We build an isometric game with AI agents and open-source our tools.

1. 

   Plan

   Real rules become data: traffic keeps left, a 4 m median, a railway 10.4 m above the road.

2. 

   Blocks

   Every object is a box in world units, checked before anything is drawn.

3. 

   Flat render

   Colors and readable shapes on the same camera: the color template the paint must follow.

4. 

   Painted, twice

   Two paint passes keep every object in place, yet change the details and even the style. That drift is why the layout file, not an image, is the source of truth.

Swipe to see every pass

1. A drawing, then a rig

   We started from both ends: a picture with the right look, and a skeleton that can move.

   

   

2. Code as a guide

   Next, the rig guided the image model. Same model, same prompt, two runs each. We measured how closely the result follows the guide's shape, how often the arms stay clear of the body so code can cut them out, and thigh width against height, which is 0.134 in the design.

   

   Color template

   • Follows the shape 87–90%
   • Arms clear of the body 53–61%
   • Thigh to height 0.10

   Follows the shape best and keeps the arms clear for cutting, but it also copies the rig's slimmer legs.

   

   Stick figure

   • Follows the shape 64–67%
   • Arms clear of the body 0–35%
   • Thigh to height 0.14

   Legs come out right, but the arms stick to the body and the dashed guides end up painted into the picture.

   no template

   

   No template

   • Follows the shape 75%
   • Arms clear of the body 37–41%
   • Thigh to height 0.12

   Closest to the design's look, with no control over the pose.

   What we learned. A guide controls the pose, but the model also copies the rig's proportions, and stick-figure lines leak into the picture. ControlNet would lock the shape, but it needs open models on a GPU of our own; the image APIs we use take reference images and text only. So code stopped drawing guides for her look.

3. One master, edited kept

   We approved one master image in a neutral pose. Every new image is an edit of that master, never of the previous edit, so small changes cannot pile up.

   

   

   What holds, what drifts

   Face and proportions hold. Hair and shirt follow the new pose on their own.

   Small details, like the seams of the shorts, can still change between edits.

   The key poses become the first and last frames for video.

4. Video between two key poses tested

   Video models fill in the motion between two frames we have already approved: the master as the first frame, a key pose as the last.

   • Pose change. Seedance 2.0 Fast lands on the last frame with a 95% shape match, feet sliding 1% of her height. Kling 3.0 Std: 94% and 2%. Veo 3.1 Lite ignores the last frame: 64% and 9%.
   • Dance loop. With the first frame repeated as the last, Seedance closes the loop. Veo 3.1 Lite does not come back (76%).
   • 360° turn. One Veo 3.1 Lite clip gives eight matching directions; back, hair, shirt and sunglasses hold.

   First and last frames are the strongest control we have, so the model choice follows the job.

5. Code keeps the video on model

   Video models drift in color, zoom in a little, slide the feet and repaint what should stay still. Each of those errors is something code can measure and undo.

   Color drift and sliding feet

   Over four seconds the model pulls her colors away from the master and lets her feet wander. Code finds her feet in every frame and pins them to one point. Then it pulls her colors, and the background's, back toward the master.

   Raw videoAfter code

   16.6 → 3.9color distance from the master, median per frame

   1.3% → 0.8%largest foot slide, as a share of her height

   Color distance from the master, per frame

   Raw videoAfter code

   Show the numbers

   Time (s)	Raw video	After code
   0.00	4.68	1.83
   0.08	10.68	3.53
   0.17	16.11	3.83
   0.25	15.85	3.79
   0.33	16.23	3.97
   0.42	18.34	4.43
   0.50	18.68	4.47
   0.58	17.55	4.33
   0.67	17.6	4.09
   0.75	15.99	3.54
   0.83	15.68	1.99
   0.92	12.32	1.94
   1.00	12.47	1.98
   1.08	15.74	3.93
   1.17	16.64	4.19
   1.25	16.52	4.31
   1.33	16.92	4.25
   1.42	18.33	4.47
   1.50	18.67	4.59
   1.58	17.55	4.36
   1.67	17.72	3.91
   1.75	15.83	4.17
   1.83	15.85	2.01
   1.92	11.78	1.93
   2.00	14.49	1.86
   2.08	15.73	4.07
   2.17	16.42	4.2
   2.25	16.39	4.35
   2.33	17.14	4.47
   2.42	17.72	4.43
   2.50	18.27	4.38
   2.58	18.52	4.68
   2.67	17.28	4.5
   2.75	15.9	2.08
   2.83	16.02	2.15
   2.92	12.76	3.15
   3.00	13.09	3.21
   3.08	13.18	2.1
   3.17	13.29	2.09
   3.25	17.25	4.15
   3.33	18.01	4.44
   3.42	18.6	4.33
   3.50	19.03	4.67
   3.58	18.87	4.66
   3.67	18.09	4.44
   3.75	17.01	4.28
   3.83	13.42	3.95
   3.92	6.47	2.04
   4.00	5.66	4.16

   A slight zoom, and a background that should stay still

   We asked for leaves, flowers, clouds and water to move, and nothing else. The model also zoomed in 1.7%, shifted the frame by 7 × 9 px and repainted every pixel. The zoom and the shift never change during the clip, so code measures them once and undoes them. Then it finds what really moves, about 15% of the frame, matches its colors to the painting and pastes the original back everywhere else.

   Raw clipRawAfter code

   • Raw clip 19.4
   • Zoom and shift undone 11.5
   • Still areas pasted back 0.07

   Mean difference from the painting in areas that should not move, per pixel, 0–255.

6. Sprites

   Frames picked from the fixed clips become sprites. What still needs our eyes: expressions that change on their own, crossed legs, and the gap between the legs when we cut her out.

   

   A dance loop: one beat, 8 frames

   One second of the fixed loop, every third frame.

   

   8 directions from one turn

   One 360° clip, so every angle matches.

Six image models, one set of tasks

Same prompt, references and size for each model, one sample each, so take single results with a grain of salt. No model won every task, so we pick a model per job.

New masterPose editStick figureNight

A new master from the design and a written pose. All six look right. None put the head at 10% of the frame as asked, so framing is left to code.

Trial E1: can an agent run Godot?

A builder agent loaded IsoBlock's sample scene into Godot 4.7.1. The maintainer agent ran it again in another container and matched every number.

• 0.49 pxlargest gap from IsoBlock's renders, 5 scenes, 213 objects
• Identicalrenders, byte for byte, across runs and sessions
• 0.00046largest error on 13 golden projection vectors
• 28.3 MBdebug Android build, made without Gradle
• 10 MBgzipped web engine, single-threaded build
• 2.3 sone edit, run and compare loop

A 2D light reading a normal map worked too. We chose Godot 4.7 with GDScript. IsoBlock stays outside the engine, and a small GDScript loader reads its files.

Popular routes we skipped

1. Spec first

   A maintainer agent (Claude) writes the specification and the expected test results.

2. Build

   A builder agent (Claude Code) implements one stage on its own branch.

3. Review

   The maintainer checks the branch's behavior against the expected results.

4. Merge

   One of us reads the review and merges the stage.

Notes from shipped games

Studios solved several of our problems with structure long before image models. We borrow their ideas, not their budgets.