Chasing 8.5

I started this game with a number in the prompt. Claude is supposed to score each part of its work out of 10 and “iterate until it’s 8.5+”. Three days in it hasn’t gotten there. The best average so far is 7.98, and the latest is 7.22.

The prompt went in on Sunday. I want a web-based 2D game in a cyberpunk setting similar to Shadowrun, with the game loop of Phantasy Star IV on the Genesis. A few things are spelled out (a world map, a town map, a small squad, turn-based combat). For the rest I wrote “all other core decisions I leave up to you”.

I have fond memories of playing both on my Sega Genesis back in the 90s, and I’ve always liked the mix of technology and magic as a framing device in fantasy stories.

Sunday produced 12 commits: an engine, a battle system with a balance simulator, a 14-track soundtrack, one chapter of content, and a test that plays the chapter start to finish and passes. Claude drew all the art in code. The title screen says Shadow Jog.

Quality rounds

Yesterday was the first quality round. Reviewer agents scored 11 areas of the game, the scores ran from 3.5 to 7.6, and the average was 6.36. Claude fixes what they name and they grade it again, and yesterday alone had 134 commits. Round 12 landed today with an average of 7.98 and nothing under 7.0.

The reviewers never play the game (they read screenshots, code, test output and measurements). The automated loop stops at round 12, and the next gate is me.

My first playthrough

I got partway through the chapter today and came back with 25 notes. Some are small, like selling junk in one go, or what the “2” on some enemies means (a squad number for two of a kind). A few are about pacing: attacks go by too fast, and Sinkline random battles came about every 24 steps. The bar and the gun shop have names too close together, and Rook is now a wounded veteran, which was one of three options I picked from.

Claude worked through all 25, and then I asked for one more round of scores. The average came back at 7.22. The scorecard blames the drop on three things: new bugs from the fixes, trade-offs I asked for (slower battles, finer creatures next to a chunkier party), and older problems scored more strictly. That last group includes a story gap that scored 8.2 in round 12 and took Narrative down to 6.0 in round 13.

The new bugs are already fixed, but I haven’t had the game re-scored since, so 7.22 is the number for now. The last reviewer went back over my 25 notes: 21 are done, 4 only partly. The gun shop is Last Rites Arms now.

From here the gate is me playing the game, and my notes are the work queue.