin made the prop meshes. That was wrong. Meshy made six of them, and the three Rodin props (crew trailer, water truck, wind machine) are in the game. There is no Rodin water tower.

SCENE 08one person, a team of agents

Where the human decided, where the agents did

I picked the concept, the look, what ships and what is cut, played the game by hand and submit it. The agents wrote the code, ran the tests and made rulings on the way. Relay blocks naming me: 312. Blocks recording a coordinator-agent ruling: 75. Of 1013 blocks.

I decided

  • Nov 11 In between, the agents wrote idea notes against the theme and I picked one.
  • Nov 11 At 1:31 PM I said "Yes, lock it" and Jump Cut became the design doc.
  • Nov 11 The decision I made: "Yes, lock it." Three concepts were scored, then a review tried to break them.
  • Nov 12 The decision I made: No more posting to itch until the very end.
  • Nov 13 B13 to B13c added a curved road, an opening ramp, a Director's Cut take with a clapper slate, wide shots that earn no film, and a cut in on-screen text (I said it was "way too busy with a lot of text").
  • Nov 13 At 20:15 I asked whether the game was being saved to Diversion and GitHub.
  • Nov 13 I said "Yes, push it", and a save went in at 20:30.
  • Nov 13 The hub could not talk it past that; I typed the OK myself.
  • Nov 13 Decisions I made: I kept the AI music, with the credit line on the page.
  • Nov 13 The HUD icons: I asked a Claude helper for a set of simple icons as vector code, looked at three versions, and told the hub "go with your v3 pick".
  • Nov 14 I said "i want you to hit 30/35", and "good night keep working and dont stop no idleness".
  • Nov 14 My overnight rule, in the words I typed: "things have to make logical sense....if a car impacts something on the ground it probably shouldnt flip, if its not in the air".
  • Nov 14 I chose scoring option B, answered "Yes, grade the cut", and said "No, keep C" to a no-crash B.
  • Nov 14 After playing, I said "textures and materials wil definitely need more polish but great work over night the game is looking much better!" and "that's a wrap end screen still has too much text though".
  • Nov 14 Save 164 is the "Clearer cue" I asked for: WAIT, then a bar, then NOW.
  • Nov 14 I also added a rule: anything that moves needs a real-life physics reason, and a "makes sense" review runs as a gate.
  • Nov 14 I dropped a HUD clamp follow-up because redoing it changed nothing.
  • Nov 14 After the cut I allowed one more save, text only, because the judge-style review said the top loss was too many words on screen.
  • Nov 14 At 14:39 I answered "Yes, polish overnight" to a card that moved the game freeze to Sunday 11:00 PT.
  • Nov 15 08:40: I said the credits in the game menu looked out of date.
  • Nov 15 I said yes at 08:47, on one rule: any failure, or no gate pass by 10:40, and 183 ships.
  • What AI made and what it didn' I picked it from the candidate cars on Nov 12.
  • The TV board On Nov 13 at 12:47 PM PT I said "tv board hasnt been staying up to date".

The agents ruled or caught

  • Nov 13 The agents ruled out stick drift by measurement.
  • Nov 13 A hub read the clock code first: the cut window ends before the gold, so players would have been taught to release on gold, into the action worth +0.
  • Nov 14 The wet ford failed its look three times, so the ruling was to stop after two tries and ship a dry road through a pond.
  • Nov 14 What the agents got wrong, and how it was caught: after each batch of patches, three fresh agents read pictures of the build, with no notes, and answer per shot "what do I do, and which button".
  • Nov 14 A looker comparing stills with a patch on and off caught a car floating on a teal lake and ripple rings drawn on dry asphalt.
  • Nov 14 Break-it reviews caught a touch tap on the menu that switched glyphs to keyboard and a latch that kept a bright "STEER RIGHT" after a fast press.
  • Nov 14 One hub counted its own memory 200K too low and was caught by the context guard.
  • Nov 14 The first package named Shipping was really a Development build, because a script variable was lost; comparing file hashes caught it at 13:00.
  • Nov 14 At 19:48 I ruled that the submission zip is the Shipping build.
SCENE 09the misses as plainly as the wins

What broke, and what it cost

Every failure was logged. Three cost the most: the Windows-only compile error that Linux checks missed (build B8, 78 minutes), the morning the hub-orders file was found empty (Nov 13, 06:43), and the plan-usage crunch (plan 85% used on Nov 13 evening, 87% on Nov 14, then a reset and the saving rules: a conservative mode and early hand-overs, logged at 218K to 251K tokens of context instead of near 400K).

SCENE 9Bin his words

What was hard, and what I'd do differently

The hardest part is always thinking I should be doing more, or orchestrating better. The AI agent playtest surprised me by memorizing the first level.

I would streamline my workflows and try to optimize my token spending more efficiently.

SCENE 9Cthe status board

The TV board

Through the jam a big-screen status board ran on the living-room TV, so I could see where things stood without opening a window. It showed the countdown to the deadline, the theme, a "now / next / build" box, a milestone list and, for part of the jam, which agents were working. The board read one data file with the theme, the current line, the next line, the build line, 26 milestones from "Kickoff + theme reveal" to "Deadline", and the work streams. The coordinator agent updated it with a small script that costs no tokens. It writes the now, next and build lines (at most 26 characters each, so they fit the screen), ticks a milestone, and then takes a screenshot of the TV to check that the board really changed. At the end 25 of the 26 milestones were ticked. I kept 182 screenshots of the board in all.

The TV board after the submit (screenshot, Nov 15)
The TV board after the submit (screenshot, Nov 15)

It was not always up to date. On Nov 13 at 12:47 PM PT I said "tv board hasnt been staying up to date". Its data had not been written since 8:42 AM, because the hand-over notes had dropped that step. After that, updating the board became a rule at every take-over and every milestone. On the last day the script's connection to the TV dropped once at the end and came back after a reconnect, and the board data was updated throughout.

SCENE 10read this before trusting a number

How these numbers were made, and what they do not cover

One command, python3 build.py, reads the logs, saves, hand-over notes and the written drafts and rebuilds this page, with zero model calls. Only counts, sums, clocks, model names, role names, save numbers and short header words leave the logs: no prompts, no message text, no session ids.

1675 session files read, 92,487 messages counted, 88,920 repeat copies dropped (resumed sessions repeat history). First message counted 2026-11-11 00:04, last 2026-11-15 18:41 PT. Counted by role (sessions with usage): helpers 1482, hubs 141, leads 52, test handheld 211. The test handheld is counted separately, on that machine: 12,109 messages in 211 transcript files with usage, from 2026-11-12 05:47 to 2026-11-15 18:47 UTC; sums by day (PST) and by model only, matching the total. Before the window opened (Nov 11 until 9 PM PT) the jam sessions logged 792,326,185 tokens of prep, shown nowhere above and not in the headline.

What the numbers do not cover

Where each number comes from

The pictures on this page are embedded in the file itself as small JPEGs, so it stays one offline file with no outside links; larger screenshots are on the itch game page.