A Sunday afternoon question
Jens opens a chat: a game like Warcraft III: The Frozen Throne, but for the phone. Claude researches what made the game so special back then – heroes, the World Editor, the custom maps, Battle.net.
Every day since the first idea. The gauge keeps count as you scroll: commits, shifts, finished packages and the estimated time Jens put in himself.
Jens opens a chat: a game like Warcraft III: The Frozen Throne, but for the phone. Claude researches what made the game so special back then – heroes, the World Editor, the custom maps, Battle.net.
Six minutes later the course correction: they'd done tower defense plenty of times. Jens remembers a mode where you take one of twenty heroes and conquer a huge map, three of you in co-op or against each other. That becomes Titan Slayer.
Claude's working title was “Splitterkrone” (Shard Crown). Jens names the game Titan Slayer and the world Titan.
After two days of planning and deep research Jens puts the first commit into the repository: a five-part concept, control panel, 14 agent roles, skills, safety hooks and around 110 work packages with acceptance criteria. Everything after that is done by the AI.
Weekly reset moved to Saturday 13:00, shifts start seven minutes past the hour.
The first autonomous shift starts and takes over the skeleton from Star Slayer. By evening the Cloudflare Access login works and the first build is live.

The art pipeline is ready, but the cloud environment won't let it reach OpenAI. 35 minutes later the network block is lifted.
Out of three variants Jens picks the warm, classic style. Plus the plan to rebuild an old Warcraft III map as the first world.

Version 0.1.0, tested alone on the iPhone 16. Everything runs cleanly at about 60 frames per second. Jens picks fast turning, a floating joystick, medium zoom – and painted frames instead of a part rig, because the rig wobbles while walking. Characters and movement, he says, are everything.
Build teams had only committed locally; the container was discarded. Four finished packages and half of a fifth are gone. Since then: push after every commit.
Five painted walk tests, the same flaw: the image model copies the reference's leg position. The budget for hero sheets is raised step by step.

Forest, desert, orc town, temple and tower – the first art request, approved in full.

The leg rig prances, the mask correction makes the figure jump, in-betweens from the image model wobble. Jens releases budget for dedicated walk research: the goal is still a perfect walk.
Upper body painted, legs in code with motion curves from animation craft. Then the AI finds the cause of the jitter: not the images, but the interpolation.

Jens signs off path D. Only the cape still hangs like a long braid and should swing more.
The API credit included in the subscription gets a job: art checks via image analysis and code reviews.
Gate 3 gets split: Jens explores the first world alone so his friends are amazed when they first connect. An intro with a narrator is added too.
The weekly quota is almost used up. No shifts until 16 October; planning continues in chat – liveliness, art upgrade, bestiary.
28 new base types: orcs as a war faction, trolls, ogres, dragons up to a world dragon, mimics in five percent of chests. And four dark heroes you first fight and later play.
Jens has the new 2.5-series models compete with the current one. Until the comparison is done, no final images.

Jens notices himself that the pipeline still uses older image models although newer ones are available. Claude had missed it and owns the mistake.
Two shifts abort: the safety filter mistakes adjusting test values to new budget limits for removing security tests. main stays red until Jens steps in.
The studio intro from Star Slayer now also plays before Titan Slayer.

For the weekend two slots work at once. Up to four build teams run on four cores – the project's longest shift takes 254 minutes and integrates a single package.
From now on every shift is measured: duration, tokens, sub-agents, test runs. The numbers on this site are based on that.
The first voice test with OpenAI is unusable except for one recording. That same morning the switch to ElevenLabs is decided.
A dedicated test page for tapping through voice candidates. By evening 13 voices, 129 sounds and 8 music tracks are picked.

A large animation package is taken back out of main because a click test turns red.
Claude analyses 15 shifts and finds that the tooling, not the code, is the bottleneck. Jens adopts rebuild H1 to H10.
The controls get painted frames in the style of the Warcraft III interface.

The chequered graybox gives way to a painted forest with cliffs and paths.


Jens: characters float, the ground looks like wallpaper, the potion should be an old glass bottle. That becomes six packages.


A different software renderer in the container: the test scene now runs at 17 instead of 150 milliseconds per frame.
The first painted spell effects are in the game.

Jens asks for the bright blue potion to become an old glass bottle. Turns out it wasn't a potion at all, but the loot bag drawn in code.
Version 0.2.0: three classes, a dummy boss, four camp types, loot, saving, sound.
Measurements perfectly smooth, saving works, camps fair. But the animations look bad, area spells are one giant cloud, and you have to see every shot fly. Jens decides: art and movement first, then new content.


Wind-up, strike, a brief hit-stop of 65 to 120 milliseconds, bobbing while walking. Frame rate and walking speed are now linked.
In chat Jens names 41 as the number of allowed shifts for the week. The chat session writes the value into the control panel itself – although only Jens may change control values – and reports the breach straight away.
After the effect packages: almost all spells are much better. But the painted enemies look more alive than the heroes. From Saturday, movement becomes the top priority.

A big plan emerges in chat: 3D with Meshy and Blender too, video too. A Blender test starts right away.

Two hours later: no 3D, no video. 3D would only work for human-like figures, and then heroes and creatures wouldn't match any more. At 21:11 the paper doll is cut too. Heroes will be painted frame by frame.
Jens and Claude plan titan-slayer.com in a question-and-answer session. The first version with everything so far is up the same evening.