Tried the No-Code Game-Building Path: I Put WorkBuddy First
Community Discussion · General

Tried the No-Code Game-Building Path: I Put WorkBuddy First

Galaxy BrothersGalaxy Brothers2d ago2026/11/16 159 views

Title: Making games without code, the last lesson I had to add was acceptance

Making games without writing code, what really trips you up is judging whether what the AI hands back is right — the engine layer isn't the blocker. The engine's AI can already handle it; nobody's minding the acceptance layer.

Last month I ran a small demo with this setup and got stuck in the middle. The AI can generate, but I first have to be able to say clearly what I want; more troublesome is that after generation, I can't quickly spot which line is wrong. My spec doc has over a hundred entries, scattered across three Feishu docs, a pile of screenshots, and two old docx files (I've used Feishu for two months, and exporting back and forth has worn me out). If you just dump this whole mess at the AI, what it gives you is probably wrong, and confidently wrong — it won't mark which line is a guess.

So what I changed later was the acceptance step, not the prompts — three specific places.

First place is writing "what counts as correct" as column constraints. Environment is still Win 11 + WorkBuddy desktop version, may not apply to everyone. Workspace, asset directory, document processing tasks all as usual; the key is that output contract input box. The term sounds intimidating, but it's just your format requirements for the result — the more rigidly you write it, the more stable the output.

The first time I wrote "help me organize the spec," and the output table's column names were different every time, and it took the liberty of adding a "remarks" column. Later I learned: write column names one by one, hard-coded: ID / Name / Type / One-line description / Dependencies / Status; no merged cells; empty values always written as "TBD," not "none" and not left blank.

This isn't OCD. If column names drift or an extra column appears, I know at a glance it went off-script, without reading line by line. Errors become visible, and acceptance cost drops.

Second place is checkpoints — I set it to stop every 20 entries processed. A checkpoint is a mid-way gate: it runs there and stops waiting for you to click confirm; if you don't review, it doesn't proceed. I've used this for two months, setting it is muscle memory now. It's a must, and what it guards against is me failing to review. Running all hundred-plus entries at once, if one in the middle is wrong, you have no idea where to start looking. Split into 20-entry chunks, I spend two minutes scanning each, roll back on the spot if wrong — way cheaper than redoing half a night at the end.

Third place is locking down permissions. If anyone can edit, it's the same as nobody accepting. The workspace can be shared; I give the lead artist and programmer read-only on the spec table and editable on the task list. Read-only matters a lot — once everyone could edit, and the next day the table had three different numbering schemes, and I redid half a night. Acceptance authority has to be centralized in one place, otherwise everyone's editing, which means nobody's responsible for the result.

Day to day I keep just one habit: every Friday run an append. New spec entries go into assets, select append mode instead of overwrite mode in the task, and run through the checkpoint after. Overwrite mode will wipe out cells I've manually edited before. Manual corrections are the most expensive asset; wipe them once and I don't want to step on that again.

The difference between the two approaches is also in these places. At the start, one is describing to the AI and re-explaining the background each time, the other first compresses materials into a spec sheet, materials go in only once. On format, one is whatever, the other hard-codes columns and keeps them aligned. On traceability, one loses track of which version is latest after chatting a while, the other treats checkpoints as natural version nodes. On batching, one does one at a time, the other dumps a hundred at once and outputs in batches.

Everyone says no-code lets imagination turn directly into a playable prototype. There's still a gap, it's just moved back one slot — moved to whether you can say clearly what you want, and whether you can judge it did it right.

2 replies

?
Ctrl + Enter to reply
Lun Wen He

The hardcoded column names thing is so real—same when I edit tables in papers; if you don't hardcode it, it just adds a remarks column for you.

I am Dapeng
Reply to Lun Wen He

Listen to me, the remarks column is still being polite. OP's "TBD" not being written as "none" is the brutal part—I added a "none for now" and got sent back to redo it.