Tips for working with Kimi

What I learned building the game, mostly the hard way.

Give long prompts as a file

Save the task as a Markdown file in your project and type "Read prompts/task.md and do the task." When I pasted a long prompt into the terminal, only the last third arrived, and Kimi went off with no rules and no test images.

Let Kimi write the image prompts first

Ask for the full image prompts before anything is generated, read them, then let it run. Kimi wrote a shared style prefix and a framing text for clean image-to-3D input, and it reviewed its own concept images: it rejected 27 takes for shadows, cut-off parts and text, and kept 10.

"Measure, don't guess"

Put that sentence in the prompt. Kimi is strong at anything it can measure: grip offsets, frame times, VRAM, triangle counts, playtests. Without it, it guessed: the sword was turned around an assumed pivot and ended up 25 cm from the hand. With measured numbers to hit, it fixed it.

Check every visual result yourself

Kimi's own visual reviews passed a boss with a hole in its back, oversaturated colours, a tiny hero and a sword through the hood. To close the hole, it once turned the whole boss into a plain green ball and accepted that. Almost every visual problem in the game was found by me, not by Kimi. Look at the renders and play the game.

Draw three-quarter concepts for image-to-3D

Image-to-3D builds exactly what the picture shows. My first tree concept was a flat front view, and TRELLIS.2 made a tree that was twice as wide as it was deep, a slab. A three-quarter view with a clearly round shape fixed it. Use square 1024 × 1024 images, and check the mesh: depth at least 80% of the width.

More that helped

Name the exact error and file when something breaks. "Fix the MCP connection" made Kimi break the one server that worked.
Add rules against "fixing" your setup: no installs, no restarts, no edits to custom nodes; stop and report if blocked. See the rules block.
Detail you'll throw away can still hang the pipeline. TRELLIS.2 at 1536 got stuck on the cottage for over an hour; at 1024 it took about 10 minutes.
Compare benchmarks fairly. ComfyUI caches nodes, so a second run can look faster than it is. Ask Kimi to compare cached-node counts.
Ask for an automated playtest. A bot that plays the level with real inputs catches broken gameplay after every change.